My customer wants me to control some products characteristics at process, such as seam width and stitches per inch. We started using a X-R chart, but as we improved our process, our process is within calculated control limits. But if I tried to recalculate the control limits, it become too narrow. Somebody told me about setting "fixed control limits", does somebody knows how to calculate "fixed control limits" ??

I haven't heard of "fixed control limits", but please do recheck how you are doing your calculation of the new limits. By "too narrow", do you mean that you end up with a few outliers? Too narrow as compared to specifications (which this is good)? Or too narrow as in a lot of points out of the control limits (which probably indicates a calculation error)?

It's too narrow as in a lot of points out of the control limits. I will have to rverify the calculation. We do monthly SPC chart, when we have the first month complete, we calculate the control limit and we use it for the next month/period. Due to all our measurements are (can't say equal) but close to the customer specification, the control limits are too "narrow" and then several measurements get out of the limits. The question now is, when do I have to stop calculating the limits?

I am attaching a paper that describes 'Economic' control limits that may be used when the process capability is very good, control limits are too narrow with respect to specs and it does not make sense to investigate every point that falls outside of control limits.
Tell us if this helps!

I too have never heard of "fixed" control limits; and fixed seems to be the antithesis of the purpose of statistical control limits.

Too many points outside the limits after recalculation suggests, first, what Steve says - a miscalculation of the limits. Second, if the calculation is correct, it suggests a recent out of control condition (as yet undiscovered).

You should not recalculate limits unless the latest data suggests you must, e.g. a mean shift, most points close to average line, or points spread beyond limits. In each case, however, you should know WHY the data is changing. And if it is NOT changing, don't change the limits.

And if the data shows CLOSE TO SPEC measurements, you definitely should be looking for a root cause and eliminating it.

Thanks to all of you guys.

Jim: two things:

'Fixed' control limits might mean teh opposite of constantly recalculating the limits...or recalculating on a set frequency...as pointed out above, limits shoudl be claculted once a stable process is obtained and not recalculated (i.e. fixed) until a known chagne has occured...usually for the better as a result of deliberate improvement activity.

Secondly, when many averages are out of the control limits, the first thing I suspect is that the subgroup is not rational...in other words the variation within the subgroup is smaller than the varitaion between subgroups. (Grant & Leavenworth present a reasonable explanation of this all too common subgrouping error). A small average range will give you a small band between the UCL and LCL for the averages...If your averages were in control in the past and are now out of control - after recalculating the limits - take a look at your ranges. Are the current ranges significantly less than they were historically? If so, your largest componenet of variation has changed since the original charts and you might need to change your subgroup selection plan (this isn't necessarily bad. The largest component of variation can change because the old largest family was significantly reduced...)

It's too narrow as in a lot of points out of the control limits. I will have to rverify the calculation. We do monthly SPC chart, when we have the first month complete, we calculate the control limit and we use it for the next month/period. Due to all our measurements are (can't say equal) but close to the customer specification, the control limits are too "narrow" and then several measurements get out of the limits. The question now is, when do I have to stop calculating the limits?

Ah, that is the problem. In control charting, one should establish a baseline time interval (Dr. Shewhart suggested 25 sets of 4 if doing xBar) and LEAVE IT ALONE. Do not recalculate the baseline (and control limits) unless there is indication of a statistically significant change (ie points outside the control limits, 7 or 8 above average, etc). I would say you need to include several months of data in your baseline and control limits.

Please feel free to take a look at http://www.hanford.gov/safety/vpp/spc.htm

- Steve

A few random thoughts to add to what Bev D suggests ...

Do you have a set of results for each sewing line (process) or are several lines (processes) grouped together? If practical, each line should be examined separately because each of the subsidiary processes has to be both capable and in a state of statisical control before the whole production can be.
Have you identified outliers, identified and corrected the causes of each, and if appropriate removed those values from future calculations?
Are there any autocorrelation effects to consider?

Graeme


Jim: two things:

'Fixed' control limits might mean teh opposite of constantly recalculating the limits...or recalculating on a set frequency...as pointed out above, limits shoudl be claculted once a stable process is obtained and not recalculated (i.e. fixed) until a known chagne has occured...usually for the better as a result of deliberate improvement activity.

Secondly, when many averages are out of the control limits, the first thing I suspect is that the subgroup is not rational...in other words the variation within the subgroup is smaller than the varitaion between subgroups. (Grant & Leavenworth present a reasonable explanation of this all too common subgrouping error). A small average range will give you a small band between the UCL and LCL for the averages...If your averages were in control in the past and are now out of control - after recalculating the limits - take a look at your ranges. Are the current ranges significantly less than they were historically? If so, your largest componenet of variation has changed since the original charts and you might need to change your subgroup selection plan (this isn't necessarily bad. The largest component of variation can change because the old largest family was significantly reduced...)

This has come up in previous threads. .

IMHO, if your process has been improved to the point that what you are monitoring does not make sense, and you are required to monitor it anyway, I'd suggest using X and MR charts and preset your control limits at a % of BP tol. 75% = 1.33 Cpk, 50% = 2.00.

This is a normal technique used after a process has been improved (Minimized normal variation, and removal of assignable causes or sources of special variation) and has demonstrated that it is stable and capable.

This will minimize the required time and effort in monitoring. You have a process that has proved it is stable, capable and effective. Now monitor it at reasonable intervals to demonatrate conformity to specification over time.

If the process deteriorates, go back to the X-Bar/R charts until you are back in shape again.

With the X/MR data, you can still generate Histograms for reporting and monitoring.

We have used "Fixed", "80% of tol.", "Pre" (call it whatever) control limits with most of our data acquisition gage packages. With multiple cavities or fixtures, we first "prove out" the capability of each cavity or fixture and then set limits based on trhe "best" and "worst" performing. In recent history, we have added computers to a couple of these gages to be able to ferret out cavity and/or fixture and utilize the actual process limits. The "Fixed" limits aren't much for problem solving but more of a "down & dirty" feedback mechanism to the operator to let him know whether to keep running or that something may need to be looked at.

Before I get bashed, we are not strong with SPC on our floor. My fellow QEs and I have been "hammering away" for the last 10 years and we are just beginning to see some understanding from the production dept. that the process "dictates to you" and not the other way around. Hopefully we will become much stronger with SPC in the future.

Bill

Hey Jim, your problem happen to me ones, Is your process a continuos process (chemical batch for example) ?, is your sample size=1?

Don Wheeler show in his bookAdvanced Topics in Statistical Process Control that you can use the autocorrelation coefficient:

Sigma = R_Average / (d2 * sqrt(1- r^2))

being r the correlation coefficient pairing the first measure with the second ones, and so on

It worked for me, you can post a set of data!.

:magic: