That day, when the CPK report came out, the room fell silent for three seconds, and then I saw the manager's face turn green.
Do you remember that batch of goods in Q3 last year, when the yield suddenly dropped significantly? At that time, we were all over the place trying to chase the yield, checking machine parameters, material conditions, and even asking the PM from the adjacent factory, but we still couldn't find the key reason. Everyone was like headless chickens, trying to adjust each process a little, only to make things worse. Cpk dropped from 1.08 to 0.95, and DPMO surged to 6210. In that meeting, the manager looked at the report, his face ashen, and said, "This batch of goods is like Schrödinger's cat; no one knows what went wrong." That's when I started to wonder, what if it's not a problem with a single process at all, but rather a group of processes that are "very similar" and failed together?
Where's the problem? Not a single process, but a "group" of processes.
To put it bluntly, in process analysis, we often mix all the data together. But honestly, many times, different processes, even if their steps are similar, can have vastly different "behavioral patterns." Imagine you have a pile of apples and oranges; if you weigh them all together, you'll only get an "average weight of fruit," which tells you nothing about how much the apples and oranges each weigh.
So the key is that we need to "categorize" these processes first. Cluster Analysis does exactly this; it helps you find those "similar" processes in your data and automatically group them. You can think of it as a super-smart classifier that, based on various process characteristics—such as temperature, pressure, time, or even machine model—identifies those processes that "are alike."
How is it actually done? Use numbers to find your "interrelated processes."
The steps are actually quite simple, no need for any advanced mathematical formulas.
- Collect your process data: Pull out all the process parameters you believe might affect yield. For example, temperature, time, and pressure for process step A; vacuum degree and gas flow for process step B, and so on. The more data, the better, but don't overdo it; pick those you consider critical.
- Choose a suitable algorithm: The most commonly used and intuitive one is K-Means. It helps you divide your data into K clusters, where K is the number of clusters you want to set.
For example, suppose we have 100 process steps, and each step has 5 key parameters. We feed this data into the K-Means algorithm and set K=3. The results might tell you:
- Group 1: 30 processes, whose common features are "high temperature, short duration, low pressure," and their yield performance is generally poor, with an average Cpk of only 0.98.
- Group 2: 50 processes, characterized by "medium temperature, long duration, high pressure," with stable yield performance and an average Cpk of 1.25.
- Group 3: 20 processes, characterized by "low temperature, short duration, medium pressure," with fluctuating yield, where Cpk falls between 1.05 and 1.15.
This way, you'll know that when the yield of processes in Group 1 drops, you should prioritize checking the three common points of "high temperature, short duration, low pressure," instead of making random adjustments aimlessly. In other words, you've identified your "interrelated processes"—when one has a problem, others in the group might follow.
Common pitfalls: More data doesn't always mean useful, and "how many clusters?"
I've fallen into this trap before. The first time I used cluster analysis, I thought more data was better, so I threw in all process parameters, machine models, and even operator shifts. The resulting clusters made no sense, like a group of "machines operated by Engineer Wang on Monday morning shift"—which has nothing to do with yield.
So, the most crucial point is: the input features must be relevant to your problem. If your goal is to improve yield, then you should include parameters that you believe will affect yield.
Another big pitfall is "how many clusters to form (the K value)." Initially, you might not know whether to set K=3 or K=5. That's okay; you can try multiple times to see which clustering result is most insightful. The most common method is to look at the "Elbow Method," which provides a visual suggestion for the optimal K value. But truthfully, this still requires you to use your experience and understanding of the process to make a judgment.
One thing you can do today
Organize your existing process data, identify the three parameters you consider most important, and try using Excel's scatter plot function to get a preliminary look for any clustering phenomena. This is the starting point for cluster analysis.