August 2006

"Quality Focus" - monthly Genesis column

Experiments Solve Can Line Problems

There are so many variables that affect the performance of a line that it is necessary to work out which are the most influential on the quality of the can...

This month we yet again continue our discussion of the “can line equation,” but now with a focus upon actually solving production problems. 

Even at the risk of repeating myself ad nauseam, let us simply begin by re-identifying can manufacturing as a process.  In fact, most physical phenomena may be proven to follow some process.  And, as I have stated (too?) many times before, Statistics is the only way to control a process, especially in the world of mass production.

So what statistical tools are available to us to solve can production problems?  First and foremost is the ability to measure the problem.  For reference, go back and reread the April 2006 column.  Measurement is by far the most difficult part of a statistical problem solving strategy, as many production defects will require the measurement system to capture defect intensity.

A validated measurement opens the door to effective and permanent defect resolution, enabling us to gravitate away from the domain of simple Arithmetic – I refer the reader to the February 2006 column.

Traditional root cause analysis tends to focus upon one factor at a time (OFAT) in an attempt to discover what is causing the problem.  Unfortunately, OFAT takes a long time, is not very effective, and most importantly it misses the interactions of possible root causes.

Enter then the concept of Designed Experiments, or Design of Experiments.  DOE allows for the manipulation of many possible root cause factors over a much shorter time and identifies the root cause interactions.  There are several experimental designs that allow for, say only “half” the data to be collected, with results that may be statistically shown valid within a specified amount of “confidence.”

An example – referencing the July 2006 column – relating to basecoated cans.  Were we having a problem, say, with variance in the opacity on the containers, we might incorporate several possible root cause suspects into our DOE, namely coating viscosity, coating temperature, mandrel variation (for two-piece cans) and applicator roller variation.

A DOE in this case would “run” the experiment with each of these factors at two levels at a minimum – however, sometimes we need to use more levels.  It would not be necessary, nor practically feasible, to experiment with every possible combination; this would likely take longer than we wish.  By running a “fractional design” in, well, a fraction of the time, we could ascertain which factors were truly affecting the opacity plus which factors interact with each other.  Interaction is often the most important information we can glean from a DOE, as it usually leaves us with more solutions to the problem than traditional experimentation.

For example, we may find an interaction between coating viscosity and applicator roller variation.  Perhaps we are not in a position – financial or otherwise – to adjust the coating viscosity, but we can easily address the roller problem.  The DOE might show us that we can run with the viscosity with which we are dealt, provided we hold the applicator roller within some reasonable bounds.  (Keep in mind this is only a fabricated example.)

Or we might find an interaction between mandrel variation and coating temperature that allows us to run with highly variable mandrels, provided we keep the coating temperature at (or near) some discovered level.

The point is that a properly executed DOE in a can plant will almost always point to the exact root cause(s) and the interactions between input factors.  More importantly, it will pave the way toward flexible solutions that economically eliminate the problem(s) once for all.