|
The fastest, cheapest, least-risky and best way to learn about governing parametric relationships during technology, product and manufacturing process development is to use a sequential strategy for applying Designed Experiments (DOE). This article will help you understand your options to design and conduct various types of experiments.
Inferior approaches such as trial and error tests and one-factor-at-a-time experiments just cannot compare to the efficiency and integrity of the results from a well-reasoned series of designed experiments. If you do not use orthogonal (balanced) arrays, or properly justified nearly-orthogonal arrays, as the underlying architecture for running designed experiments, you risk not being able to detect (1) signal from noise, (2) independent and interactive relationships, and (3) linear or non-linear relationships from your experimental data.
Below is a list of the state-of-the-art in DOE options that are available to structure a sequential learning path:
Basic Screening Experiments
Full Factorial Designs
2 level designs
3 level designs
Multi-level designs
Fractional Factorial Designs
2 level designs
3 level designs
Multi-level designs
Plackett-Burman Designs
Taguchi Style Designs for Robust Design
Static & Dynamic Robustness Optimization experiments
Definitive Screening Experiments
Response Surface Methods for Modeling Linear & Non-Linear functional performance
Various DOE structures for use in Computer Simulations (DACE)
Mixture Experiments (Chemical)
Multi-variant Designs
Conjoint Studies
These DOE architectures can result in the following during development:
- Identification of basic factors or engineered parameters (Xs) that are statistically significant and practically significant for controlling a response variable (Y). This is usually referred to as Screening for main effects and interactions of the Xs. These results are almost always conducted under tightly controlled experimental conditions where all sources of assignable cause variation (non-random noise) are minimized or included as a Block Factor to isolate their effects.
- Generation of a linear or non-linear math model for all statistically significant main effects and interactions. This case creates the actual governing equations including the proper model coefficients. These equations can then be compared to pre-existing 1st principles math models for correlation, adjustment of the analytical model (assuming the empirical model is the correct one) and confirmation.
- Optimization of the mean and/or the standard deviation of one or more dependent output variables (Ys) under nominal, non-stressful conditions where all sources of assignable cause variation are minimized or Blocked (properly isolated) within the experiment.
- Identification and optimization of controllable factors and designed parameters that make the function of a design robust to stressful, assignable (non-random) causes of variation. Here we co-vary both noise and control factors to study their interactions. If the interactivity is strong, then we can exploit them to set the control factors to set points that leave the function of the design minimally sensitive to the active noise factors.
- Balancing and optimization of tolerances. This can be done under both nominal and stressful conditions.
Using Designed Experiments enables the sequential learning cycles that must be undertaken as one develops material, energy inputs and logic and control inputs for control and robustness of subassemblies. Next, the DOE options can be used to repeat the learning cycles on the development of subsystems. Finally, this strategy is used to integrate the system to balance and optimize system performance. This learning strategy works best when the use of DOE is sequentially applied up the ladder of system development hierarchy from subassemblies to subsystems and finally the system. Nominal and tolerance set points are developed to control system performance that is both robust and tunable in light of customer and stakeholder requirements.
The bottom line is that deep thinking between physical law and statistical analysis will drive the design of a reasonable, efficient and effective series of DOEs to close your knowledge gaps during research, technology and product/process development.
| 
