When you think of artificial intelligence, maybe you think of IBM’s Big Blue processing a million potential results of one move on a chessboard, or a supercomputer that predicts weather. Actually, you don’t have to step out of the machine tool world to experience artificial intelligence.
MC Machinery’s (Elk Grove Village, IL) recent product releases, the GX-F ADVANCED Series of fiber lasers, as well as the company’s latest sinker EDM offering, the Mitsubishi SV12P, take advantage of Mitsubishi Electric’s (Tokyo) neural network artificial intelligence. (See related stories on the GF-X ADVANCED laser and the new MC Machinery sinker EDM with Maisart.) Called Maisart (Mitsubishi Electric’s AI creates State-of-the-ART in technology), it results from about a decade of research and development.
Hank White, MC Machinery’s National Project Specialist, Fabrication, explains that a couple of years ago, “Mitsubishi Electric first came out with this technology in automotive, for self-driving cars.” Predictive elevator systems also leveraged the technology. The machine tools use a similar concept: “We give the machines parameters and put it into a real-world scenario,” says White.
“Mild steel cutting is a perfect application,” White continues. “As a fabricator, you get steel from different vendors. It varies in quality and resources used. That’s not a problem on a CO2 laser, but on fiber it’s different. Fiber laser machines can be fickle if you change the consistency in the material.” White says that with the potential for less experienced machine operators in today’s fabricating world, the company is trying to make the machine intelligent enough to compensate for the lack of experience.
He explains, “The laser uses audio and optical sensors to manage the cut. It listens for the sound of the cut; you should have a certain frequency when you cut and if it captures sputtering sounds, it stops. It’s the same with the optical portion. When you process mild steel, you should get a certain spectrum from it.
“The system will make adjustments based on what it sees or hears. If the light is brighter or senses a flash, then we need to check our processor or the nozzle. The system then looks at the nozzle to check it out. If it sees an issue, it replaces the nozzle. If the nozzle is OK, it returns to cutting, and it will decrease the feed rate. Once it sees that the cutting has become nominal, it speeds up again.”
These actions go beyond problem resolution, though: “If the normal cutting is going well, the system might even increase the feed rate, it will go up to 110% of the normal feed rate,” White notes.
Inside the system
A neural network is the core piece of Maisart. Like other neural networks, it imitates the human brain, and how neurons work together to solve problems. Maisart can learn from experience and do logical inference—which improves as it learns. Three layers make up a neural network:
- The input layer (inputs from the outside, real world);
- The hidden layer (where a lot of the processing and decision-tree direction takes place); and
- The output layer (where the result of a decision begins to play out in the real world).
Neural networks separate these layers to divide and conquer the processing. Doing so means that the system can perform high-level tasks like recognition, identification, and analysis.
One problem that crops up repeatedly in neural networking is that it requires huge amounts of processing in several steps. It’s difficult, if not impossible, to put large amounts of processing power at the point where it’s needed. Mitsubishi Electric’s answer to this was to diminish the seemingly infinite possibilities into a reduced set of potential outcomes that are solidly anchored in the real world. A new algorithm was developed. Pruning potential pathways pushed processing down to 1/30th to 1/100th of what was originally needed, reducing both time and computing power requirements (see Figure 1).
One more gain was made in the learning process. An AI device exposed to a situation will set its own rules and determine what action to take. To do this, it needs lots of experiences and lots of context—including failure. The device tries again and again, and learns because it is rewarded for the right actions. This is called reinforcement learning, and can be a painstaking process. By using proprietary technology, Mitsubishi Electric reportedly reduces the number of trials to about 1/50th of the conventional total (see Figure 2).
With a mature AI effort from Mitsubishi Electric and the dawn of AI in the MC Machinery product line, what will the near future look like? White states, “This technology will be standard on all Advanced machines, our flagship machine tools. Currently 50 percent of what we sell is our flagship line. Entry level machines will not have it, it will not even be available on that platform.”
The fabricating market, and especially the fiber laser market, is becoming a fast-moving target, and White asserts, “I could see this coming down-market one day. Five years from now, who knows? It could be on every laser we sell someday. If this technology keeps progressing, you will not need a cutting condition table, you will simply choose a material type and it will cut it.”