Myths and truths about AI enabled Automated Visual Inspection for manufacturing
AI in manufacturing (Autonomous Manufacturing) will be the big trend for the future, starting from automated visual inspection.
However many manufacturers hit roadblocks for AI adoption, they spent time and money on PoC (proof-of-concept) and then found out that AI doesn’t seem all that helpful and it is hard to use, why?
A lot of the times there are myths and truths (misconceptions) about AI which lead to those situations.
Myth 1: I solved that inspection problem in PoC and my software should be able to deploy right away.
That’s 99% of the manufacturers (if not 100%) think when they first tried to use AI for inspection. And most likely they will later find out that in deployment, the “perfect ML model” does NOT work at all on manufacturing line, and they keep fire fighting and eventually give up.
Why is that?
Truth 1: Manufacturing problems and the underlying AI problems are totally different.
When you have a manufacturing problem, for example you want to use a new glass material to make a cup. You do a PoC to make 1000 first, once you test your cup, your manufacturing recipe is almost ready to go with the right DFM/FMEA/Reliability/Quality documents. When you start to make 1 million new cups per year, your recipe will mostly likely be 99.5% (if not 100%) the same to our PoC recipe.
However your underlying AI/ML problem for detecting defects for that glass, is very different. No matter how good you designed your experiments in your PoC for ML model development and testing, your will encounter a very different problems. An detection model with 1000 samples and a detection model with a million sample will always be very, very different.
So the key here is not to only excel in PoC, the key is really to understand how to transition from an awesome PoC ML model to an awesome deployed manufacturing ML model.
Myth 2: There are so many open source ML models online, implementing AI is easy, all I need to do is just hire a few college student, took Andrew Ng’s Coursera class and I should be able to use ML.
A number of companies I talked to actually did that, and some of them even hired professional ML engineers to do it.
However what normally end up happening is that they might get some success in PoC but rarely succeed in deployment on production line. The team stayed as “innovation team” and cannot make real impact on business.
Truth 2: ML model is very, very small part of the complete ML system.
For manufacturing, below are the complete list of problems one need to solve to get a working system:
See more details in: https://landing.ai/wp-content/uploads/2020/10/LandingAI_WhitePaper_20001-02-EN.pdf
As you can see, custom model is just 5% of the scope of the complete system. And what’s more, the PoC model is not even the “custom model”. That is why , when manufacturers want to be successful in AL, they need an army of ML engineers, SW engineers and operators to work together to be successful.
With the myths and truths be told, it sounds like the cost for successful AI implementation is very hard and expensive, but is it worth it?
Let’s look at the companies in manufacturing that did adopt AI.
By implementing AI in metrology, which is the inspection at wafer level. They not only improve quality, but also unlock novel process capability for more advanced manufacturing. Today they are the best semiconductor foundry in the world.
Embrace AI for manufacturing can be hard, but it will be rewarding as well.