The difference between predictive and preventive maintenance is the trigger. Preventive maintenance services a machine on a fixed schedule of days or running hours; predictive maintenance services it when readings such as vibration, temperature or pressure show that a failure is becoming likely. Most plants need both, chosen asset by asset.
This guide explains reactive, preventive, condition-based and predictive maintenance in plain terms, compares them side by side, gives predictive maintenance examples for common plant machines, and sets out a practical path from a calendar-driven schedule to one driven by data.
The four maintenance strategies explained
Reactive maintenance (run to failure)
You fix the machine after it breaks. There is no planning and no monitoring cost, but the failure arrives at a time of its choosing: in the middle of a shift, a peak order or a night with no fitter on site. Secondary damage is common, because a failed bearing can take a shaft, seal or impeller with it.
Reactive maintenance is still a valid choice for some assets. A cheap exhaust fan in a store room or a spare pump on standby does not justify a schedule or a sensor.
Preventive maintenance (time or usage based)
You service the machine at fixed intervals: every three months, every 2,000 running hours, every set number of cycles. Typical tasks are greasing, filter changes, belt checks, oil changes and planned overhauls. The intervals usually come from the manufacturer's manual, adjusted by experience.
Preventive maintenance is simple to plan and audit, and it cuts many failures. Its weakness is that the schedule does not know the machine's actual condition. Parts are often replaced with useful life left, and a failure that develops between two services is still missed.
Condition-based maintenance
You act when a measured condition crosses a set limit. Examples: change a filter when differential pressure exceeds a value, or stop a motor when winding temperature passes an alarm point. This is a step towards predictive maintenance, but it reacts to a threshold rather than a trend, so warning time can be short.
Predictive maintenance
You watch how a machine's readings behave over time and act when they start drifting from normal, well before a hard limit is crossed. The aim is to estimate that a failure is likely within a window, for example the next one or two weeks, so the repair can be booked into a planned stop with parts and people ready.
Predictive maintenance needs continuous or frequent data, a baseline of normal behaviour for each machine, and a record of what was found when someone went to check.
Reactive vs preventive vs predictive maintenance: comparison table
| Aspect | Reactive | Preventive | Condition-based | Predictive |
|---|---|---|---|---|
| Trigger | Breakdown | Calendar date or running hours | A reading crosses a fixed limit | A reading trend drifts from normal |
| Data needed | None | Run hours, maintenance calendar | Live readings for key parameters | Live readings, history, failure and repair records |
| Cost profile | Low planned cost, high and unpredictable failure cost | Steady planned cost; some parts replaced too early | Moderate; sensors and alarm setup | Higher setup effort; lower cost per avoided failure on critical assets |
| Pros | No effort until failure | Easy to plan, staff and audit | Acts on real condition | Longest warning time; repairs fit planned stops |
| Cons | Unplanned stops, secondary damage, safety risk | Misses failures between services; over-maintenance | Short warning; limits need tuning | Needs good data, history and a feedback loop |
| Best for | Cheap, non-critical, redundant items | Wear items with predictable life; statutory checks | Filters, temperatures, levels with clear limits | Critical rotating and thermal equipment with costly failures |
Note that these are not rival philosophies. A well-run plant uses all four, and the question is which one fits each asset.
Predictive maintenance examples by machine type
The examples below are typical early signs that maintenance teams watch. The exact signals depend on the machine, its duty and the instruments already fitted.
Pumps
- Bearing wear: vibration at the bearing housing rising slowly over weeks, often with a small rise in bearing temperature.
- Impeller wear or cavitation: discharge pressure or flow falling at the same motor current, or noisy, unstable suction pressure.
- Seal problems: seal chamber temperature creeping up, or a flush flow changing.
Electric motors
- Current drawn at the same load rising over time, pointing to mechanical drag or a driven-equipment problem.
- Current imbalance between phases growing.
- Winding or bearing temperature trending up relative to ambient temperature and load.
- Vibration changes linked to misalignment, looseness or a bearing defect.
Air compressors
- Discharge temperature drifting upward, a sign of cooler fouling or oil problems.
- Longer loaded time to hold the same header pressure, suggesting leaks or internal wear.
- Oil separator pressure drop rising towards its change point.
- Specific power (kW per unit of air) worsening month on month.
Boilers
- Flue gas temperature rising at the same firing rate, which often points to fouling on heat transfer surfaces.
- Oxygen in flue gas drifting, suggesting burner or air damper issues.
- Feed pump pressure or feedwater flow behaving unusually for the steam load.
- Steam output per unit of fuel slowly falling.
Turbines
- Bearing vibration and bearing metal temperature trending up.
- Lube oil pressure or temperature drifting from its usual band.
- Exhaust temperature spread widening, or output falling at the same inlet conditions.
Cooling towers
- Approach temperature (cold water temperature minus wet bulb) widening, a sign of fill fouling or poor water distribution.
- Fan motor current or gearbox vibration rising.
- Make-up water use changing without an obvious reason.
In every case the value lies in the trend, not a single reading. A bearing at 68 °C may be fine; the same bearing moving from 55 °C to 68 °C over three weeks at the same load deserves a look.
What data and history predictive maintenance needs
- The right signals. Readings that actually change as the failure develops: vibration, temperature, pressure, current, flow, speed. Start with what your PLC, SCADA or historian already records.
- Context signals. Load, speed, ambient temperature and running state, so a change caused by higher production is not mistaken for a fault.
- Enough frequency. A reading every few seconds or minutes suits most temperature and pressure trends. Detailed vibration analysis needs much faster sampling, usually from dedicated sensors.
- A baseline. Several weeks of readings in normal operation, so the system knows what normal looks like for that machine.
- Failure and repair history. Work orders with dates, what failed and what was done. Even a spreadsheet or register helps.
- Outcomes of each warning. When a warning is raised, record what the fitter found: fixed, scheduled or nothing found. This is how warnings become more accurate.
- Clean asset identity. Every reading tied to the correct machine and tag. Data from a mislabelled or unknown source should be kept out of the trends.
Getting data off the machines is its own topic. If you are deciding how, see our comparison of OPC UA vs MQTT vs Modbus for plant data.
How to choose a maintenance strategy for each asset
Go through your asset list and score each machine on three questions.
- Criticality. If this machine stops, what stops with it? A single boiler feeding the whole plant is critical. One of three parallel pumps with a standby is not.
- Failure cost. Add up lost production, repair cost, secondary damage, safety and environmental risk, and customer penalties. A failure that costs an afternoon is different from one that costs a week and a rebuilt gearbox.
- Detectability. Does the failure give an early sign in data you can collect? Bearing wear usually does. Some electronic faults or sudden fractures give little or no warning.
Then apply a simple rule of thumb:
- High criticality, high failure cost, detectable: predictive maintenance, backed by preventive basics such as lubrication.
- High criticality but hard to detect: preventive maintenance, redundancy or spares on the shelf.
- Clear limit exists (filters, levels, temperatures): condition-based maintenance.
- Low criticality, cheap, easily replaced: run to failure, with a spare in stock.
A step-by-step path from preventive to predictive maintenance
- Keep the preventive programme running. Do not drop scheduled tasks before the replacement is proven.
- Pick five to ten critical assets. Use the criticality, failure cost and detectability scoring above. Machines that have stopped the plant before are good candidates.
- List the failure modes and signals. For each asset, write down how it has failed and which reading would have shown it coming.
- Connect the data you already have. Bring PLC, SCADA or historian tags into one place with the right machine and tag names. Add sensors only where a key signal is missing.
- Watch and alert first. Set sensible limits and trend views, and make sure alerts reach a named person who must acknowledge them. This is condition-based maintenance and it pays off early.
- Build the baseline and record outcomes. Let several weeks of normal data accumulate, and log what was found after every alert and every breakdown.
- Add trend-based warnings. Look for drift from each machine's normal behaviour, and send a plain-language warning with what to check first.
- Adjust preventive intervals. Where data shows a part is healthy at its service date, stretch the interval with care. Where failures slip through, shorten it or add a signal.
- Review monthly. Check which warnings were useful, which were noise and which failures were missed, and tune accordingly.
For the wider picture of cutting stops, see our guide on how to reduce unplanned downtime.
Common pitfalls when moving to predictive maintenance
- Alert fatigue. Too many alerts with tight limits train people to ignore them. Start with fewer, well-chosen alerts on critical machines, and make each one say what to check.
- No owner for alerts. An alert that goes to a shared screen or a group chat with no one responsible is easy to miss. Name a person, and escalate if they do not respond.
- No feedback loop. If nobody records what was found after a warning, the system cannot learn which warnings matter, and the team cannot prove the programme works.
- Too little history. Judging a machine on a few days of data leads to false warnings. Give each machine time to establish its normal pattern across loads and seasons.
- Ignoring operating context. A temperature rise during a summer afternoon or a heavier production run is not a fault. Compare like with like.
- Dropping preventive tasks too early. Lubrication, cleaning and inspections still matter. Predictive maintenance adds to good basics; it does not replace them.
- Trying to cover every machine at once. Spread thinly, the effort shows little result. Prove it on a handful of critical assets first.
Where MIE fits
MIE, the Manufacturing Intelligence Engine from ThinklytixAI, covers the early steps of this path today. It takes readings your machines already send over HTTPS, MQTT, OPC-UA or Modbus, shows trends and machine pages, and sends alerts on Slack, WhatsApp or webhook that escalate through timed tiers until someone acknowledges them. Teams can also ask plain-English questions about their plant data. You can see how it works on the MIE Manufacturing Intelligence Engine overview.
Predictive maintenance in MIE is coming soon: it is the next major feature on the MIE roadmap. It is designed around the loop described above, with early warnings when a reading drifts, acknowledgement, a recorded outcome (fixed, scheduled or nothing found) and warnings that improve as outcomes build up. Read more about MIE predictive maintenance and production goal tracking.