Predictive maintenance: a practical guide

What predictive maintenance is, how it differs from preventive maintenance, which techniques fit which assets, and how to start without boiling the ocean.

Maintenance & reliability · 10 min read · Updated 2026-05-29

Reactive, preventive and predictive

There are three broad maintenance strategies. Reactive (run-to-failure) fixes things after they break — cheap until it isn't. Preventive maintenance services assets on a fixed schedule regardless of condition, which avoids some failures but also replaces healthy parts and still misses random faults. Predictive maintenance uses the actual condition of the asset to act just before failure — getting most of the failure avoidance of preventive maintenance with far less unnecessary work.

No plant should be purely one or the other. The art is matching the strategy to the asset: run-to-failure for cheap, non-critical items; predictive for critical or expensive ones.

The main predictive techniques

• Vibration analysis — the workhorse for rotating equipment (pumps, motors, fans, compressors); detects imbalance, misalignment, bearing wear and looseness.
• Thermography — infrared imaging finds hot connections, overloaded equipment and insulation or refractory problems.
• Oil analysis — wear particles and contamination reveal gearbox and bearing condition.
• Ultrasound — detects early bearing faults, leaks and electrical discharge.
• Motor current / electrical signature analysis — finds rotor, winding and load faults.
• Process-data analytics — modelling expected behaviour from existing sensors to catch drift on assets without dedicated condition sensors.

The data and the workflow

Predictive maintenance only works if a detection turns into an action. That means three things have to connect: the condition data, an analysis that produces a clear diagnosis and severity, and a work-management process that schedules the fix. A CMMS or asset-management system is the backbone — it holds the asset register, turns alerts into work orders, and records what was found so the models improve. Detections that never become work orders deliver nothing.

Sensor-based vs analytics-based approaches

Two broad product approaches exist. Sensor-based platforms add condition sensors (typically vibration and temperature) to specific machines and diagnose from that data — fast to deploy, excellent on rotating equipment, but per-machine cost grows with scale. Analytics-based platforms model existing historian, SCADA and maintenance data to cover many assets without new sensors — better for scaling across large fleets, but dependent on the quality of existing data. Many plants use both: sensors on the critical rotating assets, analytics across the wider estate.

How to start

The common failure mode is trying to instrument everything at once. A better path:

• Rank assets by criticality and failure cost; start with the worst handful.
• Pick the technique that fits those assets (usually vibration for rotating equipment).
• Wire detections into your CMMS so alerts become scheduled work.
• Measure avoided downtime and prove the value before scaling.
• Expand to analytics-based coverage once the workflow is working.

Frequently asked questions

Related guides

Software that helps