Before I :mad:while trying to further sort out the details of Cpk and Ppk, I would like to know if there is something more useful for my situation.

The only reason I even need to think about any of this is we are considering going for TS16949 cert and the PPAP form (book!!) that customers may require of us has pages designated for statistical analysis.

Us: a small company that makes capacitors. The capacitors of interest (auto related) are large in diameter, and made in SMALL quantites (so far). Orders of 1-10 are common, orders of 100 have been seen. Every stage relies on considerable, if not total, human control - not much automation. Each run is basically a new item, tailored to the customer's requirements; so no large amounts of data on any one item will likely ever exist.

What sort of statistical data can I possibly gather that would have any analytical value?:confused:

Thanks in advance!:D

I think you might be putting the cart before the horse. In order to determine the proper (if any) type of statistical analysis to use, you first need to determine what you want to know. A few questions in this regard:


What's the likelihood that my current processes will produce nonconforming output?
What are the variables in the process that need to be controlled?
How can I measure the variables in a meaningful (repeatable and reproducible) way?
What's important to my customers?

What measureables do you have that might interest the customer? Capacitance? Sizes? Lead locations? Lead lengths? What are the specifications? Usually they focus on key characteristics - things that if wrong, the product will not work.

What's the likelihood that my current processes will produce nonconforming output?
We have data that can show this with our standard products, but when you only make 10 or so of an item how would you show this? It would be doable if I combined all data from all customers.... but is that a sound approach?

What are the variables in the process that need to be controlled?
Winding speed & tension; overlap & margin measurements; diameter (which determines capacitance); people; mixtures; placements

How can I measure the variables in a meaningful (repeatable and reproducible) way?
Calipers for some, the cap fixture for that; scale for some; and the rest vary as need arises - lots of other data that is qualitative, not quantative.

What's important to my customers?
Cap and physical dimensions top the list.

The issue is more that lack of enough sample points to do any statistical analysis on any one item. We check, we measure, we test, we verify - mostly with a pass/fail mode. :)

What's the likelihood that my current processes will produce nonconforming output?
We have data that can show this with our standard products, but when you only make 10 or so of an item how would you show this? It would be doable if I combined all data from all customers.... but is that a sound approach?
You should have some form of historical data viz rejects, scrap, etc. that will tell you something. My question was more in the form of, how worried are you?

What are the variables in the process that need to be controlled?
Winding speed & tension; overlap & margin measurements; diameter (which determines capacitance); people; mixtures; placements
And how do these variables relate to failures, historically speaking?

How can I measure the variables in a meaningful (repeatable and reproducible) way?
Calipers for some, the cap fixture for that; scale for some; and the rest vary as need arises - lots of other data that is qualitative, not quantative.
How do you know that your measurement methods are trustworthy (accurate, repeatable and reproducible)? You can do statistical analysis on attribute (qualitative) data, so that's not an issue.

What's important to my customers?
Cap and physical dimensions top the list.

The issue is more that lack of enough sample points to do any statistical analysis on any one item. We check, we measure, we test, we verify - mostly with a pass/fail mode. :)

Do you record the results? Do you have an accumulation of data?

The NIST/SEMATECH e-Handbook of Statistical Methods (http://www.itl.nist.gov/div898/handbook/) is highly recommended.

You should have some form of historical data viz rejects, scrap, etc. that will tell you something. My question was more in the form of, how worried are you?
And how do these variables relate to failures, historically speaking?
Do you record the results? Do you have an accumulation of data?


None of the mentioned variables are of high concern. Any of the factors of concern are tested for multiple times and I am 99% certain that a product of questionable quality will never reach a customer on this line.

There are occasional failures, but they are found and eliminated during testing. And the more we perfect our process the less failures there are. There is some electronic accumulated data I was looking at yesterday and I have to question its accuracy regarding failures. Investigation required. If needed I can go back to the paper sources (ugh!) and get the rest of the information.

One statistic good to publish - we have had ZERO customer complaints regarding this series, and only one return for a design change. (a couple returns for evaluation when customer overtaxed unit or used it in an environment for which it was not designed)

But I think I know where I'm headed with this. That's a step in the right direction! :thanx:


How do you know that your measurement methods are trustworthy (accurate, repeatable and reproducible)? You can do statistical analysis on attribute (qualitative) data, so that's not an issue.


The NIST/SEMATECH e-Handbook of Statistical Methods (http://www.itl.nist.gov/div898/handbook/) is highly recommended.

We have a regular calibration/maintenance schedule... but no GRR study has yet been done. I'll have to figure out how to do that, too. :mg:

:thanks: for the e-book resource. It beats buying books based on titles without having a clue what you'll find inside!
:agree1:

TS16949 requires statistical analysis where appropriate.

Do you test your product 100% on your small batches?

If so then you can supply this data to to the customer in lieu of capability data.

Capability data is to establish that the probability of a defective part in a large production batch is very small. If you have 100% tested the parts you have verified that there are no defects. Don't over complicate things.

:biglaugh:Yes! We do 100% testing on this line. Woohoo.

I was hoping something this simple would be my solution. After all, we havn't had the auto customers request ANYTHING yet in the way of PPAP - we just want to be prepared for those who eventually will.

Thanks!

Now you may still need to do some statistical calculations. For example, say you have made 59 products and all 59 tested okay. What can you say about a failure rate for future production? What is the 95% confidence limit for what failure rate could give you 59 of 59 satisfactory results?

You mentioned that your company is considering pursuit of ISO/TS 16949 and that you are a small business. Don't worry so much about your company size, but I would take a long look at your documentation processes. When setting up your QMS to be compliant, you need to always ask yourself, "what value is this adding to our business?" What is management's perspective regarding ISO/TS 16949? If it is one of ISO is a necessary evil? If so, run away! Management must understand that ISO is a system for business management, not just the quality department. You may find that considerable documentation, over and above what you are currently generating, will be required. Are you and your company ready for that?:2cents:

Thanks rmf180 - we are already ISO 9001:2002 cert. A FABULOUS documentation system is already in place (no,:notme: it was here before I was). Mgt here is all for it - sees it as essential if the company is going to grow like they want it to.
We are considering TS because we are hoping that several auto makers will consider using our Power Ring technology in their alternative-energy vehicles. We don't want lack of a TS cert to be a reason to not do business with us. :agree:

We have a pre-assessment tomorrow that will help fill in some gaps. But I thank all of you who have been such a help that I won't sound like a complete babbling fool.:thanks:

Now you may still need to do some statistical calculations. For example, say you have made 59 products and all 59 tested okay. What can you say about a failure rate for future production? What is the 95% confidence limit for what failure rate could give you 59 of 59 satisfactory results?

Is this like "knock knock..."? I have no idea - but I'll bet I can figure it our with the handy reference tool some kind soul here pointed too earlier. (NIST/SEMATECH e-Handbook of Statistical Methods):lol:

I've done what I feel shows the customer what they would want to know about us. What do you think?

Is this like "knock knock..."? I have no idea - but I'll bet I can figure it our with the handy reference tool some kind soul here pointed too earlier. (NIST/SEMATECH e-Handbook of Statistical Methods):lol:

I've done what I feel shows the customer what they would want to know about us. What do you think?

One thing I would do is convert the bar chart to a control chart (p-chart specifically) and determine if the process is in statistical control and thus is stable and predictable.

And the answer to the upper 95% confidence limit is 5%. Take the number of results without failure and divide by 60 (this is an approximation, but close enough).

Assume your company makes hundreds of thousands of small run, essentially different parts for each order. There is a lot of data to be had, and providing information about the process is what the customer is after. Think process...not the product.

Now, each one of those products produced has something in common: Each is trying to hit a target value that is ideal. Therefore, regardless of the different parts made, what can be tracked is the distance away from that ideal target no mater what its value may be. An example may illustrate:

A (very small) company has one lathe and turns parts. It is a job shop to be sure. Each order is one, two, or even ten parts. No more. They range in size from 1/2 inch to 15 inches in diameter. Question is, what do these parts have in common? Answer: Diameter. Therefore tracking the distance from the target diameter yields the process capability. One chart will accommodate many, many various sized parts. Easy.

Hope that helped.

Assume your company makes hundreds of thousands of small run, essentially different parts for each order. There is a lot of data to be had, and providing information about the process is what the customer is after. Think process...not the product.

Now, each one of those products produced has something in common: Each is trying to hit a target value that is ideal. Therefore, regardless of the different parts made, what can be tracked is the distance away from that ideal target no mater what its value may be. An example may illustrate:

A (very small) company has one lathe and turns parts. It is a job shop to be sure. Each order is one, two, or even ten parts. No more. They range in size from 1/2 inch to 15 inches in diameter. Question is, what do these parts have in common? Answer: Diameter. Therefore tracking the distance from the target diameter yields the process capability. One chart will accommodate many, many various sized parts. Easy.

Hope that helped.

Where would you start when the chart shows that the process is out of control? There are so many variables that can contribute to the OOC condition. In your example, a diameter range from 0.5 to 15 inches by itself is a major source of variation, i.e.: different materials, cutting speed, tool wear, different measuring tools, etc.

Stijloor.

Where would you start when the chart shows that the process is out of control? There are so many variables that can contribute to the OOC condition. In your example, a diameter range from 0.5 to 15 inches by itself is a major source of variation, i.e.: different materials, cutting speed, tool wear, different measuring tools, etc.

Stijloor.
I would have charts for different materials but for the same feature. In this example, diameter. A chart for diameter (for steel); a chart for diameter of brass, a chart for diameter of aluminum etc. As stated earlier, the control feature is a targeted diameter; what all the shafts have in common. Chart the deviation from the target. For example, on the .500 shaft you were .0010 over the target of .500. Chart .001. The next job is a 15 inch shaft and you were .0015 over. Chart the .0015. Last, you cut a 6.500 inch shaft and are -.0050 from the target. Chart the -.0050. This is a target chart.

Capability could be estimated against the different specifications, but I would use the tightest tolerance to proclaim my capability then remember to mention that specification width during reporting.

...Now, each one of those products produced has something in common: Each is trying to hit a target value that is ideal. Therefore, regardless of the different parts made, what can be tracked is the distance away from that ideal target no mater what its value may be. An example may illustrate:
...One chart will accommodate many, many various sized parts. Easy.

Hope that helped.

Wow... I feel like a light bulb just went off in my brain.

I have diameter and capacitance values that will make two lovely charts. So... would charting it as % deviation or as plain deviation be more representative?

Wow... I feel like a light bulb just went off in my brain.

I have diameter and capacitance values that will make two lovely charts. So... would charting it as % deviation or as plain deviation be more representative?

PLAIN DEVIATION.

sorry to yell, but % deviation is not a statistic. It's business math for those people who find numbers just too confusing. more to the point it puts the average - or target) right back into the equation and you'll have a LOT of variation and out of control points - simply because you now have a bigger part than before even when your variation is the same.

Wow... I feel like a light bulb just went off in my brain.

I have diameter and capacitance values that will make two lovely charts. So... would charting it as % deviation or as plain deviation be more representative?

Don't chart % deviation, rather, plot the actual value (X-chart) away from your target with...a target chart. Again, the objective is to determine the process capability...not the product. We use, however, the product that is generated from that process to measure this. Since you have as few as one item per order, I might suggest a moving average -range chart (mX R) chart with a subgroup size of two or three; 3 being more desireable (given a normal distribution). If it is not, simply use a larger sample space.

Frequency of sample should be determined. Perhaps every 50th piece. IE: 1,2,3 - 51,52,53 - 101,102,103 and so on. Determining ARL (average run length) is a discussion for another thread.

Dr. Donald Wheeler wrote extensively on the use of target charts for shot run capability as did others.

Assume your company makes hundreds of thousands of small run, essentially different parts for each order. There is a lot of data to be had, and providing information about the process is what the customer is after. Think process...not the product.

That is true!

Now, each one of those products produced has something in common: Each is trying to hit a target value that is ideal.

For normal or near normal distributions (those with central tendencies), that is true. It could be that your processes have such large tolerances that for variation to be significant it has to be pretty huge. In that case, your total variation (measurement error, gage error, roundness, etc.) may collectively behave as 'normal', and is not significant. Had you been doing precision machining of these diameters, however, a totally different approach would have been required (Those pondering this approach with precision machining, please don't do it. It is statistically incorrect and will ensure overcontrol.)

One chart will accommodate many, many various sized parts. Easy.

If you were doing precision machining of those diameters, and wanted to know "process capability", it is quite easy. Control the process correctly (http://elsmar.com/Forums/blog.php?b=79), and run it to 75% of the tolerance for that dimension. You are now capable. Easy.

As far as capacitance - that is a whole other question.

PLAIN DEVIATION.

sorry to yell, but % deviation is not a statistic. It's business math for those people who find numbers just too confusing. more to the point it puts the average - or target) right back into the equation and you'll have a LOT of variation and out of control points - simply because you now have a bigger part than before even when your variation is the same.
Perhaps I should have found a better term, because I don't mean in the averages sense of % dev. I mean the % that the value has deviated from the target... % "off target".

With the diameters the tolerance is a value - most of them the same value (.03) - which will define the upper and lower limits as I graph. Graphed as values this looks pretty good. With diameter ranges from 3" to 15" this graphed as % "off target" looks like an erratic heartbeat.

For capacitance the tolerance is defined as a percent. Go the % "off target" graph looks pretty good, but with cap ranging from 30 to 4000 its the value graph that looks ridiculous.

Unless there is some reason not to, I will use the value graph for the value defined tolerance, and the % graph for the % defined tolerance.

:thanx:

Perhaps I should have found a better term, because I don't mean in the averages sense of % dev. I mean the % that the value has deviated from the target... % "off target".

With the diameters the tolerance is a value - most of them the same value (.03) - which will define the upper and lower limits as I graph. Graphed as values this looks pretty good. With diameter ranges from 3" to 15" this graphed as % "off target" looks like an erratic heartbeat.

For capacitance the tolerance is defined as a percent. Go the % "off target" graph looks pretty good, but with cap ranging from 30 to 4000 its the value graph that looks ridiculous.

Unless there is some reason not to, I will use the value graph for the value defined tolerance, and the % graph for the % defined tolerance.

:thanx:

Cool.
Watch out for the % of tolerance. (Torque values in the auto inductry do the same thing). The information this yields can be mis-leading. ie 10% of 1000 is a lot more than 10% of 100.

If you track both (actual & percent) try converting the percent value into a log scale. This may provide a better picture than the erratic heartbeat data you are getting.

Good luck....let us know how it works!

alternately - post your data here and let us take a look at it.

sometimes the pretty chart is lying to you and the ugly chart is providing the truth and all of the value.

SPC is not as easy as pciking a chart that looks good and going with it.

Also % of target has the same problem as % of average...looking at teh raw delta information will tell you much more about your process. I have seen interactions where the size of 'thing' being measured, increases or decreases the variation (the so-called CV effect). It woudl be important to know that...

Perhaps I should have found a better term, because I don't mean in the averages sense of % dev. I mean the % that the value has deviated from the target... % "off target".

With the diameters the tolerance is a value - most of them the same value (.03) - which will define the upper and lower limits as I graph. Graphed as values this looks pretty good. With diameter ranges from 3" to 15" this graphed as % "off target" looks like an erratic heartbeat.

For capacitance the tolerance is defined as a percent. Go the % "off target" graph looks pretty good, but with cap ranging from 30 to 4000 its the value graph that looks ridiculous.

Unless there is some reason not to, I will use the value graph for the value defined tolerance, and the % graph for the % defined tolerance.

:thanx:

I am curious, will you be able to trace the data to individual parts?

My customer seem to focus on individual parts to the extreme regardless of run conditions.

sometimes the pretty chart is lying to you and the ugly chart is providing the truth and all of the value.

SPC is not as easy as picking a chart that looks good and going with it.

Absolutely - many times a 'pretty' chart looks controlled and normal only because it is really overcontrolled by the operator. The operator has become the process. And, that is about the worst option.

For capacitance the tolerance is defined as a percent. Go the % "off target" graph looks pretty good, but with cap ranging from 30 to 4000 its the value graph that looks ridiculous.

Here is an interesting question on capacitance - what is the process variation that is the cause of the variation in capacitance? And are they related linearly?

Here is an interesting question on capacitance - what is the process variation that is the cause of the variation in capacitance? And are they related linearly?
Envision a parallel plate capacitor - Cap value is a result of the area of the plates. Now take those plates and wrap them around a pencil. The area is now a function of thickness and diameter. That's us. The capacitors in this family are custom designed for the customer's needs - both size and cap. Because of how these are made the thickness becomes a non variable (per part run), so its the diameter that tweaks the cap.
alternately - post your data here and let us take a look at it.

sometimes the pretty chart is lying to you and the ugly chart is providing the truth and all of the value...
I would love to have your collective opinions. Excel file attached.

And yes, Boscoee, I can trace the data to the individual piece with this family. With our more standard parts... I could narrow it down to a production run (lot) which might be 100 pieces or 3000 pieces.

Envision a parallel plate capacitor - Cap value is a result of the area of the plates. Now take those plates and wrap them around a pencil. The area is now a function of thickness and diameter. That's us. The capacitors in this family are custom designed for the customer's needs - both size and cap. Because of how these are made the thickness becomes a non variable (per part run), so its the diameter that tweaks the cap.


That's why I pondered the relationship of the change of the diameter and the capacitance being linear or not. Sure, changing the diameter tweaks the capacitance, but the nature of the spiral makes me think the incremental change in diameter at 1" and 1/2" ( as well as the difference in length - if any) must be significantly different due to the area variation at each of the diameters. Without working out the trig, I bet it is not linear. Maybe it is linear enough in the range you are reviewing? Not sure. Something to consider.

That's why I pondered the relationship of the change of the diameter and the capacitance being linear or not. Sure, changing the diameter tweaks the capacitance, but the nature of the spiral makes me think the incremental change in diameter at 1" and 1/2" ( as well as the difference in length - if any) must be significantly different due to the area variation at each of the diameters. Without working out the trig, I bet it is not linear. Maybe it is linear enough in the range you are reviewing? Not sure. Something to consider.
Oh, definately not linear. More calculus than trig, as the depth must be considered also.

I would love to have your collective opinions. Excel file attached.

And yes, Boscoee, I can trace the data to the individual piece with this family. With our more standard parts... I could narrow it down to a production run (lot) which might be 100 pieces or 3000 pieces.
Still hoping one of you might voice an opinion about which charts on the above excel file best represent the truth of the data. :rolleyes: