Predictive Maintenance: From Guesswork to Informed Decisions

Keeping industrial assets healthy is a constant challenge. Whether you operate a single production line or manage a fleet of machines across multiple sites, the stakes are high. Unplanned downtime leads to lost production, costly repairs, and operational stress. To stay competitive, companies increasingly rely on diagnostics, condition monitoring, predictive maintenance, and estimates of remaining useful life (RUL). Yet in practice, maintenance decisions are rarely straightforward.

Every day, maintenance managers juggle complex decisions. Which machines need servicing first? How can we avoid unexpected breakdowns without replacing parts too early? How do we plan maintenance across different sites while keeping costs under control? We aim to provide tools that bring clarity to these decisions. By combining data from multiple sources and quantifying uncertainty, we support companies in moving from reactive fixes to proactive maintenance strategies. 

Why Bayesian Networks?

Machine health is influenced by many interacting factors: load, temperature, operating conditions, usage patterns, and environment. At the same time, maintenance decisions depend on different types of information, such as sensor data (e.g. temperature, vibration, torque), inspection logs, degradation models, and cost considerations.

Bayesian networks combine these factors and represent the relationships in a structured way. They also manage uncertainty with transparency. Rather than producing a single point prediction, the model delivers probability distributions and confidence levels. It allows users to understand not only what might happen, but also how certain (or uncertain) that prediction is.

Our proof of concept demonstrates that this integrated approach can support more informed maintenance decisions: reducing unnecessary interventions and avoiding costly failures.

From Prediction to Decision Support

Many predictive maintenance solutions focus on predicting when a component is likely to fail. That is valuable, but it is only part of the picture. Our toolbox aims to go one step further by embedding cost considerations directly into the reasoning process. Using cost-sensitive learning, the model can evaluate different maintenance actions and identify the moment where intervention makes the most sense from an operational and economic perspective.

Our proof of concept shows how this approach can help to:

• Prioritise interventions across an entire machine fleet
• Balance downtime risk against unnecessary early replacements
• Make better use of existing data, without requiring a large in-house data science team

Importantly, the system is designed to support human decision-makers, not replace them. Final decisions remain with maintenance and operations experts.

Real-World Use Cases

Jan De Nul – Dredging Equipment

Cutter suction dredgers operate in extreme and highly variable conditions. A rotating cutter head breaks up soil using high-powered cutting teeth (pickpoints). Depending on whether the soil is sand, clay, or rock, wear on these pickpoints can vary significantly.

During operation, the condition of the pickpoints cannot be inspected directly. The cutter head must be lifted out of the water, which already causes downtime, after which operators perform a visual inspection. Decisions to replace pickpoints are based largely on experience and judgement, often leading to early replacements to avoid future interventions or prevent damage.

Using sensor data and historical inspection information, we are now developing a tool to estimate wear and recommend when it is worth lifting the cutter head for inspection or maintenance. The goal is to balance inspection time against the risk of damage and extended downtime.

Based on our proof of concept, we believe that this approach can result in fewer unnecessary inspections – cutting costs and reducing overall downtime.

Offshore Wind Turbines

Maintenance planning for offshore wind farms is a logistical challenge. Weather windows, vessel availability, and long lead times all contribute to high costs. While condition monitoring can support predictive maintenance, equipping every turbine with extensive sensing is often too expensive and impractical.

Our model focuses on corrosion monitoring as one of the most critical failure modes in harsh marine environments. It integrates fleet-wide SCADA data with detailed measurements from a few highly instrumented “fleet-leader” turbines, along with weather data, to model degradation and provide reliable monitoring and prognosis of corrosion across the entire wind farm.

Our proof of concept demonstrates a cost-effective approach to fleet-wide monitoring and prognosis, showing clear potential for improved maintenance planning, extended asset lifetime, and better control over operational risk.

Get in touch to discuss your use case

Early results clearly indicate that combining Bayesian reasoning, engineering knowledge, and cost-aware decision support can help companies save time and effort in maintenance planning.

Together with end users we are now focussing on validation, industrial testing and refinement. Interested in exploring how this approach could fit your assets or fleet? Get in touch to discuss your use case.