Chapter 15. Simplified Design of Experiments
What you will learn in this chapter is how to run a simplified Design of Experiments (DOE). The intent of a DOE is to optimize a process by finding the right settings for a set of key process input variables (KPIVs). This chapter applies to the Improve step in the DMAIC process. This chapter is primarily for those involved in manufacturing or process work.
Some Six Sigma practitioners feel that DOEs do not belong in a text for green belts, because DOEs are too complicated. Also, since there are whole books written just on this subject, these people don't believe that DOEs can properly be covered in a relatively few pages. These concerns may be valid for traditional DOEs, but the simplified DOE presented here has been used many times by green belts, with successful results.
First, some discussion on what a DOE entails. It is a controlled test of KPIVs, usually done right in the production environment using the actual production equipment. It attempts to measure all possible combinations of KPIVs, rather than taking a standard setup and modifying one variable at a time, one after the other. In this way the DOE attempts to find any interaction among variables and includes this interaction in identifying the optimum settings. Many dedicated DOE software programs attempt to predict the optimum settings even if they are in between the actual test settings.
That sounds good, right? It is, but here are some of the challenges:
It is difficult to identify a limited list of KPIVs to test. After all, if the process were all that well understood, you would not have to run the DOE!
It is difficult to keep in control the variables not being tested. These could include temperature, humidity, operator skill, etc.
A large number of test variables requires many trial iterations and setups. There are many reduced-iteration DOEs, but these all sacrifice statistical confidence.
One way a DOE can reduce the number of trials is to run with KPIV settings far outside normal ranges. The problem with this approach is that many processes become non-linear and any conclusions become suspect. Some processes can't even be run outside their normal settings because of process limitations.
The results of the DOE must then be tested under controlled conditions, since the real test of a process change is its ability to predict future results.
What to use as an output goal is not trivial. What if product variation is reduced, but so is product output? What if the process settings give an excellent product, but require more operator skill? Most software programs limit optimization to one output measurement.
Now that we have listed all the reasons a DOE may scare you, here is an effective way to run a simplified DOE that will minimize these difficulties and drive process improvement.
Note: The results of any DOE are usually not the key that drives the process improvement. Instead, it is the disciplined process of setting up and running the test that gives process insight. Observations made during the DOE often trigger process breakthroughs. Serendipity becomes dominant in this kind of test.