5.1. Introduction
5.1.2. |
What are the uses of DOE? |
A common use is planning an experiment to gather data to make a decision between two or more alternatives
Choosing Between Alternatives (Comparative Experiment)
- Choosing Between Alternatives
- Selecting the Key Factors Affecting a Response
- Response Surface Modeling to:
- Hit a Target
- Reduce Variability
- Maximize or Minimize a Response
- Making a Process Robust (i.e. the process gets the "right" results even though there are uncontrollable "noise" factors)
- Seek Multiple Goals
- Regression Modeling
Supplier A vs. supplier B ? Which new additive is the most effective? Is catalyst ‘x’ an improvement over the existing catalyst? These and countless other choices between alternatives are presented to us in a never-ending parade. Often, we have the choice made for us by outside factors over which we have no control. But in many cases we are also asked to make the choice. It helps if one has valid data to back up one’s decision.
The preferred solution is to agree on a measurement by which competing choices may be compared, generate a sample of data from each alternative, and compare average results. The ‘best’ average outcome will be our preference. We have performed a comparative experiment!
Sometimes this comparison is performed under one common set of conditions. This is a comparative study with a narrow scope - which is suitable for some initial comparisons of possible alternatives. Other comparison studies, intended to validate that one alternative is perferred over a broad range of conditions, will purposely and systematically vary the background conditions under which the primary comparison is made in order to reach a conclusion that will be proven valid over a large scope. We discuss experimnetal designs for each of these types of comparisons in Sections 5.3.3.1 and 5.3.3.2.
Selecting the few that matter from the many possible
Some reasons to model a process
Often we want to "fine tune" a process to consistently hit a target
Optimizing a process output is a common goal
Processes that are on target on the average, may still have too much
variability
The less a process or product is affected by external conditions
, the better it is - this is called "Robustness"
Sometimes we have multiple outputs and we have to compromise to achieve
desirable outcomes - DOE can help here
Regression models (Chapter 4 ) are
used to fit more precise models
Often there are many possible factors, some of which may be critical and others which may have little or no effect on a response. It may be desirable, as a goal by itself, to reduce the number of factors to a relatively small set ( 2-5) so that attention can be focussed on controlling those factors with appropriate specifications, control charts, etc.
Screening experiments are an efficient way, with a minimal number of
runs, of determining the important factors. They may also be used as a
first step when the ultimate goal is to model a response with a response
surface. We will discuss experimental designs for screening a large number
of factors in Sections 5.3.3.3, 5.3.3.4
and 5.3.3.5.
Response Surface Modeling a Process
Once one knows the primary variables (factors) that affect the responses of interset, a number of additional objectives may be pursued. These include:
These response surface modeling objectives will now be briefly expanded upon.
This is a frequently encountered goal for doing an experiment.
One might try out different settings until the desired target is ‘hit’ consistently. For example, a machine tool that has been recently overhauled may require some setup ‘tweaking’ before it runs on target. Such action is a small, and common form of experimentation. However, rather than experimenting in an ad hoc manner until we happen to find a setup that hits the target, one can fit a model estimated from a small experiment, and use this model to determine the necessary adjustments to hit the target.
More complex forms of experimentation, such as the determination of the correct chemical mix of a coating that will yield a desired refractive index for the dried coat (and simultaneosly achieve specifications for other attributes) may involve many ingredients and be very sensitive to small changes in the percentages in the mix. Fitting suitable models, based on sequentially planned experiments, may be the only way to efficiently achieve this goal of hitting targets on multiple responses simultaneosly.
Maximizing or Minimizing a Response
Many processes are being run at sub-optimal settings, some of them for
years - even though each factor has been optimized individually over time.
Finding settings that increase yield, or decrease the amount of scrap and
rework represent opportunities for substantial financial gain. Often, however,
one must experiment with multiple inputs to achieve a better output. Section
5.3.3.6 on second order designs, plus material in Section
5.5.3 will be useful for these applications.
FIGURE 1.4 Pathway up the process response surface to an ‘optimum’
A process may be performing with unacceptable consistency, meaning its internal variation is too high.
Excessive variation can result from many causes. Sometimes it is due to the lack of having or following standard operating procedures. At other times, excessive variation is due to certain hard to control inputs that affect the critical output characteristics of the process. When this latter situation is the case, one may experiment with these hard to control factors, looking for a region where the surface is flatter and the process is easier manage. To take advantage of such curvature in the surface, one must use designs - such as the second order designs of Section 5.3.3.6 - that permit identification of these features. Contour or surface plots are useful for elucidating the key features of these fitted models. See also 5.5.3.1.4.
FIGURE 1.2 Process before variation reduced
It might be possible to reduce the variation by altering the setpoints (recipe) of the process, so that it runs in a more ‘stable’ region.
FIGURE 1.3 Process after variation reduced
Finding this new recipe could be the subject of an experiment, especially if there are many input factors that could conceivably affect the output.
An item designed and made under controlled conditions will be later ‘field tested’ in the hands of the customer, and may prove susceptible to failure modes not seen in the lab or thought of by Design. An example would be the starter motor of an automobile that is required to operate under extremes of external temperature. A starter that performs under such a wide range is termed ‘robust’ to temperature.
Designing an item so as to be robust calls for a special experimental effort. It is possible to stress the item in the design lab and so determine which are the critical components affecting its performance. A different gauge of armature wire might be a solution to the starter motor, but so might many other alternatives. The correct combination of factors can only be found by experimentation.
A product or process seldom has just one desirable output characteristic. There are usually several, and they are often interrelated so that improving on one will cause a deterioration of another, for example: rate vs. consistency; strength vs. expense; etc.
Any product is a tradeoff between these various desirable final characteristics. Understanding the boundaries of the tradeoff allows one to make the correct choices. This is done by either constructing some weighted objective function (‘desirability function’) and optimizing it, or examining contour plots of responses generated by a computer program.
Sometimes we require more than a rough approximating model over a local
region. In such cases, the standard designs presented in this chapter for
estimating single order or second order polynomial models may not suffice.
Chapter 4 covers the topic of experimental
design and analysis for fitting general models for a single explanatory
factor. If one has multiple factors, and either a nonlinear model or some
other special model, the compiuter aided designs of Section
5.5.2 may be useful.
