Predictive maintenance

Overview

Predictive maintenance uses data from machines, such as vibration, temperature, pressure, motor current and operating hours, to estimate when equipment is likely to fail, so that it can be serviced shortly before the problem occurs. It sits between two older strategies. Run-to-failure waits until something breaks, which is cheap until a breakdown stops a production line. Preventive maintenance services equipment on a fixed schedule, which is safer but replaces many parts that still had plenty of life.

The aim is simple: fix the right machine at the right time. Done well, it reduces unplanned downtime, avoids secondary damage, cuts spare-part stock and lets maintenance teams plan their work instead of firefighting. It is one of the most common industrial uses of machine learning and a core part of Industry 4.0.

An everyday picture is a good family doctor. Instead of waiting for a heart attack, or prescribing the same treatment to everyone each year, the doctor watches trends in blood pressure and test results and acts when the numbers start to drift. Predictive maintenance applies the same idea to pumps, motors, turbines and conveyor belts.

What it is

These approaches build on each other. Many programmes start with condition monitoring and anomaly detection, because failure labels are scarce, and move to failure prediction and remaining-life estimates as maintenance records improve.

A key idea is the gap between the first detectable sign of a problem and actual failure, often drawn as a curve from "potential failure" to "functional failure". The longer that warning window, and the earlier a model can see into it, the more value predictive maintenance can deliver.

Where it came from

Maintenance practice has moved through stages. Early industrial plants largely ran equipment until it broke. Through the twentieth century, scheduled preventive maintenance became standard, especially where failure was dangerous. Aviation, in the 1960s and 1970s, helped develop reliability-centred maintenance, which matched the strategy to how each component actually fails rather than assuming everything wears out on a schedule.

Condition monitoring grew alongside: vibration analysis of rotating machines, oil analysis, thermography and ultrasound became established engineering disciplines. What changed in the 2010s was scale. Cheaper sensors, industrial networks, cloud storage and machine learning made it practical to monitor thousands of assets continuously and learn failure patterns from data instead of only from fixed rules.

Public research datasets helped the field. NASA's prognostics work produced simulated turbofan engine degradation data generated with a model called C-MAPSS, released for a data challenge at the first International Conference on Prognostics and Health Management in 2008. Datasets like this remain common benchmarks for remaining-useful-life methods.

How it works

Step 1: choose the assets and failure modes

Start with equipment where failure is costly and where a measurable signal precedes failure. A failure mode and effects analysis, done with maintenance engineers, identifies which faults matter and which sensors could reveal them.

Step 2: collect the data

Typical sources are:

• Sensor time series: vibration, temperature, pressure, flow, current, acoustic signals.
• Control-system data: set points, speeds, loads and states from PLCs and SCADA.
• Maintenance records: work orders, failure codes, parts replaced, from a maintenance management system.
• Context: asset age, model, operating environment, production schedule.

High-frequency signals such as vibration are often processed at the edge, near the machine, and only summary features are sent on.

Step 3: build features

Raw signals are turned into features over time windows: averages, trends, peaks, root mean square, kurtosis, frequency-band energy from a Fourier transform, and differences from the asset's own normal baseline. Features must be computed only from data available at prediction time.

Step 4: label and model

With few failures, use unsupervised methods such as isolation forest or autoencoders trained on healthy operation. With enough labelled failures, train classifiers such as gradient boosting for "fails within 14 days", or sequence models such as LSTMs for remaining useful life. Physics-based models and digital twins can add knowledge where data are thin.

Step 5: deploy and act

Models score streaming data, often through a platform built on Apache Kafka or an industrial historian, and push alerts into the maintenance system as work orders. The output must reach the people who plan maintenance, with enough explanation for them to trust it.

Step 6: monitor and improve

Track alert precision, missed failures, lead time and maintenance outcomes. Feed confirmed failures and false alarms back as labels, and watch for drift when machines are repaired, reconfigured or run differently; see model monitoring.

Key concepts and terms

A worked example

This scenario is illustrative. A water utility operates 120 large pumps. Each unplanned pump failure costs roughly two days of emergency work and parts; planned repairs cost about a third as much. Over three years the utility recorded 41 bearing failures.

Each pump has a vibration sensor sampled at high frequency and a temperature sensor. An edge device computes hourly features: overall vibration level, energy in frequency bands associated with bearing defects, temperature above ambient, and their seven-day trends. The team trains a gradient boosting classifier to predict "bearing failure within 21 days", using the first two years for training and the third for testing, split by time.

On the test year the model flags 11 of the 13 bearing failures with a median lead time of 16 days, and raises 9 false alarms. Each alert opens an inspection work order with the features that drove it. Engineers confirm that most false alarms show real early wear, so they adjust the threshold rather than discard the model. The utility then plans bearing replacements into scheduled maintenance windows instead of responding to emergencies.

Strengths and limitations

Strengths:

• Reduces unplanned downtime and emergency repair costs.
• Extends component life compared with fixed replacement schedules.
• Lets maintenance teams plan work, parts and staff.
• Builds a data record that improves asset decisions over time.
• Can improve safety by catching dangerous failures early.

Limitations:

• Failures are rare, so labelled data are scarce and imbalanced.
• Sensor installation and data integration can cost more than modelling.
• Machines differ, so models may not transfer between assets or sites.
• Maintenance records are often incomplete or inconsistently coded.
• Value depends on the organisation acting on alerts.

How it compares with alternatives

Practical guidance and best practices

1. Start with a few critical asset types where failures are costly and signals exist.
2. Involve maintenance engineers from the start to define failure modes and labels.
3. Clean up maintenance records and failure codes before relying on them as labels.
4. Begin with anomaly detection if failures are rare, and add supervised models as labels accumulate.
5. Validate with time-based splits, never random splits of sensor windows.
6. Measure success in business terms: avoided downtime, lead time and false-alarm rate.
7. Integrate alerts into existing work-order systems rather than a separate dashboard.
8. Retrain after repairs, upgrades or changes in operating conditions.

Common mistakes

• Randomly splitting overlapping time windows, which leaks the future into training.
• Building models before checking that sensor data and failure records line up in time.
• Treating every alarm as a success without tracking whether action followed.
• Ignoring operating context, so load changes look like faults.
• Training one model for all assets when they behave very differently.
• Launching without a clear process for who acts on each alert.

Frequently asked questions

Do I need failure data to start?

No. Anomaly detection trained on normal behaviour can flag unusual patterns without labelled failures. Over time, inspections and confirmed failures provide labels for supervised models that predict specific faults.

What sensors matter most?

It depends on the failure mode. Vibration is central for rotating equipment such as motors, pumps and fans. Temperature, current, pressure and oil quality matter for others. Start from the failure modes and choose sensors that reveal them early.

How is remaining useful life estimated?

Common approaches include regression on degradation features, survival analysis that handles assets which have not yet failed, sequence models on sensor histories, and physics-based degradation models. Each gives an estimate with uncertainty, which should be shown to planners.

Should models run at the edge or in the cloud?

Often both. High-frequency signal processing and simple alarms run at the edge to cut bandwidth and latency. Model training, fleet-wide comparison and heavier scoring usually run centrally. See real-time systems.

How do I measure success?

Track missed failures, false alarms, average lead time, and changes in unplanned downtime and maintenance cost compared with a baseline period or a control group of assets.

Glossary

• Asset: a piece of equipment being maintained.
• Run-to-failure: operating equipment until it breaks.
• Preventive maintenance: servicing on a fixed schedule.
• Predictive maintenance: servicing based on forecast failure.
• RUL: remaining useful life.
• SCADA: supervisory control and data acquisition systems that monitor industrial processes.
• PLC: programmable logic controller, an industrial computer that controls machines.
• Work order: a maintenance task record.
• Drift: change in data or behaviour that degrades a model.

Summary

Predictive maintenance uses sensor data, maintenance history and machine learning to service equipment shortly before it fails, reducing unplanned downtime and wasted preventive work. It grew from condition monitoring and reliability-centred maintenance and became practical at scale with industrial connectivity and modern analytics. Success depends less on clever models than on the right assets, aligned data, honest time-based validation and a maintenance process that acts on alerts. Start small, measure business outcomes and keep models updated as machines change.

Related

It belongs to Industries and applications in my stack, under Machine learning. Others in the same group: