How do you make a design robust?
Taguchi considers making a design robust in the parameter design portion of product or process design. In parameter design the goal is to find values for controllable settings that minimize the negative effects of the uncontrollable settings. Experiments are used to determine the impact of particular settings on both the controllable and uncontrollable factors. The idea here is that by observing changes in a controllable factor (such as the thickness of boards), a value can be found for that factor that reduces the effect (warping) of something that can’t be controlled (the humidity outside). The ultimate goal is to find the optimal settings to minimize cost by minimizing variation.
When setting up these experiments, the factors that effect the product need to be determined. Then the factors can be separated into controllable factors and uncontrollable factors and experiments can be set up to test the effects of changing the values of each factor. There are many ways to set up these experiments. Taguchi’s method involves finding correlation between variables. He uses orthogonal arrays, with the inner array consisting of control factors and the outer array consisting of "noise" factors. Each inner array is to be run with each outer array. (If six control factor experiments and three "noise" factor experiments are needed, there will have to be (six times three) eighteen experimental trials to get all the combinations). Another method for conducting these experiments is to make no attempt to control the "noise" factors, but repeatedly run the trials for combinations of control factors. (Lochner and Matar, 152) This type of experiment allows the operator to measure process variability. The trials should be taken in an environment similar to the one in which the actual use or manufacturing of the product is going to take place. A third experimental design is to identify all the control and "noise" factors (adding the control and noise factors yields k) and run an analysis using at least k +1 trials based on eight-run experiments. (You could use an eight run experiment for up to k=7, and a sixteen run experiment for up to k=15.) This will allow the interaction between variable to be seen running fewer tests than using Taguchi's method. Further instruction as to how to use this method is found in chapter four of "Designing for Quality" by Lochner and Matar.
The data found from the experimental trials is then analyzed. The analysis will depend on the method of experimentation. Plot the effect that the variables had on your variation and/or the correlation between factors. Using this data find settings for the controllable factors that are found to lower the variation caused by uncontrollable factors.
Then after the initial experiment trails are run and "optimal" settings are found confirmation experimentation is needed. By performing a series of replica experiments at the levels that were picked, we can see if the values achieved matched that of the values the model predicted. If there is disparity, there may be an interaction or noise that we didn’t see and thus our experiment must be redeveloped.