MAHALANOBIS TAGUCHI (MT) SYSTEM
ovitAlps Electric's quality engineering efforts were recently introduced in the Nikkan Kogyo Shimbun.
The article discussed the company's efforts to eliminate outflows of defective products by adopting the MT system.
I knew little about the "MT System" due to my ignorance, but its concept was so close to the FlexInspector philosophy (or is the opposite more accurate?) that it caught my interest. I was intrigued. I immediately obtained the material and looked into what was the same and what was different.
After all, in the case of images, the amount of data is too large (not impossible if features are extracted from the image and applied to it. However, it is the FI style not to extract the feature amount by force). We found that it is not possible to apply the method properly.
However, the idea of defining a "good" as a homogeneous population (unit space) and assuming that a larger distance from it is an anomaly can be considered almost the same, and many of the advantages gained by the MT system overlap with those of FlexInspector. It is brilliant, albeit coincidental.
Once again, I thought that the current machine vision industry does not have much contact with quality engineering. Perhaps this is the reason for the deep, deep divide between manufacturers and users. I have a feeling that this is the reason why I have been feeling that "the more quality engineers talk about quality, the more they appreciate FlexInspector".
There is much to learn from Mr. Hitoshi Yamada, who is known as the "Kaizen Demon," and who has inherited the legacy of Mr. Ohno, the founder of the Toyota Production System.
He said, "If the manufacturing industry exists to "make garbage," it is only natural that manufacturing places will become workplaces where there is no motivation and no hope for workers.
"It is only natural that manufacturing will become a place where there is no desire to work and no hope for the future.
The abundance of industrial products is beginning to destroy the environment, causing energy problems, and making the earth unlivable for both people and nature.
The earth is becoming unlivable for both people and nature..."
When a manufacturer plans to produce and sell only good products, it is a contribution to people and the natural environment.
I feel that this way of thinking and FI's worldview have the same qualities.
I think the principle of the MTS method and variation are the same concept.
We used to use the MTS method for defect classification and for quantification of sensory evaluation. However, in the field of pattern recognition, the MTS method is only one of the methods, so we now use pattern recognition methods that suit our applications.
Returning to the MTS method, classification is generally performed using features obtained from images, but there are also examples of applications of the MTS method that directly use gray values for character recognition and defect detection. However, this is similar to the Taguchi method, which is based on the idea of using brute force because it is not well understood.
Even with FlexInspector, there is a threshold that determines whether what is finally detected is garbage or a defect, but is there a need to make this determination more sophisticated?
No matter what kind of image processing is done, it is necessary to at least make a black-and-white determination of whether the final image is OK or not, and I think pattern recognition technology is necessary for that.
Thank you for your comment.
I personally believe that some areas of machine vision require "pattern recognition" (i.e., image understanding) and some do not. I think that extracting feature values such as dimensions from an image is a kind of pattern recognition. I also believe that extracting information from a single image using various methods is a very meaningful study in itself.
What I would like to discuss is the fact that there are simpler areas in visual inspection, such as "If it is the same as a good product, it is OK; if not, it is not OK; the type of NG is not important. In FlexInspector, the basic policy is to make the setup as simple as possible, so the "NG type" is discarded. Instead, we have made it simple and versatile. (In FlexInspector, any defect, even if it is garbage, is treated as an NG if it differs from a good product. ) There are functions such as area thresholds, which allow for simple size-based delineation, but no advanced classification functions at all.
) But you have to understand that FlexInspector can't inspect just anything, and the good products have to be uniform. In short, it depends on the product, and we cannot be sure what the results will be until the product is in production. We recognize this as our greatest weakness.
If something doesn't work, it is necessary to "recognize patterns. Once we get into that area, I think we will be stuck in the same one-product-at-a-time situation as we have been in the past.