That Day, the CPK Report Came Out, and the Room Fell Silent for Three Seconds
I still remember years ago, our machine suddenly produced a batch of defective products; the yield rate plummeted from 99.8% to 95%. The production line manager's face turned green, and he immediately called us for a meeting. As soon as the data was presented, the Cpk value for a critical dimension dropped from 1.6 to 1.08, and DPMO instantly surged to 6210. Everyone looked at each other, not knowing what to do. That's when we initiated the Measure phase of DMAIC, which, plainly speaking, was about figuring out "what the heck Y = f(X) really means."
Where's the Problem?
To be honest, many times when we see a drop in yield, our first reaction is, "Is the machine broken again?" or "Did we change a parameter incorrectly last time?" But in reality, the "yield rate" you see is our Y value, which is the "outcome." And what you suspect—"machine conditions" or "parameter settings"—these are the potential X values, which are the "causes" that might affect the outcome. The Measure phase requires you to use scientific methods to uncover the relationship between these Y and X values and establish a hypothesis. In other words, it's about identifying the critical factor among a multitude of variables that is most likely affecting your yield rate.
How Is It Actually Done?
How did we do it back then? First, we listed all possible variables that could affect this critical dimension, roughly over twenty of them. From the batch number of a certain incoming material, machine temperature, pressure, rotation speed, and even the shift of the operator on duty were included. Next, we used MiniTab software to run a series of statistical analyses, such as Correlation Analysis or Regression Analysis.
For example, we discovered then that the machine's "reactor pressure" showed a strong negative correlation with the Cpk value of that critical dimension, with a correlation coefficient as high as -0.85. This means that the higher the pressure, the lower the Cpk value. At this point, we had our first concrete hypothesis: Y (Critical Dimension Cpk) = f(X, Reactor Pressure). This hypothesis wasn't a wild guess; it was driven by data. So the key is, you cannot just guess based on experience; you must use data to validate your intuition.
Most Common Pitfalls
Frankly, the most common pitfalls in this phase are "insufficient data" or "incorrect data." Once, we also suspected an issue with a consumable supplier. So we went to pull data, and it turned out that the consumable batch numbers entered into the production system were often incorrect and didn't match the actual batch numbers used. In such cases, no matter how good your statistical tools are, the results you get will be garbage. Therefore, before starting the analysis, you must spend time confirming whether your data sources are accurate and complete. Spending more time analyzing incorrect data will only lead to incorrect conclusions. Another pitfall is trying to find too many X's at once, pulling in all possible factors, which results in an overwhelming amount of data and makes it hard to know where to start. It's more efficient to focus on a few of the most impactful X's at a time.
One Thing You Can Do Today
Go back and look at your latest yield abnormality report, and list three X's that you find most suspicious.