Every once in a while, I return to the classics.  I was rereading some Deming recently and was reminded of the funnel experiment.  (If you don’t know who W. E. Deming is, look him up as soon as you finish this blog.) Dr. Deming used the funnel experiment to demonstrate the effects of statistical variation in a stable system as well as the difference between common cause and special cause variation.  A common cause is an inherent feature/flaw of the system itself.  A special cause is something that happens occasionally that is unplanned and unexpected.  The system in this case is a funnel fixed at a set height above a level surface through which marbles are dropped.  The variation comes from the fact that while the marbles hit very close to the same spot each time, they come to rest a short distance away from the target.  To create an analogy with a machining operation, you can consider zero distance from the target as the nominal and the range of distances (and directions) from the target as the expected process output variation.  I will remind you that when specifications are applied, they are an engineering construct and are not directly linked to the variation.  This means that the specifications we choose may or may not be contained within the expected variation.  Capability studies are used when we want to explore the process performance/variation as compared to the required specifications/tolerance.  This is important because a failure to create a part within tolerance is sometimes interpreted as an unstable system or special cause and this leads to an incorrect adjustment, even when the variation is within expected parameters for a stable system.  This is what the funnel experiment explores.

Deming uses 4 “rules” of adjustment to make his point.  Rule 1 is no adjustment.  The 50 marbles are dropped through the funnel in the initial fixed position.  This creates a roughly circular distribution centered around the target. Rule 2 says to move the funnel after each drop to compensate for the result.  If the marble comes to rest 20 cm east of the target, then the funnel is moved 20 cm west of the current position before the next drop.  This also creates a roughly circular distribution, but the diameter of the circle is about 40% larger (i.e. greater output variation).  Rule 3 says to again move the funnel after each drop, but this time move that distance from the target instead of the previous funnel location.  This results in two distributions on opposite sides of the target.  Rule 4 says to move the funnel over the resting location of the last marble each time.  This results in a smear that moves further and further away from the target.  Bottom line:  The best results, as defined by the least variation, are obtained when the funnel isn’t adjusted at all.  This means that chasing “better” results by adjusting (he called it tampering, operators call it tweaking) with no understanding or acknowledgement of variation gives you exactly the opposite of your desired outcome.

In Part 2, I will describe how this isn’t just a problem when machining a part.  The same concept applies to any system where you can measure outputs.

Don’t Adjust,