sample = 300
subgroups = 60

Target = 0.093 +/- 0.005"
Some points of 0.98 and 0.99 are seen on charts.

Bad parts detected and it's 0.105"!!!

I'm still scratching my head as how the cpk is still 1.99. I hope someone can shed some light on this. TIA.

Hi:

For Cpk to be valid, the data has to be normally distributed.

Is the distribution normal? How many out of control points are there? Can you share the actual data with us?

I just received the chart from our supplier. 9 points out of specification limit. the x-bar R chart is out of control period.

sample = 300
subgroups = 60

Target = 0.093 +/- 0.005"
Some points of 0.98 and 0.99 are seen on charts.

Bad parts detected and it's 0.105"!!!

I'm still scratching my head as how the cpk is still 1.99. I hope someone can shed some light on this. TIA.

Personally, I don't put much faith in cpk. What I suspect is happening is that you have a set of average and control limits that were established at some point in the past. Hopefully during a time when there were no out of control points (1st potential error- doing cpk calculations on a process that is not "in control" is meaningless). Perhaps recently something has introducted a change to the process leading to current out of control points. If they are sufficient to lead you to believe that there is a permanent shift in the process, the average and control limits should be shifted, and a new cpk determined (potential error 2 if this has not been done). If the current out of control points are sporadic, not justifying a shift in control limits, but are indicative of an "out of control process", then again, the cpk value is meaningless as long as you are out of control.

By the way, normality is NOT a requirement to make a control chart. One can argue whether or not normality is needed to make a cpk, but again, I don't put much faith on cpk.

I just received the chart from our supplier. 9 points out of specification limit. the x-bar R chart is out of control period.


Then the CpK is not valid. You could also verify their calculation method. It seems unlikely that CpK is high (almost 2.0) with out of control and out of spec points.

Thanks for all the reply.

when you calculate the cpk , i think first you should collect the data and make the X bar - R chart , and based on the result of X-R chart ,make sure there is no instable cause , then you can analysis the cpk .

if there r some date out of limit, what you do is first find the cause and correct it ,and eliminate this data ,confirm the process is stable,then you can calculate the cpk

i just do quality job two years, and this is only my personal thought, welcome to comment
thank you in advance

If one has 60 sub-groups of 5 samples on an X bar and R chart and somehow arrives at a Cpk of 1.99, something is wrong.

I would suspect that if you calculated the control limits, some are out of control. Frankly, they would have to be out of control. If that is the case, one cannot calculate Cp or Cpk.

Cpk of about 2 means that 6 estimated standard deviations can fit in the specification limits. How could that happen when some parts are out of spec?

Yup - I would bet my last $$ that you process is not in statistical control.

If one has 60 sub-groups of 5 samples on an X bar and R chart and somehow arrives at a Cpk of 1.99, something is wrong.

Why?
I would suspect that if you calculated the control limits, some are out of control. Frankly, they would have to be out of control. If that is the case, one cannot calculate Cp or Cpk.

Cpk of about 2 means that 6 estimated standard deviations can fit in the specification limits. How could that happen when some parts are out of spec?

See Steve Prevette's comment in this thread; the most likely prospect is that the mean and control limits need to be recalculated. The first question to ask is why out-of-spec conditions were charted in the first place. If the process is producing material that doesn't meet specifications, we know that the process is grossly incapable, regardless of statistical capability. At the point where the out-of-spec measurements took place, someone should have stopped and done something about it.

Yup - I would bet my last $$ that you process is not in statistical control.

There isn't enough data available (to us) to make that assumption, let alone to bet the farm on it.

If one has 60 sub-groups of 5 samples on an X bar and R chart and somehow arrives at a Cpk of 1.99, something is wrong.

I would suspect that if you calculated the control limits, some are out of control. Frankly, they would have to be out of control. If that is the case, one cannot calculate Cp or Cpk.

Cpk of about 2 means that 6 estimated standard deviations can fit in the specification limits. How could that happen when some parts are out of spec?

Yup - I would bet my last $$ that you process is not in statistical control.

The control charts, UCL, LCL, CP and CPK are automatically calculated by the system. They are 60 subgroups, and 8 of them are way off UCL measuring 0.099. CP is 3.08. My thought for this is, when a small portion of a good data is out of control, it will still give you good cpk. A very important rule often overlooked by the operator is that they must first look at the control charts before looking at the cpk.

The control charts, UCL, LCL, CP and CPK are automatically calculated by the system. They are 60 subgroups, and 8 of them are way off UCL measuring 0.099. CP is 3.08. My thought for this is, when a small portion of a good data is out of control, it will still give you good cpk. A very important rule often overlooked by the operator is that they must first look at the control charts before looking at the cpk.

An even better rule for the operator is to stop charting when out-of-spec conditions are encountered. Operators also need to watch for trends (excursions away from the mean in particular). The whole point is to use statistical data as a guide in centering the process variation as far away from the spec limits as is reasonably possible. Having crazy, unintelligible charts after the fact isn't helpful.

See Steve Prevette's comment in this thread; the most likely prospect is that the mean and control limits need to be recalculated. The first question to ask is why out-of-spec conditions were charted in the first place. If the process is producing material that doesn't meet specifications, we know that the process is grossly incapable, regardless of statistical capability. At the point where the out-of-spec measurements took place, someone should have stopped and done something about it.



I think that we should reflect all the sample information even if the reading is out of specification but the reaction to this out of control situation is a bit different. We certainly should correct the process and also place the parts that were made from the last good check in quarantine. The suspect parts should then be 100% sorted.

Cp & Cpk should only be calculated if the process is in control and I cannot think of how a process could be in statistical control achieving a Cpk of 1.99.

I think that we should reflect all the sample information even if the reading is out of specification but the reaction to this out of control situation is a bit different.

While I agree that in a continuous situation the charts should show everything that happens, I suspect that given a 300-piece population perhaps a PPAP submission is involved. I could be wrong about this, but if I'm right, the PPAP SPC report should show a stable, capable process, and not a lot of wacky stuff.

Cp & Cpk should only be calculated if the process is in control and I cannot think of how a process could be in statistical control achieving a Cpk of 1.99.

I still don't understand your reasoning here. Why shouldn't a process that's in control yield a Cpk value of 1.99? (Note that my question has nothing to do with this particular case.)

I still don't understand your reasoning here. Why shouldn't a process that's in control yield a Cpk value of 1.99? (Note that my question has nothing to do with this particular case.)

I am referring to a situation where one has a Cpk of 1.99 and parts out of specification in my previous answer.

Of course a process that is in statistical control can have a Cpk of 1.99.

An even better rule for the operator is to stop charting when out-of-spec conditions are encountered. Operators also need to watch for trends (excursions away from the mean in particular). The whole point is to use statistical data as a guide in centering the process variation as far away from the spec limits as is reasonably possible. Having crazy, unintelligible charts after the fact isn't helpful.


Chart is plotted automatically. Operators shall be trained adequately to analyze control charts:tg:

I've attached the control chart to clarify some confusions. Actual CPK is 1.87, not 1.99 because I received the chart from our supplier only after I posted this message, but they are close.

I am referring to a situation where one has a Cpk of 1.99 and parts out of specification in my previous answer.

Of course a process that is in statistical control can have a Cpk of 1.99.

Thanks for the clarification, but you made no mention of out-of-spec parts being a condition of your statements. Here is what you said:

If one has 60 sub-groups of 5 samples on an X bar and R chart and somehow arrives at a Cpk of 1.99, something is wrong.


Cp & Cpk should only be calculated if the process is in control and I cannot think of how a process could be in statistical control achieving a Cpk of 1.99.

It's helpful if you quote at least part of the post you're responding to so that people don't read comments like those above and not understand the underlying context.

Chart is plotted automatically. Operators shall be trained adequately to analyze control charts:tg:

I've attached the control chart to clarify some confusions. Actual CPK is 1.87, not 1.99 because I received the chart from our supplier only after I posted this message, but they are close.

It appears that whomever sent you the chart just plugged a bunch of numbers into the software, hit "Print" and didn't bother to analyze the results. It's not really even worth reviewing in detail. I've seen a lot of this--people who think that (A) no one is going to parse the data, and (B) all you're looking for is a Cpk number above some defined limit. I also note that the spec limits aren't mentioned or charted. If you have already explicitly defined what you expect to see wrt SPC reports, and this report meets those requirements, you need to revisit them. On the other hand, if you haven't provided an explicit definition, you need to reject this report and tell the sender what you expect to see.

Chart is plotted automatically. Operators shall be trained adequately to analyze control charts:tg:

I've attached the control chart to clarify some confusions. Actual CPK is 1.87, not 1.99 because I received the chart from our supplier only after I posted this message, but they are close.

I reviewed the chart and it is sooooooo out of control in the averages that the Cpk is not valid.

Here are some of the conditions that could give you such a chart:

1
Your gauge uses too much of the tolerance. One must perform an R & R study and then achieve no more than 10% of the total variation (or state of the art).

2
Your sample size is not homogeneous. If one has a 12 station machine (each station affects the dimension) and you are using a 5 piece sub group, the averages would be reflect as shown in your sample. A 6 station machine should have a sample size of 6 while the 12 station machine uses 12.

3

3 The samples were taken in different locations.

Sorry but I hit the wrong button previously.

By the way, one does not have a LCL on the ranges chart unless your sample size is very large, say 15.

Hope these thoughts help out.

By the way, one does not have a LCL on the ranges chart unless your sample size is very large, say 15.


The chart shows zero as the lower bound, which is correct. While you're correct in stating that for this case there is no LCL, the software apparently adds the label anyway, which makes no difference one way or the other because the range can't be a negative number.

let's restart the thread questions:

we really need the raw data and the spec limits to give you an answer that isnt' a guess.

the other question that I would ask you to ask your supplier is: what FORMULA was used to calculate Cpk? the traditional original formula uses the within sample standard deviation isntead fo the total standard deviation. So the original formula will give you an overestimation of the Cpk value vs what you really want to see. given the PDf you attached - the software is really old (or probably nased on older software) and I woudl suspect the formula first - but again without the raw data it's jsut a guess like any one else's guess.

That said I agree with Steve P - the Cpk thing is a waste. plot your data and look at your data.....that is the best way but it takes work.

David,

Can you share the raw data with us? If possible put it in a excel spreadsheet and attach. I am very curious about the actual data. A calculation error or incorrect formula is my suspiscion.

let's restart the thread questions:

we really need the raw data and the spec limits to give you an answer that isnt' a guess.

the other question that I would ask you to ask your supplier is: what FORMULA was used to calculate Cpk? the traditional original formula uses the within sample standard deviation isntead fo the total standard deviation. So the original formula will give you an overestimation of the Cpk value vs what you really want to see. given the PDf you attached - the software is really old (or probably nased on older software) and I woudl suspect the formula first - but again without the raw data it's jsut a guess like any one else's guess.

That said I agree with Steve P - the Cpk thing is a waste. plot your data and look at your data.....that is the best way but it takes work.

Getting the raw data is impossible. The spec limit is 0.093 +/- 0.005".
The formula used to calculate cpk is cpk=min(cpu; cpl) where cpu=(usl-xbar)/3*sigma, cpl=(xbar-lsl)/3*sigma and estimated sigma is used where sigma=Rbar/d2. I would not suspect the formula. As I mentioned earlier, CP is 3.08 and that represents the process is precise. It's just a very small portion that is out of control and that does not have a big enough impact to cause cpk to be below 1.33 or acceptable level. Correct me if i'm wrong.

Getting the raw data is impossible.
If that's the case, then someone's going to have to start over if you expect some output that makes sense. Any manipulation you do at this point is based on conjecture and guessing, which is not what you want in a statistical report.

It's just a very small portion that is out of control and that does not have a big enough impact to cause cpk to be below 1.33 or acceptable level. Correct me if i'm wrong.

Before my wife gave birth, just a portion of her was pregnant. :tg: Kidding aside, something is either out of control or it isn't. You can't just separate the point(s) beyond the limits (unless you know an anomalous outlier exists) and act like nothing happened.

What type of process is this from? I have seen similar charts from continuous processes such as extrusion. As we discussed in a different thread, some processes are autocorrelated, such as your plastic extrusion process.

When you take consecutive samples from this type of process, your control limits are based on 100% measurement variation. Therefore, any shift in the process will appear out-of-control when, in fact, it is not. The best way to handle this type of process is to use an IMR chart and space your samples apart enough that you are outside the period of autocorrelation.

If that's the case, then someone's going to have to start over if you expect some output that makes sense. Any manipulation you do at this point is based on conjecture and guessing, which is not what you want in a statistical report.



Before my wife gave birth, just a portion of her was pregnant. :tg: Kidding aside, something is either out of control or it isn't. You can't just separate the point(s) beyond the limits (unless you know an anomalous outlier exists) and act like nothing happened.

jim, thanks for all your constructive comments throughout the discussion. I really dont expect any sensible output when i saw the control chart. I almost fell off of my chair when I see it! I was asking myself how could the quality guy 'qc pass' this part like that!! Not to mention they are iso certified. this is certainly an implementation problem and I know majority of the company is doing the same thing. Trends on charts? what is it? :tg:

What type of process is this from? I have seen similar charts from continuous processes such as extrusion. As we discussed in a different thread, some processes are autocorrelated, such as your plastic extrusion process.

When you take consecutive samples from this type of process, your control limits are based on 100% measurement variation. Therefore, any shift in the process will appear out-of-control when, in fact, it is not. The best way to handle this type of process is to use an IMR chart and space your samples apart enough that you are outside the period of autocorrelation.

You are right on. It's a plastic extrusion process. Sadly enough, there is no I/MR chart option to choose from.

Are you locked into a subgroup size of 5? If you can go to a subroup size of 2, try taking one measurement then waiting beyond the period of autocorrelation to take the second measurement for your subgroup. Then your control limits will include some short term variation in addition to measurement error. Add a little more time before starting the next subgroup.

Are you locked into a subgroup size of 5? If you can go to a subroup size of 2, try taking one measurement then waiting beyond the period of autocorrelation to take the second measurement for your subgroup. Then your control limits will include some short term variation in addition to measurement error. Add a little more time before starting the next subgroup.

I cant answer you right now but will definitely check it out and try what you have recommended. Thanks.

Getting the raw data is impossible.
That is a problem...and one reason that I am opposed to Cpk. It's the raw data that helps us truly understand. Cpk is so summarized that we can't determine what's wrong when it's bad and there's no way from that single number to even know if it's real when it's good...the challenge at this point is why are you asking for Cpk information? and is this truly meeting your intent?

estimated sigma is used where sigma=Rbar/d2.
Without the data this is just an educated guess but here goes: It's the formula. Rbar/d2 is the within subgroup standard deviation calculation. It ignores the between subgroup variation. The range chart shows small ranges within subgroup but you have large variation between subgroups - Cpk calculated using within sample standard deviation 'effectively takes out' the between subgroup variation. It essentially centers the individual subgroups on the grand average. (Cp effectively centers the subgroups on the target.) That is most likey why the Cpk value is high but there are many individual values out of spec.

It's just a very small portion that is out of control and that does not have a big enough impact to cause cpk to be below 1.33 or acceptable level. Correct me if i'm wrong.
understand that being out of control does not relate to individual pieces being out of spec - these two things are not related. You can be completely out of spec and in control and in out of control and well within spec.

Cpk calculations are not based on control. A statistically controllled process is required for the Cpk value to have any predictive value...

Are you locked into a subgroup size of 5? If you can go to a subroup size of 2, try taking one measurement then waiting beyond the period of autocorrelation to take the second measurement for your subgroup. Then your control limits will include some short term variation in addition to measurement error. Add a little more time before starting the next subgroup.

Yes, subgroup size can be changed. but I also found this on the manual.

There are some who would argue that for continuous operations such as wire, cable, tubing etc., a subgroup of one, and a "moving range" should be used. This comes from the fact that when the product size is measured manually, then an entire subgroup is taken from a short length at the end of the reel, coil etc. The problem with that is as stated above. The samples measured from a short length at the end of the package do not accurately represent the common cause variation of the process. The samples will, most likely, show very little range. The standard deviation will be too small and the Control Limits too high.

That is a problem...and one reason that I am opposed to Cpk. It's the raw data that helps us truly understand. Cpk is so summarized that we can't determine what's wrong when it's bad and there's no way from that single number to even know if it's real when it's good...the challenge at this point is why are you asking for Cpk information? and is this truly meeting your intent?


Without the data this is just an educated guess but here goes: It's the formula. Rbar/d2 is the within subgroup standard deviation calculation. It ignores the between subgroup variation. The range chart shows small ranges within subgroup but you have large variation between subgroups - Cpk calculated using within sample standard deviation 'effectively takes out' the between subgroup variation. It essentially centers the individual subgroups on the grand average. (Cp effectively centers the subgroups on the target.) That is most likey why the Cpk value is high but there are many individual values out of spec.


understand that being out of control does not relate to individual pieces being out of spec - these two things are not related. You can be completely out of spec and in control and in out of control and well within spec.

Cpk calculations are not based on control. A statistically controllled process is required for the Cpk value to have any predictive value...

I see what you mean. I should take ppk into account for qualifying the product. Thank you.

Yes, subgroup size can be changed. but I also found this on the manual.

There are some who would argue that for continuous operations such as wire, cable, tubing etc., a subgroup of one, and a "moving range" should be used. This comes from the fact that when the product size is measured manually, then an entire subgroup is taken from a short length at the end of the reel, coil etc. The problem with that is as stated above. The samples measured from a short length at the end of the package do not accurately represent the common cause variation of the process. The samples will, most likely, show very little range. The standard deviation will be too small and the Control Limits too high.

The standard deviation will be too small and the Control Limits too high.

The last word should read "tight". The smaller the variation, the tighter the limits.

The first sentence begins as though the writer disagrees with the IMR concept, but then the writer proceeds to make the same argument that I did. Did the writer advise against IMR?

For what it's worth, I have actually used IMR on a rubber extrusion process. We started with Xbar/R and saw the exact same issues that you see. We learned about autocorellation and changed to IMR with samples taken 20 minutes apart, and SPC suddenly gained credibility with the work force.

The standard deviation will be too small and the Control Limits too high.

The last word should read "tight". The smaller the variation, the tighter the limits.

The first sentence begins as though the writer disagrees with the IMR concept, but then the writer process to make the same argument that I did. Did the writer advise against IMR?

For what it's worth, I have actually used IMR on a rubber extrusion process. We started with Xbar/R and saw the exact same issues that you see. We learned about autocorellation and changed to IMR with samples taken 20 minutes apart, and SPC suddenly gained credibility with the work force.

Thank you for the valuable information. I'd give it a shot @ 2 subgroups with longer sample intervals. :thanx:

fEArmE,

Attached is a little simulation done on a couple of random generated distributions.
Note that ALL parts are out of spec, yet Cpk is reported at over 9. :mg:Might seem a little extreme, but without a histogram, how would you know?

Reinforces that you have to be very carefull with this metric. As others have suggested:

Always plot the data. Look for out of control situations.
Don't rely on a single metric. Note Ppk in this example tells a truer story.

fEArmE,

Attached is a little simulation done on a couple of random generated distributions.
Note that ALL parts are out of spec, yet Cpk is reported at over 9. :mg:Might seem a little extreme, but without a histogram, how would you know?

Reinforces that you have to be very carefull with this metric. As others have suggested:

Always plot the data. Look for out of control situations.
Don't rely on a single metric. Note Ppk in this example tells a truer story.

Wow, that's extremely helpful!! I'm gonna show this to my supplier. :thanx:

fEArmE,

Attached is a little simulation done on a couple of random generated distributions.
Note that ALL parts are out of spec, yet Cpk is reported at over 9. :mg:Might seem a little extreme, but without a histogram, how would you know?

Reinforces that you have to be very carefull with this metric. As others have suggested:

Always plot the data. Look for out of control situations.
Don't rely on a single metric. Note Ppk in this example tells a truer story.

yes! and an aptly named attachement I must say...