🤖 Closing Knowledge Gaps with AI: Realities in Industrial Facilities!

[Hasan S. Cemkan]

🤔 Where Does AI Fit In?​

As the Atlas Prediction Control team, we spend a lot of time in plants calibrating advanced process control systems that have gone out of tune over the years. During this process, a question we frequently encounter in control rooms is: Where exactly does Artificial Intelligence (AI) fit into this equation?

📊 Survey Results Were Surprising!​

We asked this question to engineers, process managers, and control room personnel, and obtained striking results from four live surveys:

• AI Use is Widespread, But in the Wrong Place: 71% of participants stated they use AI for productivity tasks. However, this use was in administrative tasks and personal learning areas, rather than daily operations. AI use in critical applications like predictive maintenance or soft sensors was almost non-existent.
  
• Real Knowledge is Held by People: When asked where the real knowledge about plant operations resides, the majority (62%) said it was "in a person," not in a system. Experienced personnel carry far more operational truth than documentation. When a problem arises, the first step is not to look at procedures, but to open the trend screen and find someone who has encountered a similar situation before (53%).
  
• Hidden Knowledge Sources: This situation indicates that the real knowledge base in most plants is distributed in "tribal knowledge" and historical trend data, not in written documents. An AI trained with general manuals misses the plant's unique, often unwritten, information.
  

🔒 Deployment Issues and Solutions​

We found that users do not doubt the value of AI, but have concerns about data security and not knowing where to start (53%). Cost and trust were also significant barriers. This is a "deployment" problem, not a "convincing" problem.

• Local Solutions are Essential: This means the model must run on the plant's existing hardware, within OT (Operational Technology) boundaries. Thus, proprietary data never leaves the building. Cloud-based tools receive a "no" answer in most operational control networks.
  
• Purpose-Built AI Agents: Local AI agents designed for control rooms are different from general chatbots. Instead of a single model interpreting many documents, a design that directs a question to specific, targeted sources like P&ID, OEM manuals, or historians is more effective. This returns a specific piece of information from each source, providing precise answers instead of general text dumps.
  

🎯 Accuracy and Security are Priority​

Two fundamental rules emerge from the real needs of plants:

1. No Hallucinations: A system operating close to the control loop cannot "hallucinate." It must be accurate or clearly state that it does not know.
   
2. Local and Air-Gapped Operation: The system must operate locally or air-gapped to resolve cybersecurity objections. This is the only way to address security concerns without sacrificing the reasoning power provided by AI.
   

When considering such a tool for your own plant, think about these questions: Can it show its work, and does it need to leave your network to do so?

In conclusion, closing knowledge gaps using AI agents begins with understanding where knowledge about plant operations resides. At Atlas, we are working on precisely this problem, and we believe it is possible by carefully considering our current environment and user needs.