Robust Design Experimentation and Dispersion Effects
One of the most common methods in robust design methodology is design of experiments. This method makes it possible to identify settings of design factors that make products and production processes robust with respect to noise factors. In practice, this may be done by searching for interactions between design factors and noise factors that can be used to diminish the effect of noise factors on important performance characteristics.
In the appended Papers D and E it is argued that a dispersion effect can be viewed as an interaction between a design factor, deliberately controlled in the experiment, and an uncontrolled factor. Provided that all potentially active design factors have been included in the experiment, it is reasonable to view the uncontrolled factors as noise factors. Papers D and E propose methods for identifying dispersion effects in unreplicated completely randomized two-level fractional factorials.
Split-plot experiments are well suited in robust design experimentation as they allow for a precise estimation of effects related to interactions between design factors and noise factors. Split-plot experiments are also cheaper to conduct than completely randomized experiments. However, traditional analyses of split-plot experiments have not made use of all the information that can be extracted from the experiments. The work reported in this thesis sheds some light on alternative methods of analysis that use information more efficiently than in the traditional methods of analysis. Specifically, Papers A and B propose methods for the analysis of split-plot experiments that make it possible to find opportunities for improving the robustness of products and processes. The suggested methods offer more subtle knowledge of how factors influence location and dispersion.
Paper C presents a decision tree that links different restrictions in randomization to their consequences on the experiment. This tree facilitates deliberate choices of restrictions - choices that have consequences that are favorable with respect to the cost and information content of the experiment.
Finally, Paper F presents the results of a telephone survey on the use and knowledge of robust design methodology in Swedish manufacturing industry.
design of experiments
generalized linear models