That day the CPK report came out, the whole room was silent for three seconds
I remember several years ago, our production line introduced a new optical measurement system, boasting its powerful image analysis capabilities. After running for a week, the CPK report came out, surprisingly, at 1.08. To be honest, in our industry, this number is on the verge of failing. At that time, the production line manager's face was ashen, and the entire conference room fell silent for three seconds. Do you know that kind of atmosphere? It was more awkward than equipment crashing. Later, we discovered that the problem wasn't inaccurate equipment measurement, but rather our failure to understand the unique characteristics of image analysis when conducting MSA (Measurement System Analysis).
Where was the problem? The blind spot of "accuracy" and "stability"
Simply put, MSA is about confirming that your measurement system is sufficiently "accurate" and "stable." What you measure cannot be different every time, nor can it deviate too much from the standard. In traditional MSA, you might take a standard part, have different people measure it a few times, and check the repeatability and reproducibility of the data. However, for MSA of optical image analysis, things get complicated. Because it doesn't directly measure physical quantities like length or width. It "interprets images"! The system uses algorithms to identify features in an image, such as edges or defects.
Therefore, the key point is that when you use an image system for measurement, its ability to "interpret images" is affected by many factors unrelated to the measurement itself. For example, you might think measuring the size of a defect is easy. But if the image brightness, contrast, or even the blurriness of the defect's edge vary, the algorithm's judgment will deviate. This isn't the measurement system being "inaccurate"; it's "misinterpreting" the image.
How to actually do it? The "human" variable must be eliminated
To perform MSA effectively for an image analysis system, you first need to clarify the sources of variation. Traditional MSA's Repeatability and Reproducibility are still very important.
- Repeatability: The same product, the same operator, using the same machine, measured 10 times – the data must be stable enough. This reflects the stability of the instrument itself.
- Reproducibility: The same product, different operators, using the same machine, measured a few times – the data should not vary too much. However, image analysis has a significant pitfall: "human" judgment.
Therefore, the key is to eliminate the "human" variable. How do we do it?
- Establishment of a Standard Image Library: For different defects or features, create a standard image library. These images must be "golden standards" validated by human consensus or even more precise instruments.
- Algorithm Stability Test: Let the system run with these standard images and see if its judgment results are consistent each time. This is actually testing the stability of the algorithm itself, not the repeatability of the measurement system. If the algorithm itself has instability, for example, the determined edge position varies by 2-3 pixels each time, then your repeatability is fundamentally compromised from the start.
- Environmental Parameter Sensitivity Analysis: Before running MSA, adjust parameters of the optical system such as illumination, contrast, and focus to see how much these parameters affect the measurement results. For instance, if brightness is adjusted from 80% to 90%, and the measured DPMO changes from 6210 to 8500, then you need to consider whether to fix the brightness setting.
The most common pitfall: Non-standard standards
The biggest pitfall I've encountered is using "non-standard standards" for MSA. Once, we took several "purportedly" good wafers and had the system learn from them. The MSA results turned out quite good, and we thought everything was OK. However, after deployment, we found that the system classified some minor defects as good products, causing the DPMO to skyrocket. Upon careful comparison later, we discovered that the wafers originally used as standard good products actually had some minor nicks on the edges of a few dice, which were not obvious to the naked eye.
Simply put, your "golden standard" wasn't golden enough! Image systems are honest: what you feed them is what they learn. If your standard image library itself contains errors, then no matter how impressive the subsequent MSA figures for measurement results are, it's merely self-deception.
One thing you can do today
Re-examine your image analysis standard library to ensure they are true "golden standards."