“Predictive maintenance” takes two words to say and two years to build. Most projects that start with a sensor catalogue end in month three with “too many alarms, so we switched them off”. We start from the other end: which machine stops the line when it stops?
Critical assets first, sensors second
If a plant has a hundred motors, perhaps five are critical: production stops when they stop, there is no spare, and repairs take weeks. Predictive maintenance starts with those five. Periodic maintenance is enough for the other ninety-five; fitting sensors to them is waste. Projects that set off without making this distinction lose both the budget and the operators’ interest.
The decision before the sensor
01 / 04
100 motors
A plant may have a hundred motors; they are not all equal.
02 / 04
5 critical assets
The five that stop production when they stop, have no spare and take weeks to repair. Predictive maintenance starts with them.
03 / 04
Three signals
Vibration tells you about bearings and imbalance, current about load and friction; temperature warns late, but reliably.
04 / 04
Normal and deviation
For the first weeks you only measure, and work out each machine’s own normal. A deviation becomes a work order in the maintenance team’s hands.
- critical machines to start with
- 0critical machines to start with
- signals: vibration, current, temperature
- 0signals: vibration, current, temperature
- weeks of measuring only, first
- 0-6weeks of measuring only, first
Which signal?
On rotating equipment, vibration and motor current are the two signals that tell you the most. Vibration speaks about bearings and imbalance; current about load and mechanical friction. Temperature warns late, but it is cheap and reliable. We usually take all three together; deciding on a single signal is like judging distance with one eye.
Threshold or model?
Everyone wants machine learning; for most plants a well-set threshold is enough. But the threshold cannot be fixed: when the machine’s load changes, so does its vibration. Our approach: for the first four to six weeks, only measure; work out each machine’s own normal for each of its operating modes; then alarm on deviation. Some cases do call for a model, such as slowly developing bearing damage, but you get there only after setting the threshold up properly.
Who will read the alarm?
The question most often skipped. If an alarm is flashing on a SCADA screen and the maintenance team is not in front of that screen, there is no alarm. An alarm that doesn’t reach the maintenance work order system (or at least the maintenance team’s phones) becomes a record, not an action. The software side of predictive maintenance is less about the algorithm and more about building this flow.
- Start with five critical machines; widen the scope over six months.
- Raise no alarms in the first weeks; just measure and learn what normal looks like.
- Give every alarm an owner and a matching work order.
- Keep sensor data in the same place as production data (shift, product, speed); that is usually where the cause of a deviation lies.