Challenging human quality inspection
Automated systems still struggle to replicate human efficiency in some visual inspection applications, but there are ways to get machine systems up to speed. Automated machine vision systems can improve over time; see three ways machine vision is catching up.
Automated machine systems are improving the quality and efficiency of what manufacturers can accomplish, but they are not always a perfect analog for humans. One major area where automated systems struggle to replicate human efficiency is in the field of visual inspections, said Corey Merchant, vice president-Americas at Kitov.ai, in the presentation “Challenging Human Quality Inspection with Modern AI and Deep Learning Technologies” at Automate Forward, an A3 online conference and exhibit during the week of March 22.
It is estimated more than 80% of all visual inspection is still done by humans, Merchant said. But human inspection has its drawbacks – humans can be costly, inconsistent and error-prone.
Visual inspection: Low volume, high-mix applications
“It has been said the manual inspection is really the Achilles’ heel of manufacturing,” Merchant said. “So if it is the Achilles’ heel, why not use robots or other automation instead of humans? Well, it turns out humans are actually pretty darn good at human inspection, especially when used in low-volume, high-mix operations.”
The bottom line is humans still offer many advantages over traditional machine systems. Humans are excellent at adapting to new products, they’re product agnostic, they have 3D vision, they have hand-eye coordination, they learn from experience, and, perhaps most important, they can deal with variation.
Variation has always been the big problem for machines. Merchant compared it to inspecting a car. You have the tires, paint job, glass, interior and more. Each presents different inspection challenges. How many parts, and thus inspection points, are there in an automotive engine? In addition, there is variability in the kinds of defects. Is that a scratch in the paint or a human hair? Machines still struggle with these kinds of differentiations.
The human eyes and brain are a powerful image system. Humans offer advantages like 3D vision, vision memory, contextual understanding, generalization and conceptualization. They can be trained and can adapt to situations. But they also have limitations – emotions, inconsistency and an inability to manage large amounts of data – that machines can offset.
Three ways machine vision is catching up
So what’s involved in creating automated systems that have the ability to replicate what humans do for a lot of the final product inspection? There are a number of different factors, according to Merchant, beginning with an easy, intuitive set-up. This includes quick and semantic teaching methods, automatic switching between different products, 3D viewing and optimal inspection for different product sizes.
1. Flexible and variable machine vision
“It’s one thing we’ve heard through a lot of the speeches and presentations this week from all different manufacturing segments when asked what is the most important thing with automation: It’s flexibility,” Merchant said. “Flexibility and variability, dealing with that. And we can’t emphasize that enough. You can build something to fit today, but is it going to adapt with your process? That’s the question.”
2. Smarter machine vision logic
The second piece manufacturers need is a robust and powerful detection engine. This must comply with all inspection requirements, offer complementary algorithms for a wide range of defects, and be able to handle various materials and textures.
3. Machine vision with machine learning
Finally, Merchant said, the machines must have the capacity to learn, which typically requires a large volume of data. This means they need to adapt to manufacturing variability, require very few samples and adapt to a dynamic production environment.
The solution, Merchant said, is a hybrid model that includes deep learning and 2D and 3D classic computer vision with artificial intelligence. It’s all about increasing yield. That’s really what everyone is after. So how do you block defective parts while keeping out phantom defects? That’s where the system will earn its stripes, Merchant said, but it’s also where things get difficult.
The challenge of deep learning is it requires tons a data. Merchant said one predictive variable can be learned for every 10 events, which means manufacturers must collect a lot of events.
To find the time required for systems to get up to speed, you must combine technologies using this hybrid approach, Merchant said. When sufficient data is accumulated, things can actually be better with automated systems than they are with human inspection. Inspection accuracy should be at a higher level, and the system can improve over time.
“You can think of autonomous driving being so popular now; you’re teaching that car how to drive autonomously,” Merchant said. “It’s not a binary decision path. It’s not all 1’s and 0’s. It’s not yes or no or black and white. It’s actually based on a history of widespread examples that it’s seen. And then those are used to constantly adapt and improve that performance.”
Gary Cohen, senior editor, CFE Media and Technology, firstname.lastname@example.org.