Taguchi S/N Ratios and Direct Robustness Measurement for Computational Robust Design

Robust design methods are becoming widely adopted in automotive industries because of their cost- and time-efficiency for quality improvement in a vehicle development process. These methods were pioneered by Dr. Genichi Taguchi of Japan and introduced to the U.S. in 80’s. Typical Taguchi-class robust design approaches apply S/N ratios of various types to measure the robustness performance of the output responses of target systems. Next designed experiments are applied to identify key control factors and interactions to maximize the selected S/N ratio in order to make the systems insensitive to pre-selected noise factors. Commonly used Taguchi-class S/N ratios (e.g., Nominal-the-best and Dynamic types) are based on some underlined statistical assumptions: 1) linear proportionality between location effects and dispersion effects, 2) availability of adjustment factors in the chosen control factors, which is insignificant on the system robustness but significant on mean output, and 3) a single S/N ratio based on ideal function to evaluate the overall robustness of a target system. Because of the growing complexity of modern engineering systems, the assumptions above may not be valid for major robust design applications, especially those through computational simulations. The author would like to illustrate some important assumptions behind Taguchi S/N ratios and suggest the following three recommendations for the usage of robustness measurement for computational robust design: 1) direct measurement for location and dispersion effects instead of a single relative measurement like Taguchi S/N ratio, 2) engineering (equality or inequality) constraints to meet engineering targets and requirements, and variable design specs to satisfy numerous specifications of complicated systems, and 3) multiple robustness measurements for the responses of the target system instead of one summary S/N ratio for the overall robustness of a multi-output system. These three recommendations will make robust design more adaptable, efficient, and less sensitive to assumptions for complicated computational applications than Taguchi S/N ratios.