Could anybody explain the Taguchi Methods in layman’s terms, particularly with regards to how to choose the variables around which experiments are designed?
TIA.
Could anybody explain the Taguchi Methods in layman’s terms, particularly with regards to how to choose the variables around which experiments are designed?
TIA.
I’m not sure; I only know about the general concepts.
Taguchi is (basically) just a purely statistical approach to quality control. It is primarily used to reduce variation in manufactured components, but it can (theoretically) be used for anything, even TQM.
Yeah, I understand the basic concepts too, it is used to design experiments for product design (“robust design” is the jargon). How does one choose which variables to perform experiments on, though?
Having had formal experimental design training, I can answer that. The variables chosen require that one perform one or more “pilot” experiments. In these experiments, one piles on as many possible variables as one can imagine–yes anything that doesn’t sound outright silly is acceptable. Your only guidance is empirical experience and/or the current literature on the subject.
Then subject the results to analysis. One popular form is principal components analysis. Essentially, one tries to find out which variables “cover” the majority of change in the results. Then build a model that uses as few variables as possible but as many as necessary–how many is that? That’s what the next step is for. Try to use your model to predict outcomes of some test experiments. If your predictions fall within your arbitrarily chosen limits, then remove a variable. Keep repeating until you fall outside your arbitrarily chosen limits. Then go back to the last model that worked. This can be expensive and take a lot of expertise. The idea is to nail things down right and proper. Then you write up the “standard procedure” that gets used in general.
There is no generally accepted a priori method for pre-choosing variables in a completely new system. Fortunately, almost no experiments are truly “completely” new. There is usually something somewhat related that somebody has reported in the past.
How would one compute the “noise level” and what purpose does that serve?
:smack:
Thanks for your reply
:o
Anyone?