Predictive Maintenance
Also known as: condition-based maintenance, CBM, PdM, AI maintenance prediction
Predictive maintenance is the practice of using data from sensors, operating logs, and machine learning models to forecast when a piece of equipment is likely to fail or require service: allowing maintenance to be performed at the optimal time before failure, rather than after it (reactive) or on a fixed calendar schedule (preventive).
It is widely used in manufacturing, aviation, energy, and facilities management to reduce unplanned downtime, extend equipment life, and optimize maintenance labour costs.
Predictive Maintenance, explained at three levels
At home
Predictive maintenance is like having a doctor who can tell you when something is about to go wrong with a machine before it actually breaks. Instead of waiting for a factory machine to break down (which stops production) or replacing parts on a fixed schedule (which might replace them too early), the machine's sensors measure things like vibration, temperature, and pressure constantly. An AI watches those measurements and learns to recognize the subtle patterns that appear before a failure happens: like a doctor noticing early warning signs of illness.
At work
Predictive maintenance helps operations teams avoid two expensive failure modes: unexpected breakdowns (which cause unplanned downtime and emergency repair costs) and over-maintenance (replacing parts that are still serviceable, wasting money). By predicting when a component will actually need attention, it enables the maintenance team to schedule work at a convenient time: during planned downtime, when the right parts are in stock and the right technician is available.
At home
Predictive maintenance systems typically combine sensor data streams (vibration, temperature, pressure, acoustic emissions, oil analysis) with operational data (hours run, load levels, environmental conditions) and historical maintenance records (when failures occurred, what parts were replaced). Machine learning models are trained on this combined dataset to identify the multivariate patterns that precede specific failure modes. The model outputs a risk score or remaining useful life (RUL) estimate for each monitored asset.
At work
The economic case for predictive maintenance depends on the cost profile of the equipment being monitored. For assets where unplanned failure is extremely expensive (aircraft engines, offshore oil equipment, hospital critical equipment), even imperfect prediction that reduces failure rates by 20-30% is highly valuable. For low-cost, easily replaceable equipment, the cost of instrumentation and monitoring may exceed the savings from avoided failures. Asset criticality analysis should precede any predictive maintenance investment.
At home
Predictive maintenance data architectures typically use edge processing to handle the volume of raw sensor data at the source, transmitting only aggregated features or anomaly alerts to a central platform rather than raw sensor streams. This reduces bandwidth requirements and latency. Cloud-based machine learning platforms then analyse the aggregated data and push updated risk scores or alerts back to maintenance teams through CMMS (Computerized Maintenance Management Systems) integration.
At work
From a practitioner standpoint, the most common predictive maintenance project failure is insufficient or poor-quality historical data. A model trained on only three months of sensor data from a machine that fails once every two years will have seen very few examples of the pre-failure patterns it needs to learn. Successful implementations typically require at least 12-24 months of historical data including multiple failure events for each failure mode, or use transfer learning from similar machines. Starting with the most common and costly failure mode, rather than trying to predict all failure modes simultaneously, reduces scope and improves model quality.
Real-world examples
At home
Nadia got a notification from her electric car's app saying the brake pads were estimated to need replacement within the next 3,000 miles based on braking pattern analysis from the car's sensors. She was able to book a service at her convenience rather than being stranded when the brakes wore out completely.
At work
A paper mill's maintenance team deployed predictive maintenance on their 40 largest paper machines using vibration and temperature sensors on critical bearings and motors. The AI model generated a daily risk score for each machine. In the first six months, the team avoided four unplanned production stoppages that historical data suggested would have lasted an average of 18 hours each: a saving of roughly 72 hours of lost production.
How it compares
| Feature | Predictive Maintenance | Preventive Maintenance | Reactive Maintenance |
|---|---|---|---|
| When maintenance occurs | When sensor data / AI predicts a failure is imminent | On a fixed calendar or usage-hour schedule | After a failure has occurred |
| Data required | Real-time sensor data + AI/ML model | None: based on manufacturer's recommended intervals | None: triggered by failure event |
| Downtime risk | Low: maintenance is planned before failure | Medium: schedule may replace parts too early or too late | High: unplanned downtime after failure |
| Cost profile | Higher initial investment; lower long-term cost for complex assets | Predictable but may waste money on early replacements | Low initial cost; highest long-term cost for critical assets |
| Best for | High-value complex equipment where failure is expensive | Low-cost equipment or when sensor data is unavailable | Non-critical equipment where failure is acceptable |
Predictive maintenance reduces both unnecessary maintenance and unplanned failures but requires upfront investment in sensors and AI. Preventive maintenance is simpler but may over- or under-maintain. Reactive maintenance is cheapest upfront but most expensive when applied to critical assets.
For analysts
In BA artefacts, predictive maintenance requirements surface in manufacturing, facilities, transportation, and energy sector projects. Key requirements to specify: which assets will be monitored, what sensor data is available, what failure modes are being predicted, what the alert threshold is (at what risk score does maintenance get scheduled?), and how alerts integrate with the existing CMMS or work order system. The most important acceptance criterion is often the lead time: 'The system must provide at least 72 hours of advance warning for Failure Mode X with a false positive rate below 15%, verified against 12 months of holdout historical data.'
History & origin
The concept of condition-based maintenance (performing maintenance when equipment condition indicates it is needed, rather than on a fixed schedule) has existed in industrial engineering since the 1960s. The availability of low-cost sensors, industrial IoT connectivity, and machine learning tools in the 2010s made predictive maintenance at scale practical for the first time. Technology providers including GE (Predix platform, 2015), IBM, SAP, and Microsoft Azure IoT built commercial platforms specifically for predictive maintenance use cases. Aviation has been a leading adopter, with jet engine manufacturers monitoring engine sensor data in real time to predict when components need replacement.