Researchers at Ghent University use an approach called Bayesian inference to help study how engineers control plasma inside future fusion devices. Now, the very same strategy is helping frozen food companies optimise their production lines.
Mathematical thinking
Different teams around the world are working hard to build experimental fusion reactors. But not all fusion research involves megasites and hardware. Some of it relies on probabilities and maths.
“Ghent University does a lot of research related to nuclear fusion,” says Tomas Van Oyen of PropheSea, a company that recently worked with researchers from the Belgian university. “They use particular methodologies to observe the state of the sensors and also the state of fusion experiments. This methodology is based on Bayesian theory.”
A Bayesian approach is a way of reasoning about uncertainty. You start with a prior belief, update it as new evidence comes in, and arrive at a revised belief that reflects both what you already thought and what the new data tells you. In a real-world scenario, such as a fusion reactor, this can be applied to a system where data is collected to assess the likelihood of a change in the system – and whether that change represents an anomaly.
“Based on this probability distribution, you can actually gain information on whether the state you are measuring is normal or anomalous,” Tomas adds. “You can exploit this methodology, this way of thinking, towards anomaly detection.”
Researchers at Ghent University use Bayesian modelling to detect anomalies in the magnetic fields used to control and stabilise plasma inside a tokamak. This approach is particularly useful when sensor data is limited or uncertain, allowing researchers to estimate the state of a system from incomplete information. With only a small number of fusion devices operating around the world, such approaches can help researchers study fusion systems without having a reactor operating next door.
“They use this methodology a lot within their research because they do not have enough sensor data to completely cover all of the parameters,” adds Tom Neels, from Ikologik, a Belgian company developing manufacturing software that brings together production processes and operational data to improve how factories are managed and run.
Tom saw potential in applying this methodology outside the realm of nuclear fusion. If you could detect anomalies in a complex system using this approach, then other industries could benefit too.
Beyond fusion research
PropheSea and Ikologik jointly applied to use the Bayesian methodology developed through Ghent University’s fusion research in a pilot project with a pioneer customer: Ecofrost, a food processing factory in Belgium.
Ecofrost operates in the frozen fried potato business. The company cleans, cuts, blanches, freezes and packages potatoes so they are ready to fry when they reach the customer. Throughout this process, the factory uses a wide range of resources. Therefore, it’s important that the company monitors how each one is being used and ensures that no energy or water is wasted.
“Ecofrost is a fast-growing business and they wanted to fine-tune their production line. One of the things where they sometimes lose money is detecting things like water usage or other issues in production. They were detecting these things too late,” explains Tom. “If they could have something running that can find the anomalies, that would be a great opportunity for them.”
In theory, having a sensor collecting data at every step of this long process would be the ideal solution. But that isn’t entirely feasible. Tom explained that it would be very costly to build specific software to collect data from every type of appliance in Ecofrost’s process. So they considered another approach – and this is where the fusion research came in handy.
Using the same Bayesian approach developed by the fusion researchers, Ikologik and PropheSea created software that uses data collected by a limited number of sensors across Ecofrost’s production line. The system then infers the state of different parts of the process while accounting for uncertainty. As it receives more observational data from the sensors, it assesses how likely it is that a particular state has changed – and whether that change represents a problem worth reporting.
Much like researchers at Ghent University don’t need (and can’t realistically have) every data point describing how plasma behaves in a fusion reactor to identify potential anomalies, Ecofrost doesn’t need a large number of sensors covering every part of its processing line. With this approach, the inference system can use the available data to identify anomalies without requiring a sensor at every point in the process.
“We collect, I think, between ten and twenty thousand different parameters throughout the factory. So, individual values of, for example, the status of a machine – on or off – a measuring point, temperatures, etc.,” Tom adds.
Sensors collect thousands of different parameters throughout the factory. Copyright: Ecofrost
Significant savings
All of this helps Ecofrost manage its utilities more efficiently and save on energy and water. When using this approach, the company achieves significant savings in water consumption and electricity.
“The consumption of water is very high in Ecofrost’s factory, it’s really huge. Having a system that lets them predict something as simple as whether all taps of water are properly open and properly closed, for example, is already more efficient and leads to savings,” Tom explains.
So far, the technology has only been applied to this specific case in food processing, but Ikologik and PropheSea would like to take it further. “We’d like to test our solution with other datasets,” Tom concludes. “And with other datasets, we mean other data points coming from sensors in production lines in different factories and eventually rolling it out to other industries.”
This work is part of EUROfusion’s efforts to encourage the use of fusion research beyond the fusion sector itself. In this case, a methodology developed to deal with limited and uncertain data from fusion experiments has found an application in a very different setting: a frozen food production line.
More projects are underway to explore how technologies and methodologies developed for fusion can be applied in other areas and, ultimately, help make everyday life easier. The current call is open for applications until early 2027.