Trying to determine what type of control chart to apply...

We two CNC Lathes that run small batches. These batches may be of different material and configurations which require setups between runs. There is a lot of variation between runs (due to setups) and very little variation within runs (variation in lathe).

I would appreciate a recommendation. If you successfully implemented SPC in this circumstance an example would also be appreciated.

We had a similar situation and used some of the "short-run" SPC techniques to help us. It requires, among other things, to know the sigma of each "within set-up" run. By normalizing the "gross sigma" for the process, you can have one set of modified control limits for each NC machine. This technique allows you to focus on the machine and not just the part that you are fabricating. We used this technique years back and it worked. You do lose some resolution on the control limits, but we found that the gains outweighed the loss due to resolution. Anyway, check out some books on this topic in the ASQ Publications catalogue.

Trying to determine what type of control chart to apply

I am not sure I can help here. Control chart applications are very dependent on usage. It sounds as if your chart is varying across setups, but I need more details.

There are some alternatives to the traditional control chart in this type (small run) which may be of help. Under the assumption that each run is different, you may want to look at a pre-control type of technique.

Perhaps I could be more help if additional details were provided. Or, perhaps someone else has had this experience. Personally, I have found control charts for short runs rather, in a word, lacking.

Regards,
Don

------------------
Just the ramblings of an Old Wizard Warrior.

Are there any statistical formulas that relate precontrol data to process capability indices?

dmac - I am not sure this is what you are looking for but here it is anyway for what it is worth.

To establish a pre-control chart, it must be assumed that the process is normally distributed with a Cpk of at least 1.0. It can then be statistically shown that 86% of the product will be within the control lines with 7% in each of the other sections.

As there is no statistical data taken or analyzed, it is important to only use pre-control on a process known to be stable.

I am sure there are others who can add a lot to this but at least it may help to get you started.

Dave

1- Check normality by using a Histogram of your data.

2- You can have a ROUGH estimate of process capability by taking standard deviation of your data and calculating Ppk.

3- I don't know to what extent you can rely on the above calculation, because using a Pre-control chart you cannot be sure your process is under control (a prerequisit to determining capability)

Any other ideas?

MSAFAI

Seeking any other details on Pre-Control. Found this here FYI:

http://Elsmar.com/Forums/showthread.php?t=5156

Steve: We two CNC Lathes that run small batches. These batches may be of different material and configurations which require setups between runs. There is a lot of variation between runs (due to setups) and very little variation within runs (variation in lathe).

I have a similar problem, but unable to use any SPC that I know of. We take a batch of material, mold it, cure it in an oven under clamped pressure, grind it, scorch it, then take 5 pcs/batch (batch = approx 3000 parts) and do various physical tests on it. SPC is not done on the process, but the end result of these operations.
So for example, the density of the part is measured per batch as a subgroup. This charateristic is affected by every operation above beginning with the material batch, potentially processed on 1 of 3 presses, through 1 of 6 ovens, 4 clamping stations,etc. It is not possible to keep these separate and plot individually. Each control chart subgroup is from a different batch of material and off a multitude of processes and equipment.

Limits are very tight from Rbar, which makes the charts look terrible from subgroup to subgroup variation, but still very controlled compared to the specifications. This is hard to explain to vistors that do not understand the complexity of the problem.

To me, it seems that these could be considered different setups, processes, or population just as you would expect a mean shift for a different lot of steel in a fineblank press.

When things appear to go out of control, there is not much you can do to react to the problem. There is no adjustment at this point in the process because the parts are already made. Because this happens very frequently and because of the huge number of influencing factors, you can't investigate everything and record the reaction as required for OUC conditions. Some real assignable causes can be determined after a week or two if your lucky, others are just ellusive despite DOE after DOE. I have began to think that this level of variation is normal, little more can be done to improve it, and that a lot of my problem is the use of control charts that shouldn't be applied to this situation. I think standard Xbar & R charts are inappropriate here. Anyone agree or am I out in left field?

Thanks,
Chemman

D.Scott said:

dmac - I am not sure this is what you are looking for but here it is anyway for what it is worth.

To establish a pre-control chart, it must be assumed that the process is normally distributed with a Cpk of at least 1.0. It can then be statistically shown that 86% of the product will be within the control lines with 7% in each of the other sections.

As there is no statistical data taken or analyzed, it is important to only use pre-control on a process known to be stable.

I am sure there are others who can add a lot to this but at least it may help to get you started.

Dave
So.... Under what circumstances does pre-control make sense?

D.Scott said:

dmac - I am not sure this is what you are looking for but here it is anyway for what it is worth.

To establish a pre-control chart, it must be assumed that the process is normally distributed with a Cpk of at least 1.0. It can then be statistically shown that 86% of the product will be within the control lines with 7% in each of the other sections.

As there is no statistical data taken or analyzed, it is important to only use pre-control on a process known to be stable.

I am sure there are others who can add a lot to this but at least it may help to get you started.

Dave

I thought I had read somewhere that precontrol does not assume normality, and is free of distribution shape.... I can't recollect the source, so I may just be wrong!

Offcourse if normality was not assumed there would be no estimates of the Alfa and Beta errors.

Has anyone else come across this reference of Precontrol being free of the normality assumption?

All the references I have for Pre-control assume a normal distribution. I would be interested in the reference Ravi mentions. If anyone comes across it, please be sure to post a link. I am never too old to learn something new.

Dave

I too thought normality was not required for pre-control to be effective, but can't find the reference. It may have been the write-up in Juran's Quality Handbook Version 4, which I no longer have. If someone has it perhaps they could give it a look.

If one has a non-normal process and wonders about pre-control's effectiveness, a good, old-fashioned experiment can be done. Use a traditional X-bar and R chart and whatever else you want to try, in parallel with a pre-control chart and see what happens! Let the data speak for itself! (But share the results if you try it!)

IMO, the confusion is caused because we are trying to understand / explain Pre-Control using the terminology of conventional SPC (which comes with its own meaning)

I see there are a few things that indicate that Pre-control does not insist on normality of data.
1. Control limits are not derived from data, but are based on Specification width.
2. Pre-Control does NOT calculate Process Capability/Cp/Cpk etc.
3. Pre-control does not require as many samples/subgroups as SPC to characterize the process.

Most (All?) of Shainin's techniques are shrouded in a kind of mystery because no theoretical explanations are readily available. Exponents of pre-control say - here is a great, math-free, visual tool to monitor your process, this is how you do it, don't bother about the underlying statistics. Many of us cannot digest this argument. The one way to prove/disprove pre-control is, as Mike suggests, to try it and compare with conventional SPC results.