Unlocking Predictive Power: How Bayesian Networks Enhance Maintenance Decision-Making

Interview with Erik Hostens, Senior Researcher at Flanders Make

At the upcoming Asset Performance Conference, Erik Hostens, researcher at Flanders Make, will demonstrate how Bayesian networks can significantly reshape maintenance strategies in asset-intensive industries. In this article, Hostens walks us through the science and real-world value of applying probabilistic models to predict Remaining Useful Life (RUL), optimize maintenance timing, and manage associated risks and costs.

Flanders Make is a strategic research center that bridges academic research and industrial application, similar in spirit to institutes like Fraunhofer or TNO. Hostens, a senior researcher at the organization, is involved in both hands-on research and the coordination of industrial projects. “Our mission is to enhance the competitiveness of Flemish companies through innovation,” he explains. “We work closely with universities and industrial partners, translating scientific advancements into practical tools for companies.”

Understanding Remaining Useful Life

At the core of Erik’s work lies a deceptively straightforward question: when will a component fail? But as Hostens emphasizes, predicting asset failure is far from trivial. “We talk about Remaining Useful Life, or RUL, which is the time left before a component becomes non-functional or underperforms. It’s not just about knowing how long something will last—it’s about understanding how confident you are in that estimate.”

Traditional maintenance strategies often rely on statistical averages provided by manufacturers—such as L10 or L50 lifetimes under ideal conditions. But field data tell a more complex story. “In practice, components are usually replaced before failure. That means the real moment of failure is often missing from the data. We call this censored data,” says Hostens. “But even if you don’t see the failure itself, the degradation history leading up to it is extremely valuable.”

Bayesian Networks: A Holistic Approach

That’s where Bayesian networks come in. “A Bayesian network is a graphical model that shows probabilistic relationships between different variables,” he explains. “You start with something as intuitive as a whiteboard drawing—linking load, temperature, vibration, and so on—and you quantify the relationships between these variables. Eventually, this model allows you to calculate the probability that something will fail within a certain time frame, under specific conditions.”

Unlike traditional approaches that treat statistics, condition monitoring, and degradation models separately, Bayesian networks combine all this information. “We’re not replacing existing models like Weibull distributions or machine learning algorithms,” he clarifies. “We’re integrating them. For example, neural networks can be used to identify features in vibration data, and those features feed into the Bayesian model.”

A Real-World Case: Solenoid-Operated Valves

Hostens will illustrate this with a case study on solenoid-operated valves. These components were tested in the lab, some up to one million cycles, capturing operational and failure data. The Bayesian model developed from this dataset was able to predict RUL based on real-time measurements. “And then we layered a cost model on top of it,” he adds. “What’s the cost if you replace the component too early? What’s the cost if it fails unexpectedly? What’s the cost of a service visit? These are all stochastic variables—they have their own probabilities.”

Supporting Real-World Maintenance Decisions

What makes this approach especially relevant for technical maintenance managers is its applicability to complex, real-world decisions. “Imagine you manage a fleet of machines. You have limited resources and need to schedule maintenance visits efficiently. You want to avoid downtime but also avoid unnecessary replacements. Our model supports that decision-making process in an objective and data-driven way.”

The use of Bayesian networks also supports opportunity maintenance—planning multiple interventions in a single visit to reduce overall costs. “It quickly becomes very complex,” says Hostens. “If you’re replacing one component, should you also replace another that might fail soon? How do you account for lead times and logistics? The model calculates all of that based on the current state of the equipment and the cost scenarios.”

While Bayesian networks may sound technically daunting, Hostens emphasizes their accessibility. “They allow engineers who aren’t data scientists to still leverage complex statistical insights. And that aligns with our philosophy at Flanders Make: knowledge transfer. We don’t sell software—we help companies build the capabilities to do this themselves.”

What to Expect from the Conference Session

What can attendees expect to learn from his session at the Asset Performance Conference? “First, they’ll understand how to use all available information—sensor data, statistics, operational conditions, and cost models—to support maintenance decisions. Second, they’ll see how Bayesian networks offer a practical and transparent way to combine this information. And third, they’ll leave with a realistic idea of how to start applying this approach in their own organization.”

Hostens’ session is especially relevant for reliability engineers, maintenance managers, and asset strategists looking to move beyond fixed schedules and reactive maintenance. As industries continue to digitize and optimize, the ability to predict and plan with greater accuracy is not just a technical advantage—it’s a competitive necessity.

Join Erik Hostens at the Asset Performance Conference to explore how probabilistic modeling can bring clarity to complex maintenance challenges, and help your organization make smarter, more cost-effective decisions.

Erik Hostens,
Senior Researcher at Flanders Make