That Day, Yield Dropped to 88%, and Adé's Face Turned Green
"Senior, we have an issue!" Adé, the rookie, rushed in, face pale, his report trembling in his hand. "Last night's batch, the yield is only 88%! The customer's DPMO has skyrocketed to 12500, and the PM is already fuming in the office!" I put down my coffee and glanced at the report. Hmm, it's been a while since I've seen such "exciting" numbers. Usually, at times like this, the boss starts asking, "Does anyone know which parameter is causing this?" Then everyone starts shooting in the dark, adjusting this up, that down, and the result is usually... even worse.
Where's the Problem? You Can't Fix It By Random Adjustments
To be frank, many times when we encounter problems, our first instinct is to guess based on experience. Adjusting the machine temperature, found it didn't work; increasing the pressure a bit, seemed even worse. This is the typical blind spot of "changing only one parameter at a time." But do you know? Sometimes the problem isn't caused by a single parameter at all, but by "Parameter A" and "Parameter B" combining to cause trouble! This is what we often call "interaction effects."
Imagine you're out shopping with your wife on the weekend. Your wife asks, "For dinner tonight, Japanese or Italian?" You reply, "Japanese." She then asks, "Ramen or sushi?" You reply, "Ramen." As a result, you're unhappy with your meal. After getting home, you realize that you like Japanese food, but specifically dislike ramen; you prefer sushi. But if your wife had asked, "Ramen or Italian?" you might have chosen Italian. This is an "interaction effect" between two factors (cuisine type, specific dish).
It's the same in the factory. Temperature 100 degrees is fine, pressure 5 Bar is also fine, but if "Temperature 100 degrees + Pressure 5 Bar" are combined, problems might arise. Therefore, what we need to know is not just the "impact of individual parameters," but also "what happens when parameters are combined."
How Is It Done In Practice? 2^k Full Factorial Experimental Design
This is where 2^k full factorial experimental design comes in handy. Simply put, it involves identifying the parameters (factors) you believe "might" affect the result, and then setting two levels for each factor: high and low.
For example, let's say you suspect that "Temperature (A)" and "Pressure (B)" affect your product yield.
- Factor A (Temperature): Low level (e.g., 100°C), High level (e.g., 120°C)
- Factor B (Pressure): Low level (e.g., 5 Bar), High level (e.g., 7 Bar)
This gives you 2^2 = 4 combinations:
- Low temperature, low pressure
- Low temperature, high pressure
- High temperature, low pressure
- High temperature, high pressure
You run through all four of these combinations and record the yield data. Suppose the data comes out like this:
- Combination 1 (Low Temp, Low Pressure): Yield 95%
- Combination 2 (Low Temp, High Pressure): Yield 90%
- Combination 3 (High Temp, Low Pressure): Yield 92%
- Combination 4 (High Temp, High Pressure): Yield 88%
From this data, you can calculate the "temperature effect," "pressure effect," and the "interaction effect between temperature and pressure." The calculation method is not difficult, but the key is that you will obtain the independent impact of each factor, as well as their combined impact.
Therefore, the point is that you don't need to guess blindly; instead, you can systematically test all possible factor combinations to truly identify the root cause of the problem.
The Most Common Pitfall: Thinking DOE Costs a Fortune
When I first started, I also thought DOE was a "big production," requiring several days and consuming a lot of machine time and materials. Once, the boss asked us to investigate an anomaly on a machine where the yield dropped from 99% to 97.5%, and the CPK also decreased from 1.5 to 1.08. Several veterans in our department, including myself, thought a certain machine part was broken. We disassembled and inspected it for half a day, but found nothing.
Later, the boss told me to try a 2^3 DOE, identifying three parameters we deemed most critical, running each at high and low levels. We only spent two days, using an online machine, choosing off-peak hours for testing. The result showed that two of the parameters, A and C, caused a significant drop in yield when "A was high and C was low," but adjusting A or C individually had no effect! Before, we would only adjust them individually and never discovered this "combination bomb." After that, I truly believed that DOE is not a luxury, but a powerful tool for problem-solving.
To be frank, many people find experimental design intimidating, thinking it requires complex statistical software and a lot of time. But in reality, for initial problem diagnosis, a simple 2^k DOE combined with Excel can help you clarify most issues.
One Thing You Can Do Today
Think of the most troublesome problem you currently face, list 2-3 factors you believe might affect the outcome, and then plan a 2^2 or 2^3 experiment to try on a small scale.