My company has a generic reaction plan for SPC. The plan discusses the different reaction to Cpk lower than 1.33 for the different types of KCC's. I have attached the procedure to help in this discussion. We are a company making fuel injectors so the processess are high precision grinding, honing, and matching; measuring down to the micron level.

I am looking for help in rewriting this procedure. Currently, the procedure strates that any process below 1.33 and is an SC has to be sorted 100%. Needless to say, this is not happening nor do I feel it necessary (I've only been at the company 1 month so they are looking for fresh ideas).

I would like help iun defining an action plan that can be implemented and with allow for our registrar to see the improvement process. As far as I know, there is no requirement in ISO/TS or in the customer specifics that require 100% sort for low Cpk. We are only discussing low CPK, not out of spec conditions. We do have a CAPA(coorective action / preventive action) meeting to discuss these items but it is not very structured. SO a detailed prcess will help.

If you could look at the current plan and help with a more feasible one, I would appreciate it. Iwas thinking along the lines of opening an imporvemtn project involing QA, Manufacturing, and Design and Process engineers

:thanx:Thanks in advance to all who can help.

I looked at the plan and found it to be quite rigid and maybe a bit extreme. I would never suggest sorting product that is in control and having a Cpk above 1.33. Who ever does that? We are wasting $$ here.

Having product with a Cpk above 1.67 should have a reaction of "RUN" and even with a smile. That bit about improving the process is extreme. I would love to be a 3rd party auditor and reviewing this area. I would simple ask how you are improving the process on products with a Cpk above 1.67 per your reaction plan and watch someone squirm. IT IS NOT HAPPENING!

If your product has a Cpk below 1.33 but above 1.0 and is in control, I believe that the TS standard states that the product should be 100% sorted. If one is a supplier of the the big 3, I guess that they would have to follow this requirement but is it being performed??? I have a hard time with this and would tend towards making this process one that requires improvement and there should be some documented attempt to improve it.

Certainly, a Cpk below 1.33 and above 1.0 and suddenly becomes out-of-control, I would sort back to the last good check that was in control and change the process.

All processes with a Cpk below 1.0 DO require 100% ongoing sorting.

Using stats in the automotive had become a bit of a game for the last 25 years. There is a lot of "show" going on and it hinders efficiency. We place stats on the wrong characteristic that aren't really vital to its fit, function or safety and it has cost us a lot of $$$. In some cases such as flatness, as an example, we could show stats on a small number of constant points having a great Cpk but if someone swept the full surface, the part could be out-of-specification. Using stats appropriately really does help but sometimes it is not applicable.

Dave's correct about reaction to the various Cpk levels.

but the bigger issue for you may be that many of your processes are Not normal and you may be seriously underestimating the Cpk values... you may want to start by looking at your actual distributions and use more distributionally appropriate capability calculations...

as for the last comment that dave made about the auto industry - I couldn't agree more!

Yes, it has turned into a game at times with CPK versus form, fit, and function for automotive. I have actually had calls from OEM's asking me what our CPK is, and I asked for what specific dimension or product and they would just say they need a CPK value. Disheartening to say the least.

Yes, it has turned into a game at times with CPK versus form, fit, and function for automotive. I have actually had calls from OEM's asking me what our CPK is, and I asked for what specific dimension or product and they would just say they need a CPK value. Disheartening to say the least.

Consider the shape they're in...:mg:

Stijloor.

Thaks all for the quick reply to the post. I wanted to add one thing.

The biggest reason that we have so many low cpk values is two fold.

First, as Bev D stated, we are not using the right moitoring techniques. We have many utilateral tolerances and are measuring GD & T dimensions sush as Perpendicularity, flatness, runout, and profiile. If any body can provide a better technique then X bar R or Individulas charts, please help me.

Second, I know for a fact that the process were in no way under control when the limits were set. You cannot set up control charts until the process is under control. Now it is just a matter of fixing it. Any suggestions?

Third, we are using forced control limits instead of calculated control limits. Are limits are set to allow up to 70% of the specification tolerances. Any pros or cons to this would help in getting it changed or improved.

Thanks again for youre help. If you can also post comments on the above I will be forever indebted.:lol:

:thanks:

timference:

Boy, are you in a mess.

I've had a fair amount of experience in GD&T and they are not really the best features to apply stats. I realize that we can get measurements from the GD&T symbols but if the values are not credible or do not reflect the whole surface and in a repeatable method, what good are they? Love to see their respective R & R values or maybe those might have been "spiced up" to appease your customer. Remember, one must have no more than 10% of process on critical characteristics.

You placed what I call "Cheater Limits" rather than calculated limits. The process may not be capable of statistically holding 70% of the tolerance and in other cases, 70% is too large. Please replace with normal statistical limits but only if you have the correct chart and the process is normal. I would use Modified Control Limits on non-normal processes. Get rid of the cheaters.

Good luck on convincing management here. Don't be too forceful though or you will be pegged as a person with an attitude.

First, as Bev D stated, we are not using the right moitoring techniques. We have many utilateral tolerances and are measuring GD & T dimensions sush as Perpendicularity, flatness, runout, and profiile. If any body can provide a better technique then X bar R or Individulas charts, please help me.
If you have a subgroup size of 4 or 5 pieces, the central limits theorem takes care of the Unilateral tolerance issue and the X bar R chart is fine. If you need to collect and plot individuals data then you run the risk of a false alarm. You can transform the data or use software to plot non-normal control limits.

Third, we are using forced control limits instead of calculated control limits. Are limits are set to allow up to 70% of the specification tolerances. Any pros or cons to this would help in getting it changed or improved.

Forced "control limits" derived in some way from specifications have nothing to do with statistical control limits and will only serve to give you false alarms or a false sense of security. It is common practice, but not what Dr. Deming and Dr. Shewart intended.

... you may want to start by looking at your actual distributions and use more distributionally appropriate capability calculations...

as for the last comment that dave made about the auto industry - I couldn't agree more!

And, don't forget, if you rubber stamp Cpk calculations on your data, you may want to direct folks to AIAG PPAP 4th Edition to ponder if that might be a bad idea:

2.2.11.5 Processes with One-Sided Specifications or Non-Normal Distributions

NOTE: The above mentioned acceptance criteria (2.2.11.3) assume normality and a two-sided specification (target in the center). When this is not true, using this analysis may result in unreliable information.

I love that section...:cool:

Consider the shape they're in...Stijloor.

Let me be the first to jump in with the best manufacturers no longer find a for SPC because of robust design.

I think we can pretty safely ignore all this automotive TS stuff, Quality is always the first against the wall when the hard times come. It will take years/decades before any surviving Quality people can build things back up to the level they were at before the troubles.

Forced "control limits" derived in some way from specifications have nothing to do with statistical control limits and will only serve to give you false alarms or a false sense of security. It is common practice, but not what Dr. Deming and Dr. Shewart intended.

This may be true for form features, but it is absolutely not an absolute. There is a common distribution where it is absolutely statistically correct. In fact, the processes the OP has listed - high precision grinding, honing - utilize that distribution. In fact, I bet the reason why their capability values are low is because they are using the wrong statistics. If they are using X-bar-R charts for those processes, you can bet they are using the wrong charts and statistics. :cool:

Let me be the first to jump in with the best manufacturers no longer find a for SPC because of robust design.

I think we can pretty safely ignore all this automotive TS stuff, Quality is always the first against the wall when the hard times come. It will take years/decades before any surviving Quality people can build things back up to the level they were at before the troubles.

Capability is typically just a shortcut for customers to rubber stamp approval of a process without taking the time to understand what they are approving.

However, SPC does have value in spite of robust design - of either the part or the process - when used correctly. :cool:

Using stats in the automotive had become a bit of a game for the last 25 years. There is a lot of "show" going on and it hinders efficiency. We place stats on the wrong characteristic that aren't really vital to its fit, function or safety and it has cost us a lot of $$$. In some cases such as flatness, as an example, we could show stats on a small number of constant points having a great Cpk but if someone swept the full surface, the part could be out-of-specification. Using stats appropriately really does help but sometimes it is not applicable.

The biggest problem is being forced to use statistics by the standard, but having no idea what they are implementing. This generated a rubber stamp of X-bar R charts all over - whether applicable or not, then choking on the results.

Not sweeping the surface is a measurement error common in the precision machining industry. This is the point of and corrected by using the X hi/lo-R charting. (See SPC for Precision Machining (http://elsmar.com/Forums/blog.php?b=79)):cool:

...but the bigger issue for you may be that many of your processes are Not normal and you may be seriously underestimating the Cpk values...

Bev is absolutely correct. For precision machining, if you are calculating Cpks, it is clear evidence that you are using the wrong statistics to control the process. You have likely been victimized by the Xbar-R chart that is the worst possible chart for your processes. Using the wrong statistics generates the contempt for statistical process control that you may have read here. Your processes are non-normal, and if your customers are automotive, consider that AIAG PPAP 4th Edition states Cpk calculations for your data are incorrect:

2.2.11.5 Processes with One-Sided Specifications or Non-Normal Distributions

NOTE: The above mentioned acceptance criteria (2.2.11.3) assume normality and a two-sided specification (target in the center). When this is not true, using this analysis may result in unreliable information.

Do not feel bad, is has been going on for years, much to the dismay of people doing precision machining everywhere. Chances are it would be nearly impossible for precision machining to truly be so far out of control that you would have to sort if they were controlled properly. You have that gut feeling, you need to understand why your gut feeling is true - and how to get there correctly. Now, you will have to sort for special causes, such as broken tools. That is a given.

Rather than focusing on the reaction plan, I would suggest revamping your SPC to a system more applicable to you processes. See SPC for Precision Machining (http://elsmar.com/Forums/blog.php?b=79).

Not sweeping the surface is a measurement error common in the precision machining industry. This is the point of and corrected by using the X hi/lo-R charting. (See SPC for Precision Machining (http://elsmar.com/Forums/blog.php?b=79)):cool:

Whenever one gets into GD&T symbols such as flatness and any of the 3 angles, one must sweep the complete surface for the data to have any credibility and I do agree that measurement error could be rather high.

Did we set up the same way each time. Was it a 3 point set up for flatness or was it "best fit" per CMM. It does makes a difference. Did we just use strategic points on the surface? How did we zero off on the surface? I think that this is more important than the hi/lo approach since we could have different results from the same part.

On any of the angles relative to a datum, did we use a 3 point set up on the datum or use a datum simulator (flat surface) which could make the part rock. Where was our first reading taken and do we emulate that each time?

I would love to see the R & R study taken on flatness? Is it no more than 10% of the process? I don't think so.

I guess that I am not a great proponent of stats used in GD&T but if you have to, I would suggest a chart that is not used that often - modified control limits on the averages and R chart. This would absorb all the runs and trends acknowledging that the process and the way it is measure may not be normal.

Whenever one gets into GD&T symbols such as flatness and any of the 3 angles, one must sweep the complete surface for the data to have any credibility and I do agree that measurement error could be rather high.

You are right, flatness is going to be a particularly difficult form function to deal with - and you have aptly described many reasons why. But, because it is difficult does not exempt it from control - just makes it more challenging. Controlling the measurement process (not just the gaging) will clearly be a part of capturing meaningful data. Your points illustrate those issues well. I would love to see some studies with a well controlled measurement process and machining process to get a clear idea of what the true distribution of this variable would be.

The X hi/lo-R chart is a rather simple approach to control roundness and parallelism (and can be adapted for taper). You are actually controlling the GD&T of a length or diameter feature using that approach. X-bar-R ignores the effect of roundness or parallelism, causing a gross misrepresentation of the process variation, as well as being based in the wrong statistical distribution. :cool: