Using AI to Capture Industrial Expertise

Summary

Industrial “brain drain” is real. As skilled workers retire, they take their valuable operational technology (OT) expertise with them. The manufacturing industry is losing decades of experience at an alarming rate as personnel who know how to run plant processes are taking that know-how with them into retirement.

Several statistics from LNS Research Associates show dramatic changes in just a few years. In 2019, the average tenure of manufacturing sector workers was 20 years and the average time in a position was seven years. The average three-month retention rate of new manufacturing personnel in 2019 was 90 percent. As of 2023, the average tenure of manufacturing workers was three years and the average time in a position was only nine months. The average three-month retention rate of new manufacturing personnel dropped to 50 percent that year.

Bryan DeBois, director of industrial AI at RoviSys, a global automation systems integrator, cited these statistics in a virtual presentation associated with the Los Angeles section of the International Society of Automation. “We are losing the expertise that made us great in manufacturing in the United States at a rate that we are not replenishing,” he said. “This is knowledge about how to run and maintain our equipment that we may never get back. It’s critically important that we try to capture this expertise. AI [artificial intelligence] is a way we can do it.”

DeBois explained the different types of artificial intelligence and demonstrated how autonomous AI, through the process of machine teaching, can capture and preserve critical knowledge before it’s lost, “empowering manufacturers to retain and scale their most valuable skills.”
 

Different types of AI

DeBois explained the differences between the three types of industrial artificial intelligence in use today and why they are not equal (Figure 1). Traditional AI, autonomous AI and generative AI differ based on the machine learning (ML) algorithms they use, making each good for different applications.

Traditional AI

Traditional AI algorithms have been in use for 15 to 20 years and come in two subcategories: unsupervised and supervised learning. Unsupervised learning is anomaly detection. An unsupervised machine learning model associated with a process can learn what normal is and, over time, detect and report when the process becomes abnormal. However, this type of AI can’t determine why things are becoming abnormal.

Supervised learning is when an ML model is trained based on large volumes of very clean, very correlated data. Then that ML model can learn to predict a single value. Concerning predictive quality, supervised learning can predict the final quality of a certain batch well before that batch is completed; however, it will predict only one value.

Also built into supervised learning is the assumption that someone knows what to do with that prediction. If an operator can predict (with the aid of traditional AI’s supervised learning) with 95 percent certainty what the final quality of a specific batch will be, the operator can predict that the batch will be off-spec. If known soon enough, steps can be taken to bring the batch back into spec.

Autonomous AI

In addition to recognizing a problem, autonomous AI knows how to fix the problem by using human-like reasoning to make human-like decisions.. Autonomous AI is based on a different type of ML technology called “deep reinforcement learning.” It can act, not just predict. It can continuously learn from input data and past experiences. “In this sense,” explained DeBois, “autonomous AI’s meaning can also stand for solutions that are self-learning and have the ability to become more efficient over time.”

Most autonomous AI solutions use a range of technologies that include ML algorithms, deep learning and physical sensors such as cameras, microphones and scanners to gather data, generate insights, and perform actions independently. While no two autonomous AI solutions are the same, DeBois said they all share core characteristics:

• They can automate complex tasks.
• They don’t require human input to perform tasks and process data.
• They can continuously learn and improve task performance by generating insights from real-time and historical data.
• With the right data, they can be trained to complete tasks with 100 percent accuracy.
• Automated operation means they are available to complete manual tasks 24 hours a day.

An autonomous AI agent is a neural network built on deep reinforcement learning. It can build a long-term strategy. Through a machine-teaching process, expertise can be captured within the agent. DeBois said that autonomous AI brains, or agents, can be useful for dynamic and highly variable systems where traditional control systems have struggled to manage the process due to excessive variability.

“We tend to have a human being in the loop [for those highly variable processes]. Brains or agents adapt to a wider range of scenarios,” DeBois said.

The scenarios autonomous AI agents can adapt to include advising on competing optimization goals or strategies, production scheduling, adapting to unknown starting or system conditions and/or adapting to unknown system conditions. Here are some questions autonomous AI agents can answer:

• Competing optimization goals or strategies. I can make this today, but will I be able to make this tomorrow? Given this mix of raw materials, what can I make out of it and what should I make out of it? “We can train these agents to consider things like market signals that say, ‘You should make more of this because that’s what the market is demanding,’” said DeBois.
• Production scheduling. What’s the best way to optimize a production schedule? “Those are the types of decisions that human beings are making today. We can build an autonomous AI agent to make those decisions in the same way,” explained DeBois
• Unknown starting or system conditions. In certain batch food product manufacturing, bad product will be made for a while after startup until an expert operator with many years of experience adjusts the process until good product is made. However, when the next batch is started, the materials will likely have a different profile, so more bad product will be made until the operator makes more adjustments. An AI agent can make the adjustment based on other raw materials.
• Unknown system conditions. DeBois explained that sometimes you can’t directly measure exactly what’s happening inside something like a polymer reactor because there’s no physical way to do it. The best operators have to use proxy measures, things like changes in pressure and temperature, and from that data build a mental model of what’s happening inside. “And of course, there’s operator overload,” DeBois said. “That’s an unknown system condition and we can build an AI brain that can do that same thing. These brains never go to sleep, they never call off and they never take vacation.”

 
Generative AI

Generative AI has received a lot of attention lately given the rise of ChatGPT and all of the image- and video-generation tools that have emerged. Generative AI is a form of large language model (LLM) that excels at natural language processing, but it can’t reason or analyze. “While generative AI looks like it’s doing reasoning behind the scenes, it’s just very good at stringing words together,” DeBois said.

In October 2024, Apple exposed the limitation of LLM models, explained DeBois. “Apple’s study showed that LLM models can’t do reasoning. What we recognize as reasoning in LLM models is simply pattern matching.

The study found that there is no way that LLM models were able to reason about simple mathematical theories that even a grade school student would be able to solve,” he said.

“The future will be not just generative AI agents by themselves, but a combination of the natural language capabilities of generative AI married to the decision-making capabilities of autonomous AI. That’s what the future of AI agents should and will look like.”
 

Caution

DeBois cautions that a problem with generative AI in the industrial space is hallucinations. These are when “generative AI steps over its skis when it reaches too far, and it basically makes things up out of thin air. The problem with that is that you can’t recognize the difference between real life and a hallucination unless you already know the answer,” he said.

“Many AI vendors are talking about bringing generative AI into the plant floor to tackle the problem of losing all of our expertise around maintaining equipment on the plant floor,” DeBois said. But, ”using generative AI on the plant floor can have disastrous results.”

DeBois used a hypothetical scenario to explain: Some traditional AI vendors say, give us all of your standard operating procedures, equipment manuals and maintenance records for the last 10-plus years and we’ll create a chatbot for you. Then you put your least-experienced maintenance person in front of that chatbot and they say, “This piece of equipment is down. Do you know how to fix it?” Generative AI will say, “Absolutely, I know how to fix it: You do this, you tweak this, you torque this bolt, and you rev it back up.”

And you could blow up the plant, DeBois said: “You could kill somebody. It’s not appropriate to use generative AI on the plant floor.”

“We have to give it time,” DeBois continued. “[Some generative AI vendors] will solve this problem. In many ways, generative AI is already starting to solve problems on the carpeted space of those manufacturing companies. And that’s great. Use it to solve the knowledge problems in the carpeted space but keep it away from the plant floor. It is not ready.”
 

AI from a different perspective

ISA Fellow Jonas Berge, senior director of Applied Technology at Emerson, provides additional perspective in a guide to industrial AI and digital transformation.  The scope of industrial automation and control systems (IACS) includes distributed control systems (DCS), safety instrumented systems (SIS), machinery protection systems (MPS), manufacturing execution systems (MES) and energy management information systems (EMIS), just to name a few, Berge said. “Industrial AI plays an important part in all automation and control systems to support production and maintenance, as well as in the deployment of the automation and control systems themselves.”

Berge puts the technologies used in industrial AI tools into four broad categories: causal AI, machine learning, deep learning and generative AI (GenAI). “It should be noted that none of these are better than the others for all tasks,” he said, “just like a hammer is not better than a spanner for all jobs.

Each AI technology has specific strengths and suitable applications:

• “Causal AI embeds already well-established first principles of physics and chemistry into models showing how processes and equipment work and embeds already well-known cause and effect relationships describing how they fail into ‘agent’ functions. These types of models and agents are built into ready-made apps that quantify efficiency and predict problems,” Berge noted.
• “Machine learning in its training phase uses one of many statistical algorithms and applies it to historical data from the plant, such as process history, maintenance records and lab results, to find correlation for normal operation and abnormal conditions like failure, process upsets and quality issues. In the operational ‘inference’ phase, these correlations are then used in agent functions to predict abnormal conditions. There is also deep learning, which is a subset of machine learning based on neural networks,” he said.
• Generative AI uses LLM-type artificial neural networks (ANNs) such as generative pre-trained transformers (GPT), which are statistical algorithms. These systems are pre-trained on existing software code to generate new code based on prompts,” Berge explained.

Getting started

DeBois recommends starting with use cases and specific AI algorithms. “Use cases should drive every digital transformation project, never technology,” DeBois said. “Use cases lead; technology follows. Start small and build on your successes.

“Build a snowball of ROI as small wins lead to big wins. We have to realize that these are iterative projects. We’re solving unsolvable problems and it’s likely no one has ever tried to solve this with AI.”

_{This feature originally appeared in the March 2025 issue of Automation.com Monthly.}