Informational Control Chart Interpretation - General "Rules"

bobdoering said:

For Shewhart charts, and those subsets of processes whose distributions apply, yes. But, that is not always applicable, and therefore not hard and fast rule. We have discussed those exceptions in detail elsewhere in the forum. Precision machining is a classic example of such an exception.

Click to expand...

Well, this thread is about Control Chart Interpretation, so it would seem to apply in this case. And I am not sure I am in agreement on precision machining being an exception, either. Being willing to learn, could you give me a link to the discussions of the exceptions? I don't seem to be having any luck finding them by searching.

Geoff Withnell said:

If you can describe to me ANY scenario of a process which is adjusted based on measurement, where the measurement variability does not end up as a portion of the process output variability, I will concede your point.

Click to expand...

As I stated, it is there, but it had better be insignificant. Making decisions based on data fraught with error puts you into the "garbage in, garbage out" scenario.

Geoff Withnell said:

There was some degree of variation inherent in the measurement, and this is part of the control loop of the process.

Click to expand...

Control decisions can be no better than the validity of the data. Making decisions based on measurement error holds little value at all.

Geoff Withnell said:

If you can describe to me ANY scenario of a process which is adjusted based on measurement, where the measurement variability does not end up as a portion of the process output variability, I will concede your point.

Click to expand...

It is perfectly acceptable for it to exist, as long as it is statistically insignificant compare to the actual transformation process variation. That is a fundamental basis of gage choice and control.

Geoff Withnell said:

Never said you should accept inadequate measuring systems.

Click to expand...

I agree, you are on the right track there.

Geoff Withnell said:

Just that, in the real world, production application of SPC, the measurement error is in the process, it's part of the process control loop, and it is in the process output.

Click to expand...

But an adequate gage system holds statistically insignificant error compared to the transformation process variation, or it is - by definition - inadequate, and misses the whole point of the measurement exercise.

Geoff Withnell said:

Well, this thread is about Control Chart Interpretation, so it would seem to apply in this case. And I am not sure I am in agreement on precision machining being an exception, either. Being willing to learn, could you give me a link to the discussions of the exceptions? I don't seem to be having any luck finding them by searching.

Click to expand...

You can start here.

Ah, sudden flash of insight. Gage error is, at least in my experience, seldom statistically insignificant. Meaning if I do a chi square test on the variance of the total process with and without gage error, I will find that the variances are in fact different, I will be able to reject the null hypothesis of no difference in the variances. However, in a well designed system, the difference will be practically insignificant. meaning that the measurement error will not lead to making a different decision then what would be made could it be eliminated. Take the frequent case of a lathe operator running to "shop tolerances" measuring a diameter with a mic. is the measurement error statistically significant? Almost certainly. Is it practically insignificant? Again, assuming a well trained operator, almost certainly. And one of the reasons for running the chart is to detect if this changes.

bobdoering said:

You can start here.

Click to expand...

Well, if one is running a process sufficiently well controlled that tool wear is your main source of variation, and you can clearly identify the sawtooth trend pattern, you are almost certainly using sophisticated NC processes. Why haven't you built adjustment for the sawtooth pattern into your program, eliminating that assignable cause of variation? I've seen it done even without NC, with a centerless grinder operator adjusting for wheel wear every X pieces, and resetting after each wheel dress. And then you are back to a normal distribution, with a smaller spread.

Geoff Withnell said:

Well, if one is running a process sufficiently well controlled that tool wear is your main source of variation, and you can clearly identify the sawtooth trend pattern, you are almost certainly using sophisticated NC processes. Why haven't you built adjustment for the sawtooth pattern into your program, eliminating that assignable cause of variation? I've seen it done even without NC, with a centerless grinder operator adjusting for wheel wear every X pieces, and resetting after each wheel dress. And then you are back to a normal distribution, with a smaller spread.

Click to expand...

Using charting allows for identification of special causes (tool breakage), tool wear rate evaluation, etc. There is no guarantee that it is being used on a sophisticated NC process, and the processes where operators make adjustments and tool changes, this technique assures minimal overcontrol and also provides leading indicator of tool failure - so the data tells you when to change tools, instead of guessing. It's like driving a car...you don't need a gas gage, but it helps if you have one. This charting is the "gas gage" in that example.

If you see a normal distribution, either your sampling is too frequent, your gage is the variation, or your operator has become the process my adjusting too much. If you do not see a sawtooth, you are out of control. It is pretty straight forward.

Hi Bob,

I have reviewed the data you entered into the Hi/Lo chart and to me the process looks very random for that feature.

I did feel that Dimensions 18,21 and the Sphere diameter all showed a clear trend upward over time, however I am concerned by some things as well...

Between parts 6 and 7 there is a clear shift in several of the features between setups. I am told that the process should repeat and it does not seem to.

I don't feel that the roughing tool breaking at 0.0013" range for dimension #1 gives me very clear definition of a problem since there seems to be a fair amount of variation in the range with the new tool - going as high as 0.0010 and back down to 0.0007

Geoff,

Can you please further explain with some detail the process of the Chi Square test, and how this relates to the null hypothesis? I assume this is in some way connected to a normality test - perhaps different than the standard AD test for normality.

Thanks for all of the input - I appreciate seeing the different perspectives to this approach!

Coleman Donnelly said:

I have reviewed the data you entered into the Hi/Lo chart and to me the process looks very random for that feature.

Click to expand...

Could be the sampling is too frequent, yet not over long enough time for the actual tool wear to show up. It is like the effect of roughness and waviness. You are looking for "waviness", so the "roughness" is likley just noise.

So when a tool breaks after 9 pcs, how can I reduce inspection frequency?
I am checking 100% at this point trying to identify trends, I just don't think I am comfortable reducing inspection at this point.

Coleman Donnelly said:

So when a tool breaks after 9 pcs, how can I reduce inspection frequency?
I am checking 100% at this point trying to identify trends, I just don't think I am comfortable reducing inspection at this point.

Click to expand...

Good point. The data did not indicate that, which is why control chart notes are critical.

Next point is why? You mentioned working with tough material. Are you taking off the recommended amount for that tool? Speeds and feeds appropriate? 9 pieces is pretty much warm-up for most machine/material combinations - which is special cause, or not at steady state and stable. That would explain to some degree the variability you are seeing - warmup variation. It can vary in just about any manner.

On the other hand, if the tool breaks at 9 pieces, your inspection rate is pretty much set. Sampling has little to offer.