Milko Gergov, a 45-year plastics veteran, is known for the IntelliMold process he developed in the early 1990s, which controls injection molding melt pressure and temperature using transducers.
His motto sounds like an Industry 4.0 mantra: "If you cannot measure it, you cannot control it. If you cannot control it, you cannot improve it."
Plastics News senior reporter Bill Bregar posed a series of questions to Gergov on one his favorite topics: artificial intelligence vs. machine learning. Both are hot buzzwords, but Gergov said they are far from the same thing.
"Machine learning is the front runner of AI. Machine learning is the workhorse of AI," he said.
He also discussed the concept of deep learning. And he addressed implications for society and the workforce — human beings.
Gergov gives the analogy to a cruise control for a car. Cruise control is evolving into intelligent cars after adding radar and cameras and self-parking ability.
"Skilled human decisions are still an integral part of the setup and the use," he said.
Q: Artificial Intelligence makes some people nervous. They are afraid it will take over human decisions.
Gergov: The complexity is vast and multidirectional. When AI is to be in an adviser's mode to help the factory worker or a decision maker is still a hot topic and debatable. Deficiencies in AI software and hardware in the past and in current projects cannot be a reason to stop improvements. Misrepresentations like AI is good Robocop turning bad Robocop is for Hollywood.
Q: A segment of the factory worker population, or before that the skilled craftsman like the cooper that made barrels, always feared technology.
Gergov: Finally, that's the big one. That artificial intelligence will put most of us out of work. For sure, there will be integration of the components of AI that, in my opinion, is needed to drive people for continuous improvement in education and ability. Currently, most of the components of an AI package are fragmented; they are like separate industries with no sense of integration, which is slowing down implementation. For example, components like drivers, software and sensors. Better integration of those components will speed up implementation of AI.
Q: What are your views on machine learning?
Gergov: The first practical implementations of machine learning in plastics started to emerge in the beginning of the 1990s, when limit switches were replaced with position sensors, melt pressure/temperature sensors that became available. Also, closed loop-controlled hydraulics and temperature controls were implemented. The need to implement process control in conjunction with machine control was obvious. Different modes of closed-loop control (PID) injection phases developed as alternative.
Q: How did this evolve?
Gergov: Slowly the results from these steps did show the importance of the feedbacks for process control to be deeper in the melt site. First, soft sensors were developed that calculated process variables from real-time measurements from physical sensors, and software that solves mathematical problems, were developed and used as feedback for closed-loop control of injection in real time.
With the first all-electric injection molding machines came new modes of controls like fast-feed forward (FFF) injection control using expert algorithms. Statistical process control and data pull-up tables control packages was implemented, also.
All of the above today — after major improvement in the speed of machine controls electronics — delivers a significant amount of data collections in real time, as capability and use for control, is called machine learning.
Q: You've talked about something called deep learning. What is that?
Gergov: In the last several years, the use of more and smarter sensors in the manufacturing systems demanded way more powerful analytical tools to process the gigantic volumes of information collected in real time. A new generation of hardware and software came to life called deep learning.
These very powerful computing servers use sophisticated mathematical modeling, simulation modeling software such as flow simulation and finite element analysis. There are vast databases for the materials, systems and components involved that have been able to generate characterization patterns. And from those patterns, the system can select functional or behavioral models.
Q: Why are those models important?
Gergov: The improvements in requirements for priorities and definitions, models training and evaluations are becoming a solid base for finding patterns in the data, correlations to the target and algorithms for predictive accuracy. These achievements finally gave deep learning the ability to produce reasons and decisions for changes.
Q: How does this all come together in Industry 4.0?
Gergov: As much as the machine learning and deep learning cover the technical side in artificial intelligence, there is a tremendous need of administration, standard rules of engagement and industrial applicational support is obvious and very important.