The difference between Classical Taguchi Methods, DOE, and RSM is elaborated as below.
Taguchi Methods: To estimate an optimum setting for a product/process, in terms of its control factors and their design levels. e.g., if y= f(A1/2/3, B1/2/3, C1/2/3), identify the combination of A, B & C at their levels 1/2/3 for an optimum y The objective is always to maximize the S/N ratio of the response. Fully saturated, fractional factorial designs. The information space for studying interaction effects is utilized on studying additional factors within the same set of experimental runs. Dr. Taguchi advocates the use of multi-levels (3, 4 or 5 levels) to identify the curvature in the response and identify the near-optimum settings of the factors. Factors could be both continuous or discrete. Using the ‘Average Effect Plots’ the factors are set at their respective levels where the average-effect of the S/N ratio is maximum.
Classical DOE: To develop a linear (1st order) function to describe a product/process performance and use the equation to set the product/process at its optimum setting. e.g., if y=f(x1,x2,x3), estimate the equation: y=a0 +a1x1 +a2x2 +a3x3 +a12x1x2 +a13x1x3 +a23x2x3 +a123x1x2x3 Based on the objective of design, y is set at minima/maxima or its nominal value. Full factorial designs. These plans incorporate information space for studying two factors and three-factor interactions. These are two-level designs for estimating the planar equation (1st order) for a product/process performance. Factors could be continuous or discrete. Using multiple linear regression, the mathematical model is developed. Using the equation the optimum setting for the product/process is arrived at.
RSM: To develop a quadratic (2nd order) function to describe a product/process performance and use the equation to set the product/process at its optimum setting. e.g., if y=f(x1,x2,x3), estimate the equation: y=a0 +a11x12 +a22x22 +a33x32 +a1x1 +a2x2 +a3x3 +a12x1x2 +a13x1x3 +a23x2x3 +a123x1x2x3. Based on the objective of design, y is set at minima/maxima or its nominal value. Factorial designs augmented by center points and axial points (Central Composite Designs – CCD) which offer the flexibility of sequential experimentation OR Box-Behnken Designs for directly estimating the 2nd order quadratic function for a product/process performance. In the CCD, the center point runs are required to detect the presence of any curvature of the response (indicating the start of the quadratic region) and the axial point runs are required for estimating the coefficients of the squared terms. Factors must be continuous variables! In the case of Central Composite Designs (CCDs): If the response could be described by a linear model, the “Method of Steepest Ascent/ Descent” is deployed. On detecting a presence of curvature, the quadratic model is developed by conducting the additional runs of axial points. Box & Behnken Designs: These designs are deployed when the product/process performance is known to follow a quadratic equation. The quadratic equation is directly estimated using the regression tool.
Summing up, when conducting experiments is an expensive affair, Taguchi Methods present a cost-effective alternative. They are easy to plan and the analysis involved requires simple computation. If one is interested in studying interactions, then the Classical DOE presents itself as an appropriate tool. If the engineer is not sure whether the product/process performance follows a linear equation or quadratic equation, he/she could use the CCD. If the engineer is confident that the product/process performance follows a quadratic equation, he/she could use the ‘Box & Behnken Design’ plan.
Recommended path:
Start the development work with many factors (short-listed after brain-storming)
Use Taguchi Methods as ‘Screening Experiments’ to reduce the list of potential factors and reduce the experimental effort.
Use Classical DOE (Full Factorial) to verify whether the product/process performance follows a linear function.
If Yes, deploy the method of steepest ascent/descent
If analysis indicates a presence of a curvature, conduct additional runs as per RSM
I want to do statistical optimization of the preatment of my raw material (citrus peel waste). I want to choose 4 factors with 3 or 4 levels and I read on articles that Taguchi method is more used in this case. Why ? and which DOE is better to my case?. Thank you so much.