Statistical process control has been widely used in the manufacturing companies. However, with the need to reduce product to market time, the concept of quality has moved to the design stage of the product. In the design stage, the manufacturing imperfections for the product are taken as tolerances for the design parameters and assumed to follow some probabilistic distributions such as the normal or uniform distributions. With these variations in design parameters, the corresponding design responses will also change. When matched against the design specifications, a design can be classified as to have passed, that is conformed to the specifications, or failed otherwise. This process of finding the optimum nominal values to ensure that the behaviour of the design satisfies the given specifications with the greatest probability is called design centering.
If a design is represented in the parameter space with X = (x1,
x2, , , , xn) and have a joint probability
distribution for the design parameters f(X), the yield is
then defined as the n-dimensional integral
Design centering, part of the parameter design stage in Taguchi’s approach to robust design, can primarily approached through approximating the geometric constraints, modeling the yield function and through the use of heuristics methods. Ellipsoidal and simplicial based techniques are usually use in the modeling of design constraints. On the other hand, for heuristics methods, center of gravity is one of the popular approaches. However, most geometric modeling algorithms requires the assumption of convexity of the constraints regions and is usually computing intensive. On the other hand, the CoG algorithm is not able to reach the high yield region effectively and often get stuck in local optimum.
My strategies for design centering is through the use of a new heuristics based algorithm called momentum based CoG methodology and using neural networks as response modeler. More details of my work will be posted as soon as I submit my thesis.