AI on the Plant Floor Is Not What You Think It Is

In 2016, a computer program called AlphaGo beat professional Go player Lee Sadol. Google DeepMind built the system using deep neural networks trained through reinforcement learning. Nearly a decade later, that same underlying technology shows promise in optimizing processes throughout chemical plants and filling knowledge voids left by retiring industry veterans. 

But the great AI hype of today has given way to misconceptions about the current reality of AI and its applicability in chemical processing. 

While generative AI, the large language models (LLMs) like ChatGPT, commands most of the attention, it is autonomous AI that holds the most immediate potential for the plant floor, said Bryan DeBois, director of industrial AI for systems integrator RoviSys.

Members of the RoviSys team will be in attendance and exhibiting at Chemical Processing’s eChem Expo conference taking place in Kingsport, Tennessee, April 8-9. The company has been working with manufacturers from various sectors to integrate AI into their existing operations.

Unlike LLMs that recognize patterns, autonomous AI can make decisions, operate within defined constraints and deliver deterministic results. 

That’s a critical distinction because LLMs are prone to hallucinating and providing varied results. That may be OK for booking travel reservations but could be catastrophic in a safety-sensitive environment like a chemical plant, DeBois said. 

“On the plant floor, the stakes are much higher,” DeBois explained. “If you give the wrong information to the wrong person at the wrong time on the plant floor, you could kill somebody; you could blow up the plant. Right now, our message to customers is keep an eye on generative AI. If you want to use it to solve knowledge problems in the carpeted space, great. But we need to keep it away from the plant floor for now.”

Instead, RoviSys is focusing on training autonomous AI systems to operate alongside workers in a decision-support capacity. Ideally, the system would serve almost as a mentor looking over the shoulder of a less-experienced worker. 

The key to enabling this type of human-machine collaboration is simulation but not first-principles simulation, characterized as modeling processes based on established equations and assumptions of a system. These simulations are based on real-world data in the historian.

 “We're using that data to build data-driven models,” DeBois said. “Now, we can hook the autonomous AI up, and it can learn inside that simulation."

Building the simulation begins with exploratory data analysis, a process that involves auditing the data to ensure there’s enough data to tell the whole story and identify missing measurements from sensors and instruments. The resulting data model doesn’t need to perfectly mirror reality, just close enough for the autonomous AI system to learn, DeBois said.

But the technology is only as good as the humans behind it. Interviews with expert operators are critical for the “machine teaching” phase of autonomous AI development. Over the course of weeks, the systems integrators meet with the experienced operators to understand their process.

“We're saying, ‘Help us understand. Teach us like you would a new operator,’” DeBois explained. 

That information gets built into a workflow that’s delivered to the autonomous AI system. It’s a trial-and-error training process, similar to what a business might experience with a new hire. The system integrators will retrain the system based on feedback from the human operators during the pilot phase, DeBois said. 

“By the end of that, we have mapped out multiple scenarios that a typical operator needs to be able to handle, and we don't go easy on the autonomous AI,” he said.

The AI trainers will give the system challenging tasks to complete, DeBois added.  

“Once it does that, we vet it, [and] we validate it with those experts,” he said. “So we take it into a conference room, we hook it up to the simulator, and we have it go through a typical workday and they're [the operators] watching all the decisions that it's making.”

When the system is finally live, operators are sometimes surprised by the results. DeBois related one situation in which a glass bottle manufacturer couldn’t understand why the autonomous AI system was telling him to adjust a plunger that drops gobs of molten glass during the production process. DeBois and his team asked a 20-year veteran to try it. When he did, the bottle was produced according to specifications. 

"It's completely counterintuitive to everything that he had thought about how the process works, but it knew, the autonomous AI, knew that that was the right way to do it,” DeBois said. “It will find those innovations. It's going to find those changes that no human has ever thought to try.”

When it's on the plant floor, the AI system is disconnected from the internet. It instead operates on a plant control network. The decisions the AI system makes appear on a display next to the HMI, which the operator oversees. The operator can override the system if something doesn’t look right, according to DeBois. 

The integrators can also program constraints that prevent the autonomous AI system from operating outside those parameters. One example might be the introduction of a new product on a line. The system will recognize the new SKU and turn the control back to the human operator, DeBois said. 

In terms of deployment time, the process can take anywhere from eight to 12 months, depending on the scope of the project, DeBois said. Autonomous AI projects are complex, which is why RoviSys typically focuses on projects that will deliver an annual return on investment of $750,000 to $1 million. 

In terms of implementation challenges, DeBois said a common hurdle is a lack of data history or disparate historians. That’s why it’s important for organizations to focus on improving their OT data infrastructure before considering autonomous AI. 

“Those AI projects that you want to get to are downstream of those OT data infrastructure improvements, so do those things. Don't wait,” DeBois said. 

The process also requires some patience. 

“Come into these projects with an open mind,” DeBois said. “These are going to be oftentimes iterative projects. These are projects where people never tried to do these things in the past. These are problems that have been unsolved. And so come into this knowing that this is going to be a journey, and we're going to have to maybe try some different things and that's going to be part of the process.”

RoviSys will be among the companies exhibiting at the eChem Expo taking place at the Meadowview Convention Center in Kingsport, Tennessee, April 8-9. To register, visit www.echemexpo.com.