In most industries, maintenance is a waiting game. Things are fixed when they break. But in the 21st century, an age defined by data and automation, that approach no longer makes sense. The solution could be predictive maintenance. This is an approach that uses sensors and software to analyse equipment performance in real time and predict when it might fail.

Edward Khomotso Nkadimeng, a lecturer and researcher in artificial intelligence and data systems in nuclear/particle physics at Stellenbosch University, has researched how a predictive maintenance model can help keep critical systems running – from research equipment to national infrastructure. He explains why this approach could be a practical tool for resilience across Africa.

What is a predictive maintenance model and why did you build one?

For decades after the global industrial boom, many industries relied on a simple rule: wait for a machine to break, then repair it. That made sense when machines were simpler and downtime was just part of the routine.

Periodic maintenance is common too, but still inefficient and often based on time, not actual machine condition. That approach costs time, money, and sometimes even safety. Modern systems are more interconnected and expensive to halt.

A predictive maintenance model is a data-driven system that forecasts equipment failure before it happens. It predicts when systems are degrading, rather than just reacting. It monitors a variety of systems, from industrial pumps, compressors and turbines to scientific instruments, by collecting real‑time data like vibration (which measures how much a machine physically oscillates), temperature, pressure and voltage.

These measurements come from Internet of Things (IoT) or condition-monitoring sensors. Even machines that aren’t ultra-cutting-edge can be instrumented to provide this data. Once collected, the data feeds into machine learning models that learn to recognise patterns associated with slow drift towards failure.

The model monitors a broad range of systems: industrial pumps, compressors, turbines, and high-precision scientific instruments (cyclotrons, vacuum pumps, beamline diagnostics). It is designed for systems where sensor data can be collected – any instrument that generates measurable signals. It uses live data vibration, the physical oscillation of a machine component, where subtle changes in vibration amplitude or frequency often precede mechanical failures, such as bearing wear or rotor imbalance, as well as temperature, pressure and voltages.

While advanced machines may produce richer data, even legacy machinery can benefit with added sensors. The method is therefore broadly applicable to recognise when they’re slowly drifting towards failure.

At NRF-iThemba LABS, a South African national nuclear and accelerator research facility, and Stellenbosch University, I built a system like this out of necessity. Our teams include physicists, engineers and computer scientists who collaborate on high-precision experiments in nuclear and particle physics.

The research instruments are complex, expensive and often one of a kind. When they fail unexpectedly, experiments stop, data is lost, and public funds go to waste. For example, we work with 70 MeV cyclotrons for isotope production, superconducting magnets, radiofrequency acceleration cavities and vacuum systems. These are one-of-a-kind instruments, sensitive to downtime.

So, the goal was to make an affordable, self-learning system that can scale from our research equipment to the industrial infrastructure that keeps African economies running pumps, turbines and power grids. Similar predictive maintenance systems are applied in industrial power plants, water utilities and aviation, reducing unplanned downtime by 20%-40%. Our adaptation for African labs and industrial systems uses low-cost Internet of Things sensors with cloud-based AI.

What did you learn from the model? Why is this useful?

The first thing I learned is that machines whisper before they scream. Long before a breakdown, they show tiny signs like slight vibrations, small voltage drops, or subtle changes in speed.

With enough data on vibration, temperature, pressure, voltage and motor load, for example, these data streams form the input for AI models. These patterns form a kind of language, and artificial intelligence becomes the translator.

By training the model on real operational data like pump vibration over time and other readings, we discovered that failures aren’t random: they follow recognisable signatures. Once the system learns these patterns, it can predict what’s coming and even suggest what to do next. The real benefit is timing, scheduling maintenance exactly when it is needed and not too early, which wastes parts and labour, and not too late (which risks catastrophic failure).

Instead of over-servicing equipment or waiting for something to fail, maintenance can happen exactly when it’s needed. That saves resources, reduces downtime and keeps operations running smoothly. And because the principle is universal, it applies just as well in factories, hospitals and water systems as it does in research labs. For example, detecting a failing motor before a line shutdown in a manufacturing plant, or ventilator sensors predicting pump failure in a hospital, or monitoring municipal pumps to prevent water shortages.

What are the practical implications of applying the model?

The practical impact is huge. Predictive systems help avoid blackouts, water shortages and unplanned shutdowns – issues that affect daily life and essential services. An example can be seen in South Africa’s blackouts: the power utility Eskom’s transformers are monitored for predictive faults. In Cape Town, predictive maintenance of water systems reduces pump downtime. They also make workplaces safer and budgets more efficient.

For African countries especially, where technical resources are often stretched, predictive maintenance is a form of resilience. It replaces firefighting with foresight. By using affordable IoT sensors (small devices collecting data like temperature), cloud-based AI (online software that analyses this data in real-time), and self-learning algorithms, maintenance becomes continuous, automated and smart.

It’s the quiet side of AI, keeping the lights on, the pumps running and the economy stable. Physics, data and engineering can quietly work together to keep important systems alive and reliable.