That day, when the CpK report came out, the room fell silent for three seconds before I dared to say: "This result, there's a problem with it."
I recall an instance when we had just introduced a new visual inspection machine, touted for its AI-powered judgment and sky-high precision. The machine's acceptance CpK report, however, came out at a mere 1.08! The entire conference room went silent instantly, and even the air seemed to condense. The boss's face was grim, and everyone exchanged glances, no one daring to speak. I thought to myself, this value is far too low; logically, a new machine like this should achieve at least 1.33. It was then that I boldly spoke up: "Boss, I suspect this CpK might not be an issue with the machine itself, but rather with our measurement system."
Where's the problem? Are your "eyes" accurate?
To put it plainly, many times when we measure a low CpK, it's not truly due to an unstable process, but rather that the "ruler" you're using for measurement is itself inaccurate. In automated visual inspection, this "ruler" is your AOI machine, your algorithms, and your optical system. When you're judging defects on wafers or measuring component dimensions, if this visual system itself "sees wrongly," how can you possibly obtain accurate process capability data? This is the core problem that MSA (Measurement System Analysis) aims to solve. It's not about whether your product is good or bad, but rather whether the tool you use to "judge if your product is good or bad" is itself good enough.
How is it actually done? Let the machine and humans "take a test" together.
To evaluate a visual inspection system, the most common approach is to perform a Gage R&R (Repeatability & Reproducibility). Simply put, you find several "standard samples" (with defects, without defects, oversized, undersized, etc.), have your machine measure them repeatedly many times, and then have several operators measure them manually.
- Repeatability: Take the same sample and have your machine measure it 20 times. If it makes the same judgment every time, for example, consistently judging it as NG, then its repeatability is good. However, if the same sample is sometimes judged as NG and sometimes as OK, then the machine's "trembling hand" (instability) is significant.
- Reproducibility: Have several operators measure the same sample to see if their judgments are consistent. For instance, if three operators all judge the sample as OK, then reproducibility is high. If one person judges it as OK and another as NG, it indicates significant human judgment variation. In visual automation, this part is more akin to evaluating the consistency of judgment for the same wafer across "different machines" or "different shifts."
Ultimately, you will obtain a %GRR value. If %GRR < 10%, congratulations, your measurement system is very reliable. Between 10% and 30%, it's marginally acceptable, but improvement is highly recommended. If it exceeds 30%, your measurement system is utterly chaotic, and the data it provides is unusable. I encountered a case where a new machine's %GRR was surprisingly as high as 45%! No wonder the CpK was unbelievably low. That wasn't a process issue at all; the machine itself was "lying."
The Most Common Pitfall: Blaming the Process While Forgetting to Check the "Eyes"
When I first started my career, I often saw poor CpK reports and would immediately rush to the production line to find the process engineers, asking them, "What's gone wrong this time?" As a result, the process engineers were hounded by me and found no clear explanation after half a day of checking. Later, it was discovered that the measurement machine's light source had issues, leading to unstable image quality, which in turn caused the AI judgments to be erratic. Therefore, whenever I see abnormal data, my first reaction is never to blame people, but to start by checking the measurement system. Frankly, engineers often place too much trust in instrument data, forgetting that instruments are also designed and operated by humans, and they too can make mistakes.
One Thing You Can Do Today
Go immediately and check your equipment's MSA report to see what the %GRR is!