Three ways to look after a machine
Machines can be maintained in roughly three ways.
Reactive maintenance means fixing a machine after it breaks. It costs nothing until a failure, but failures tend to come at bad moments, stop production and sometimes damage other parts.
Scheduled maintenance means servicing on a fixed calendar or usage interval, such as every thousand hours. It prevents many failures, but some work is done too early on parts that were fine, while other parts still fail between visits.
Predictive maintenance means servicing when the evidence suggests a problem is developing. Sensors watch the machine, and software looks for signs that a part is heading toward failure. The aim is to do the right work shortly before it is needed, at a time that suits production.
Predictive approaches are often presented as the main benefit of connecting machines, because unplanned downtime can be expensive. But the approach has real requirements, and it does not suit every machine.
What is measured
Failures rarely happen without warning signs, though the signs may be subtle. Common measurements include:
- Vibration, which changes when bearings wear, shafts become misaligned or parts loosen.
- Temperature, which may creep upward as friction or electrical resistance increases.
- Motor current or power draw, which can shift when a machine works harder to do the same job.
- Sound or ultrasonic emissions, which can reveal leaks or bearing problems.
- Oil condition and pressure, in systems where fluids matter.
- Operating context, such as speed, load and product type, which helps explain why a reading changed.
The wider topic of condition monitoring covers these measurements in more depth. Predictive maintenance is what you do with them: turning readings into a decision.
From readings to a prediction
The simplest version uses thresholds. If vibration passes a set level, raise a flag. This is easy to understand and often useful, but a fixed limit can miss gradual change and may fire when conditions are merely unusual.
More advanced versions compare current behavior against the machine’s own history, using trend analysis or anomaly detection to notice when readings drift from normal. Some systems use machine learning to estimate how much useful life remains in a part. These methods need enough past data, including examples of actual failures, to learn from. Many machines fail rarely, which means there may be few examples, and the resulting predictions carry a lot of uncertainty.
Readings are typically stored as time series data so trends can be examined over weeks or months. A digital twin, a software model of the machine, is sometimes used to compare expected and actual behavior, though a basic setup works without one.
Where the savings come from
When prediction works, the benefits can include:
- Less unplanned downtime, since repairs are scheduled rather than forced.
- Better use of parts, because components are replaced closer to when they need it rather than on a rigid calendar.
- Smaller secondary damage, since a failing bearing caught early may not wreck the shaft.
- Safer work, since fewer emergency repairs reduces rushed, risky interventions.
- Better planning, since spare parts and technician time can be arranged in advance.
How large these gains are depends heavily on the plant, the machines and how costly downtime is. Figures quoted in marketing material should not be assumed to apply elsewhere. A machine that sits idle half the time, or one that is cheap and easy to swap, may not justify the effort. A bottleneck machine whose stoppage halts the whole line is a much better candidate. Improved availability also shows up in measures such as equipment effectiveness.
The caveat: data quality
Predictions are only as good as the data behind them. Typical problems include:
- Sensor drift or poor mounting. A vibration sensor that is loosely attached reports noise, not machine behavior. Regular sensor calibration and secure installation matter.
- Missing or gappy data. Network drops and sensor failures leave holes that can confuse analysis. Broader issues are covered under data quality.
- No failure labels. If maintenance records say only “repaired motor” without date, cause or symptoms, there is nothing to teach a model what a failure looked like.
- Changing conditions. A new product, a different supplier’s material or a seasonal temperature swing can shift readings without anything being wrong.
- False alarms. Too many unnecessary warnings make technicians stop trusting the system. Missed failures are the opposite problem. Tuning the balance takes time.
Predictions should be treated as guidance for human judgment, not as certainty. An experienced technician’s observation that a machine “sounds different” is still valuable and can complement the sensors.
Starting practically
A reasonable path looks like this.
- Pick one critical machine with a known, costly failure mode.
- Review maintenance history to see what actually goes wrong and how often.
- If the machine is old, machine retrofitting with external vibration and temperature sensors can add data without changing the machine.
- Collect data for a while, and record every intervention accurately.
- Start with simple thresholds and trends before attempting complex models.
- Run it as one of several pilot projects, and compare results with a baseline: number of unplanned stops, repair hours and parts cost.
If the results are encouraging, expand slowly. If not, the data collected still gives a better picture of how the machine behaves, which has value of its own.