DOE continuous improvement of a released manufacturing process

wendytseng

I have a customer who is interested in learning more about approaches to use for experimentation on a released manufacturing line with the least amount of impact to the line … either in terms of execution logistics or risk of unintended consequences (hitting a local hot spot for scrap or something like that).

One approach would just be to run something like a screening design or RSM within the qualified range, but he has seen studies where a triangle is used rather than orthogonal settings (at least that is his interpretation).    In those examples, the process is continually iterated until the maximum is found by augmenting the original triangle space with something like a “star point” off of the leg of the triangle between the two best process points.  

Has anyone heard of these types of designs?  What are ways he can consider experimenting in a released manufacturing process setting efficiently?

Dan_Obermiller

It sounds like Sequential Simplex Optimization to me. For two factors, the simplex is a triangle. Of the three vertices, the worst vertex is "reflected" around the line formed by the other two vertices. The process continues and the idea is to slowly iterate to the optimum area. The step sizes are usually much smaller than a designed experiment because the goal is not a model. Instead, the idea is to continue to make viable product while slowly getting better over time.  The process is slow because you often stay in one spot for awhile to ensure which vertices are the best. 

There is a book called Sequential Simplex Optimization by Stan Deming, Walters, Morgan, and Parker. 

I worked with Mark Bailey to write a script to do this when I was at Dow Chemical back in 2000 or so. Unfortunately, it was so long ago that even if I had the script, it likely would not work.

For a different view of this type of work you could also consider Evolutionary Operations by Box and Draper (which is an older technology). That is based more on designed experiments as it uses small full factorial designs. The idea is the same that the designs are small in size to ensure viable product is made, but you "reflect" the design towards a better response and continue to improve over time. Same idea, but a different approach that requires more data than the Simplex approach. This approach does a better job of allowing a model to be developed to describe the process.