Every machine eventually fails. The question is whether you find out when it breaks down in the middle of production, or weeks earlier from the data. Predictive maintenance uses sensors and analytics, increasingly machine learning, to predict failures and schedule maintenance at the right time.

Three maintenance strategies

  1. Reactive: fix it when it breaks. Simple, but failures are costly and unpredictable.
  2. Preventive: service on a fixed schedule. Reduces failures but replaces parts that still had useful life.
  3. Predictive: service when the data indicates degradation. Aims for the best of both.

The data

Predictive maintenance relies on condition monitoring — measuring signals that change as equipment degrades:

  • vibration, the classic signal for rotating equipment like motors, pumps, fans and gearboxes;
  • temperature of bearings, windings and fluids;
  • acoustic signals, including ultrasound;
  • electrical current and power drawn by motors;
  • oil analysis for wear particles;
  • operational data such as load, speed, pressure and cycle counts;
  • maintenance records of past failures and repairs.

The models

Anomaly detection. Learn what normal behavior looks like and flag deviations. Useful when few failures have been recorded — often the case.

Failure classification. Identify the type of fault, such as bearing wear or imbalance, from signal patterns.

Remaining useful life estimation. Predict how long until a component reaches failure, enabling maintenance planning.

Physics plus data. Combining engineering knowledge of how equipment degrades with machine learning often works better than either alone.

Typical benefits

Organizations adopt predictive maintenance to reduce unplanned downtime, cut maintenance costs, extend equipment life, improve safety and manage spare parts better. Results vary widely with equipment criticality and data quality, so business cases should be built on the cost of downtime for specific machines.

How to start a project

  1. Choose critical assets where failures are expensive or dangerous.
  2. Understand failure modes with maintenance engineers: how does this machine fail, and what signals precede it?
  3. Instrument and collect data — retrofitting wireless sensors is often affordable.
  4. Start with anomaly detection and alerts reviewed by experts.
  5. Close the loop: record what technicians find when they inspect an alert, so models improve.
  6. Integrate with maintenance systems so predictions become work orders.
  7. Measure downtime, maintenance costs and false alarm rates.

Pitfalls

  • too many false alarms, causing staff to ignore alerts;
  • lack of failure data to learn from;
  • poor sensor placement or data quality;
  • predictions that don’t fit maintenance planning cycles;
  • ignoring technicians’ knowledge.

Beyond factories

The same approach is used in wind turbines, aircraft engines, trains, elevators, data centers and power grids.

A worked example: a pump in a water treatment plant

A water utility installs vibration and temperature sensors on a critical pump. For the first month, the system learns its normal operating patterns at different flow rates. In month three, the model detects a gradual increase in vibration at a specific frequency associated with bearing wear. The alert goes to a maintenance engineer, who confirms the trend with a handheld analyzer and schedules a bearing replacement during a planned low-demand period two weeks later.

Without the alert, the bearing might have failed unexpectedly, possibly damaging the pump shaft and interrupting service. The cost of the sensors and software is compared against avoided downtime, emergency repairs and secondary damage.

Choosing between building and buying

Options range from sensor vendors’ turnkey platforms, to maintenance software with built-in analytics, to in-house data science. Turnkey solutions are faster for common equipment like motors and pumps; custom models make sense for unique, high-value assets where the organization has deep process knowledge.

What to watch next

  • Generative AI assistants that explain alerts and suggest repair steps.
  • Cheaper wireless sensors for retrofitting older machines.
  • Foundation models for industrial time-series data.
  • Combination of digital twins with predictive models.

Key takeaways

  • Predictive maintenance services equipment based on its actual condition.
  • Vibration, temperature, acoustic, electrical and operational data are key inputs.
  • Anomaly detection, fault classification and remaining useful life models are common methods.
  • Start with critical assets and close the loop with technicians’ findings.
  • False alarms and scarce failure data are common pitfalls.

Frequently asked questions

What is the difference between preventive and predictive maintenance?

Preventive maintenance follows a fixed schedule; predictive maintenance acts on data showing actual equipment condition.

Do I need years of data to start?

No. Anomaly detection can start with a few weeks of normal operating data, improving as failures and inspections are recorded.

Is predictive maintenance only for factories?

No. It is widely used in energy, transport, buildings and infrastructure.