A setter wants to find the best settings for a finishing operation. He changes the cutting speed and keeps everything else constant. Then the feed. Then the coolant. After a dozen trials he has an answer, and it may still be wrong, because factors often influence each other.
Designed experiments test several factors at once, in a planned pattern.
The two-level factorial
Choose two levels for each factor, low and high, and test every combination. With three factors that is 2 × 2 × 2 = 8 runs.
In the bore example the response was the deviation from nominal in µm:
| Run | Cutting speed | Feed | Coolant | Deviation µm |
|---|---|---|---|---|
| 1 | 180 m/min | 0.08 mm | 5 percent | 6 |
| 2 | 220 m/min | 0.08 mm | 5 percent | 8 |
| 3 | 180 m/min | 0.12 mm | 5 percent | 12 |
| 4 | 220 m/min | 0.12 mm | 5 percent | 14 |
| 5 | 180 m/min | 0.08 mm | 8 percent | 5 |
| 6 | 220 m/min | 0.08 mm | 8 percent | 7 |
| 7 | 180 m/min | 0.12 mm | 8 percent | 11 |
| 8 | 220 m/min | 0.12 mm | 8 percent | 13 |
Reading the effects
The main effect of a factor is the average result at its high level minus the average at its low level.
- Feed: (12 + 14 + 11 + 13) / 4 minus (6 + 8 + 5 + 7) / 4 = 6 µm
- Cutting speed: 2 µm
- Coolant: minus 1 µm
Feed dominates. The team set the finishing feed to 0.08 mm and locked it in the program. Every result was used to estimate every effect, which is why eight runs are enough.
The same design also estimates interactions, for example whether the effect of feed depends on cutting speed. In this example they were close to zero.
Six steps
- State the goal and the response, measured with a checked gage.
- Choose two to five factors and levels far enough apart to show an effect, but safe for the process.
- Build the design and plan repeats, so that noise can be estimated.
- Randomize the run order and record everything unusual.
- Calculate the effects and look for the few large ones.
- Confirm the best setting with separate runs before changing the standard.
When to use it
Designed experiments fit when several factors may matter and they can be set deliberately: machine parameters, recipes, assembly forces. They are less useful when the main causes are people, methods or the measurement system. Fix those first.
With more than four or five factors, fractional designs test a carefully chosen half or quarter of the combinations. That is the point to involve a black belt or a statistician.
The payoff
One factor at a time often needs more trials and misses interactions. A small factorial design answers more questions with fewer parts, and the result can be explained on one page: these factors matter, this much, and here is the setting we chose.