What predictive maintenance actually predicts

The moment equipment stops, losses start accumulating. So plants work hard to prevent failures in advance, but setting the standard for “in advance” is not easy. Predictive maintenance is an attempt to set that standard from data rather than instinct.

The limits of preventive maintenance

Many sites still run maintenance on a calendar: replace and inspect parts every few months. It is safe, but wasteful. Parts with life left get thrown away, and equipment still fails suddenly between service intervals. Time, as a criterion, has nothing to do with the actual condition of the machine.

Predictive maintenance reads the early signs

Predictive maintenance is not magic that guesses failures in advance. It is closer to reading, in data, the signs a machine leaves on its way toward failure. As a bearing wears, the vibration pattern shifts gradually; as a motor is strained, current and temperature trace a different course than usual. These changes show up in the trend, not in any single reading.

What has to be in place for prediction to work

The quality of a prediction depends more on the data beneath it than on the model. First you need a history of how that machine normally behaves. Without a baseline to compare against, you cannot even tell whether the current value is unusual. And you need context for which machine a signal belongs to and what it connects to — an ontology — before a cause can be narrowed down. If data is scattered and its provenance unclear, even a good model will chase the wrong thing.

Measuring vibration with wireless sensors

If vibration is the signal that reveals early signs first, the next question is how to measure it consistently. Sending someone around with a handheld meter is hard to do often, and the records vary from person to person. Wiring a sensor to every machine, meanwhile, is no small construction job.

So we use wireless vibration sensors. A sensor like the Advantech WISE-2410 simply mounts on the motor, and with no wiring work it can go onto equipment that is already running. A built-in 3-axis accelerometer measures vibration along with temperature, and the data goes out over LoRaWAN. Because LoRaWAN is low-power, the battery lasts and the signal carries far, so data from sensors spread across a large plant can be collected by one or two gateways such as the WISE-6610.

The measured 3-axis values are assessed against an international standard such as ISO 20816. That lets you answer “is this much vibration normal?” against a standard rather than a hunch. The sensor itself is built to an IP66 rating across a wide temperature range, so it holds up in humid, hot and dusty conditions.

Wireless sensors lower the threshold for predictive maintenance considerably. With no wiring and no walking rounds, vibration data starts accumulating simply by attaching a sensor to the machine you want to watch.

In a form a person can act on

A risk score alone does not move the shop floor. More than “this pump’s risk is 0.87,” it is a sentence like “vibration is rising in a pattern similar to the run-up to the last failure; bearing wear is suspected; inspection recommended” that produces action. An AI agent turns the analysis back into human language like this, and presents the data and history the judgment rests on alongside it.

Where to start

Do not try to cover every machine at once. It is more realistic to pick a single machine with a clear failure history and a high cost of stopping. Put that machine’s history in order, connect the related signals through an ontology, confirm that early signs are actually being caught, and widen the scope from there.

In closing

For the groundwork → Five principles for turning industrial data into an asset
For how AI decides with evidence → AI agent whitepaper