How Does a Machine Know It’s About to Fail?

In Part 2 of our predictive maintenance series, we followed a simple scenario: condition data suggests that a bearing on a critical production asset will probably require intervention within approximately 30 days.

That early warning gives Maintenance valuable time to identify the right spare part, check availability and prepare the intervention. But where did those 30 days come from in the first place?

A machine cannot tell you that it is starting to fail, but changes in its physical behaviour can provide early indications that something is wrong. A component may begin to vibrate differently, generate more heat or show other signs of wear before its function is actually compromised.

Condition monitoring captures these changes. Predictive maintenance goes a step further by interpreting the available information to assess what may happen next.

Machines Leave Clues Before They Fail

Let's return to our bearing. A healthy bearing has a characteristic operating behaviour, but as it begins to deteriorate, that behaviour can change well before the bearing stops performing its function.

Vibration is one of the most familiar examples. Changes in vibration patterns can indicate developing problems in rotating equipment, while temperature, lubricant condition, acoustic signals or electrical measurements can provide additional information depending on the equipment and potential failure mode.

This does not mean that manufacturers simply need to install as many different sensors as possible. Different failure modes produce different physical effects, so the relevant question is which information can reveal the type of deterioration you are trying to detect.

This principle is reflected in ISO 17359, the international standard providing general guidelines for machine condition monitoring. It covers parameters including vibration, temperature, tribology, contamination, power and speed, and connects the choice of monitoring approach with the failure modes being investigated.

The starting point, therefore, should not simply be which sensors can we install?, but rather what could fail, and what evidence would that failure leave behind?

A Sensor Doesn't Predict Failure

We introduced this distinction in Part 1, and it is worth returning to because it is easy to overlook when discussing predictive maintenance.

A vibration sensor measures vibration and a temperature sensor measures temperature. Neither measurement, on its own, tells you that a bearing will require replacement in 30 days. The useful information comes from understanding how those measurements are changing over time and what those changes mean in the context of the equipment.

For example, today's vibration may differ significantly from the machine's normal behaviour or may have been increasing gradually for several weeks. The pattern might also resemble behaviour observed before previous bearing failures. At the same time, operating context matters: a change in vibration caused by deterioration is not necessarily the same as a change caused by a different load or operating speed.

Condition monitoring, diagnosis and prediction are therefore connected steps rather than interchangeable concepts. Sensors provide the measurements; analysis helps determine whether the changes are meaningful; predictive maintenance uses that evidence to assess what is likely to happen next.

This Is Where AI Becomes Useful

This is also where AI can add value. Modern industrial assets can generate large volumes of condition data, and across a production network that can quickly mean analysing information from hundreds or thousands of assets.

Machine-learning models can help identify patterns and anomalies in this data and continuously update assessments as new information becomes available. A 2025 systematic review by Abdulrahman Alamin and his co-authors analysed 60 peer-reviewed predictive-maintenance studies published between 2020 and 2024. The authors found machine learning being applied to areas including fault detection, diagnosis and Remaining Useful Life estimation, while also highlighting ongoing challenges around data quality, interpretability, scalability and real-world deployment.

AI therefore does not remove the need for engineering knowledge or reliable condition data. Its value lies in helping organisations analyse more information, more frequently and across more assets than would be practical through manual analysis alone.

Does Every Machine Need More Sensors?

Not necessarily. More sensors do not automatically lead to better predictive maintenance, and the appropriate monitoring strategy depends on the asset, its criticality, likely failure modes and whether detecting deterioration earlier would actually create useful time to act.

ISO 17359 reflects this broader approach by considering factors such as equipment criticality, failure modes, operating conditions and the feasibility of different monitoring methods when establishing a condition-monitoring programme.

For some critical assets, continuous monitoring may be justified. For others, periodic measurements, inspections or information already available from existing equipment may provide what Maintenance needs. The objective is not to collect the maximum possible amount of data, but to collect information that supports better decisions.

From Machine Signal to Maintenance Decision

Let's return once more to our bearing. Over time, its vibration behaviour changes sufficiently to warrant investigation. Analysis of the available condition data suggests that the bearing is deteriorating, and as further measurements reinforce that assessment, Maintenance concludes that an intervention will probably be required within approximately 30 days.

At this point, a technical signal has become operational information. And that information immediately raises questions beyond the condition of the machine itself: Which bearing will be needed? Is it available? How long would it take to source? Could the expected intervention also influence how that spare part should be planned in future?

We explored the first questions in Part 2. Later in this series, we will look more closely at the last one: how information about equipment condition could contribute to more dynamic spare parts planning. Knowing more about when a component is likely to be needed becomes particularly valuable when deciding whether that component should already be available.

The Sensor Is Starting to Move

So far, we have largely imagined condition data being collected by sensors attached to individual machines. Increasingly, however, manufacturers are also exploring ways to bring the monitoring technology to the equipment rather than installing dedicated monitoring hardware everywhere.

Drones, computer vision and autonomous inspection robots can collect condition information across different assets and reach areas that may be difficult, dangerous or simply inefficient for people to inspect regularly.

And some of those inspection robots have four legs.

Next in the series: Robot Dogs, Drones and Smart Sensors: Meet the New Maintenance Team.

Missed the other parts of the series?

• Part 1 – From Maintenance Schedules to AI: We looked at how predictive maintenance evolved from historical reliability methods and condition monitoring to AI-supported prediction.
• Part 2 – Predictive Maintenance Gives You 30 Days. Now What?: We explored what needs to happen after an early warning, from identifying the correct spare part to checking availability and current supplier lead times.
• Part 4 – Robot Dogs, Drones and Smart Sensors: We looked at how mobile inspection technologies are expanding the ways manufacturers can collect condition data across their assets. (to be published soon)
• Part 5 – From Predictive Maintenance to Predictive Spare Parts Planning: We explore how condition data and expected maintenance demand could help make spare parts planning more forward-looking. (to be published soon)
• Part 6 – The Autonomous Factory: Can a Machine Order Its Own Spare Part? We bring the series together and ask how far the process from detecting deterioration to identifying, locating and procuring the right spare part could eventually be automated. (to be published soon).