AI-Driven Manufacturing Efficiency Tool

Ed Brown

Tech Briefs: What got you started on this project?

Qing “Cindy” Chang: My background is in manufacturing systems — before I joined academia, I worked for about 11 years at the General Motors R&D Center. At that time, our work was focused on how to improve overall plant operations — quality and efficiency. We built a math-based model, and we also used data-driven methods. We tried to understand the connections between the factory physics and what the data told us.

Tech Briefs: Could you explain what you mean by “the relationships between the factory physics and what the data tells you.”

Chang: Physics assesses the overall manufacturing system. When we build a model, it’s a math-based model. We have some basic understanding of the connections because a plant is a really large system. We have machines, we have buffers, we have material handling, we have maintenance. So, we need to have a model to describe these things. From the model, we can explain some of the phenomena. We had lots of data at that time; even 20 years ago, we had already collected lots of data from the plant floor. From those data and from what the model told us, we could build some kind of understanding of the connections.

For example, our very early work was on how to identify bottlenecks in the plant. We used math principles to help find them, but that is very difficult for an engineer or a plant worker to understand. However, since we had collected all that data, we could use indirect methods using the data to find the bottlenecks. But that result has to match the math-based models.

Starting from that time, about 2007, AI has been getting really hot — there have been some great advancements. Also, by that time I'd joined academia, and we started to do more research about how to make further improvements using AI. We found that multi-agent reinforcement learning (MARL) is very useful for solving many problems that we couldn’t solve earlier.

Our aim was to fill the gap between individual processes and overall systems. Although systems include multiple processes, the earlier focus had just been on an individual process without understanding the big picture of the overall system. System people tended to just look at a high-level kind of system efficiency without considering how their high-level decisions can propagate down to individual process parameters. So, our work has been to bridge this gap.

Tech Briefs: That sounds like a complicated task.

Chang: Yes, it is. For the past 10 years. We worked on both the process level and the system level. So, we had a good understanding of both ends. Now with this machine learning tool we think we can indeed, bridge the gap.

Tech Briefs: Can you explain a little about how your model works in practice.

Chang: We didn’t want to simply solve the problem — we wanted to know how and why, what are the underlying reasons. Based on our experience, we built a mathematic model that gives us a deeper understanding of the overall plant-floor multi-stage processes. So, we had the model — the basic understanding. However, at that time, we didn't use machine learning. So, now we want to see how we can embed that understanding into machine learning. That is one way we can solve the problem that we weren’t able to solve earlier — it strengthens the machine learning. That way, the machine learning is a tool, not just a black box. We embed our knowledge into the tool so that the result is more convincing — we know what’s going on.

The plant has multiple processes, and each process has its own parameters, for example, its capacity. There is the designed processing time, but also lots of randomness, such as random downtime and varying material arrival — the overall system is therefore highly nonlinear, it’s stochastic. So, we built a state-space model to link all the multiple processes, which considers all the uncertainty, all the nonlinearity, in the context of the control concept.

However, we couldn’t use classical control theory. But we could solve it by feeding in the real data and developing a quick recursive algorithm. There's no closed-form representation, but reinforcement learning can help us solve the control problem.

Tech Briefs: What do you mean when you call this a physics-based model?

Chang: “Physics based” means understanding the relationships using the model. We view the manufacturing system as a typical engineering dynamic system and there is a model to describe that. From that model we can predict performance and derive some important properties of the system.

Tech Briefs: Are you still working in coordination with General Motors?

Chang: We just finished one project with them, and we also had a grad student interning there last summer, and we're probably going to continue to do that this coming summer. We will also be discussing continuing collaboration.

Tech Briefs: Have you applied your model in actual practice?

Chang: It was used to solve some real-time bottleneck problems in a General Motors plant based on the work of my Ph.D. student intern last summer, who developed a new tool and technology — although it was based on our previous work. It was used to address some new challenges on their assembly line.

Tech Briefs: What are you working on now?

Chang: We are continuing to work on improving overall manufacturing system efficiency, while also including an energy savings aspect. How do we consider alternative energy in the overall formula?

We also expect that advanced manufacturing systems will become even more flexible. In a future factory, we can imagine there would be lots of mobile robots for jobs like material handling. And there will be a lot of human-robot collaboration issues. So, we are spending a lot of time working on human-robot collaboration; training the robots to work well with humans and learn from them to improve overall efficiency. We see robotics in automation as a tool for humans, not a replacement. So, we are working to understand those issues and how to solve them. This can be described as human-centered AI or automation.

Tech Briefs: Are you planning to apply this to other industries outside of automotive?

Chang: Yes, currently we are working with the Commonwealth Center for Advanced Manufacturing in Virginia, which is a state-funded center that connects university researchers with manufacturers. So, by our close collaboration with this center, we are hoping to make manufacturers aware of our work.