Simulated sawmill reveals where AI falls short in industry
In a simulated sawmill at RISE, AI agents were tasked with monitoring a timber drying kiln. When a moisture sensor gradually began giving incorrect readings, none of the models managed to identify the fault. The experiment shows why AI needs more context and clear boundaries before it can be allowed to act in industrial settings. Petri Luomala and Joakim Eriksson talk about the work.
What does it take to bring AI into industry?
“In industry, safety and security also encompass people’s health and safety and the environment. It is important to understand the consequences a failure can have in these settings,” says Petri Luomala.
He describes how predictive AI models are already used to detect signs of impending machinery failures. Allowing an AI agent to act on a prediction, however, is a bigger step. Slowing down or stopping a machine can affect other parts of the process. How do we know the action is safe? Who is responsible if something goes wrong?
Introducing these solutions requires several kinds of expertise. IT teams know the networks and systems. Operational technology (OT) teams understand the machinery and industrial processes. Suppliers provide equipment, software and maintenance.
So how can these teams work together in practice? That was one of the questions raised when RISE brought several sawmill companies together in Borås. The meeting laid the groundwork for OTTO (Operational Technology for Reliable Interoperability). In the project, RISE, industrial companies and technology suppliers are working together to connect industrial control systems to modern IT systems without compromising safety, security or day-to-day operations.
Is AI ready for industry?
Yes, but there is more to do. Before AI can take action, we need to be able to trust that it interprets information in the right context.
“An agent is exactly as good as what its tools let it see, and exactly as dangerous as what its tools let it do,” says Joakim Eriksson.
AI needs to understand the context
Giving AI access to large amounts of measurement data is not enough. A temperature of 50°C might make sense when drying timber, but be completely wrong in a fridge. To tell the difference, the system needs to know what the value refers to.
“Data needs to describe itself: what is being measured, in which unit and where the sensor is located,” explains Petri Luomala.
This is where metadata and semantics come in, areas RISE is working on through Semantic Data Readiness. Metadata can describe which sensor a measurement comes from, when it was taken and whether the temperature is in Celsius. Semantics make the meaning and relationships explicit: that the sensor is in a fridge storing food, for example, and which temperatures are appropriate there.
With that context, AI could be better equipped to raise the alarm when something seems wrong. Understandable data also helps the people working with these systems. When IT specialists, operators and suppliers can interpret information in the same way, it becomes easier to collaborate and identify misunderstandings.
Test before allowing AI to act
But how do we know AI will respond correctly when something actually goes wrong? Simulators allow researchers to introduce faults and disruptions deliberately and examine how systems handle them, without putting ongoing production at risk.
One example is a simulated sawmill and timber dryer built by Joakim Eriksson. AI agents were tasked with monitoring operations, diagnosing faults and requesting actions within defined limits. In one experiment, a humidity sensor was made to give increasingly inaccurate readings. No alarm was triggered, yet the timber was damaged in the simulation. None of the models tested managed to diagnose the fault.
Experiments like these can reveal limitations that might otherwise be difficult to spot. They also make it possible to investigate which information and tools AI needs to perform better, ahead of further evaluation in a real industrial environment.
The actions AI is subsequently allowed to take need to be governed by clear permissions, safety limits and responsibilities.
From the cloud back down to earth: a growing trend in industry?
“For a while, everything was supposed to move to the cloud. Now some are starting to bring it back down to earth,” says Petri Luomala.
He and Joakim Eriksson see growing interest in handling data and running digital solutions locally. One example is sawmill company Fiskarheden, an industrial partner in OTTO, which is investing in its own data centre and local operations to strengthen its resilience.
This is about both cybersecurity and availability. An external service can be attacked or stop working. An internet connection can fail. Functions that are critical to safety and operations still need to work.
Running systems locally can reduce these dependencies, but also requires in-house expertise, security and maintenance. Each organisation therefore needs to decide which functions must run close to the process and how they should behave during an outage.
Introducing AI safely into Swedish industry requires technology experts and those who understand industrial processes to work together. As Joakim Eriksson emphasises, an agent’s capabilities depend on the information and tools it can access. A warning from an AI system can give an operator time to intervene before a machine breaks down. But what happens if the system is allowed to stop the machine on its own? In RISE’s simulated sawmill, researchers can test how AI responds to faults and what actions it proposes. The results can help industrial companies decide when to let AI intervene and when the operator needs to make the call.
Example of how RISE contributes to safer AI in Swedish industry
At RISE, research in AI, cybersecurity and industrial processes comes together to develop safe and reliable AI solutions. The focus is on protecting both operations and the people working in production.
Two examples of what we do:
OTTO is a project in which RISE and industry partners are exploring how industrial control systems can work together with modern digital solutions without compromising safety or operations. The project also addresses how responsibility should be shared between IT, operational technology and suppliers.
In the Semantic Data Readiness project, methods and ways of working are being developed to make industrial data understandable to both people and machines. How does the AI know how to interpret the data? Clarifying what data actually means, and in which context it should be used, is crucial for developing reliable AI that makes the right decisions.