AI predictive maintenance uses sensor data, typically vibration, motor current, temperature and pressure captured by a data acquisition (DAQ) system, and machine learning models to detect early signs of wear and estimate when a machine will need service, so maintenance happens before a failure instead of after it or on a fixed calendar. The AI is only as good as the data behind it: choosing the right sensors, sampling them correctly and turning raw signals into features is most of the work.
This article is part of our guide to industrial AI, and it leans on our data acquisition guide for the measurement side. Here we cover sensors, sample rates, features, model types and how to run a pilot.
Four maintenance strategies
| Strategy | When service happens | Trade-off |
|---|---|---|
| Reactive | After a failure | No monitoring cost, but unplanned downtime and secondary damage |
| Preventive | On a fixed schedule (hours, cycles, calendar) | Predictable, but replaces healthy parts and still misses early failures |
| Condition-based | When a measured value crosses a limit | Needs sensors; simple thresholds give late warnings |
| Predictive (AI) | When a model forecasts degradation | Earliest warning and best scheduling, but needs data, models and upkeep |
Predictive maintenance pays where downtime is expensive, failures develop gradually and can be measured, and the machines are critical enough to justify the sensors: spindles, pumps, compressors, gearboxes, fans, presses and conveyor drives.
Which sensors to use
| Sensor | What it detects | Typical sampling |
|---|---|---|
| Accelerometer (vibration) | Imbalance, misalignment, looseness, bearing and gear faults | Kilohertz to tens of kilohertz |
| Motor current (CT or shunt) | Rotor bar defects, load changes, mechanical binding, pump cavitation | Several kilohertz for current signature analysis |
| Temperature (thermocouple, RTD) | Friction, lubrication loss, electrical overload | Once per second or slower |
| Acoustic emission / ultrasonic | Very early bearing damage, leaks, arcing | Hundreds of kilohertz, usually processed to an envelope |
| Pressure and flow | Pump wear, leaks, valve and filter problems | Tens to thousands of samples per second |
| Process data from the PLC | Speed, load, cycle counts that explain normal variation | Per cycle or per second |
Context matters as much as the fault signal. A vibration level that is normal at full load can be a warning at idle, so log speed and load alongside the sensors. For sensor interfacing, see signal conditioning and, for temperature, thermocouple data acquisition.
Sampling rate: a worked example
Bearing faults produce impacts that show up as high-frequency vibration long before they are audible or hot. Suppose you want to analyze vibration up to 10 kHz on a motor:
- Analysis bandwidth: 10 kHz
- Sample rate: vibration analyzers commonly use 2.56 × the bandwidth to leave room for the anti-aliasing filter, so 2.56 × 10,000 = 25,600 samples per second
- Data per channel: 25,600 samples/s × 4 bytes (24-bit samples stored as 32-bit) = 102,400 bytes/s, about 102 kB/s
- Eight channels: about 819 kB/s, or roughly 2.9 GB per hour
Streaming that to the cloud from every machine is impractical, so most systems compute features at the edge and send only the features, plus raw snapshots when something looks wrong. Sampling too slowly is the opposite mistake: frequencies above half the sample rate alias into the band and corrupt the analysis. Our article on sampling rate, Nyquist and anti-aliasing filters covers the filter side.
From raw signals to features
Models rarely look at raw waveforms directly. Engineers first compute features that track the physics of wear:
| Feature | What it indicates |
|---|---|
| RMS (overall level) | General condition; rises with imbalance, misalignment and looseness |
| Peak and crest factor (peak ÷ RMS) | Impacting; crest factor climbs early in bearing damage |
| Kurtosis | Spikiness of the signal; sensitive to early impacts |
| Spectral band energy (FFT) | Energy at shaft speed, its harmonics, gear-mesh and bearing defect frequencies |
| Envelope spectrum | Repetition rate of bearing impacts hidden under other vibration |
| Current signature sidebands | Rotor and load problems seen in the motor current spectrum |
| Temperature rise over ambient | Friction and lubrication problems |
Model types, from simplest to most advanced
- Statistical baselines and trends. Learn normal ranges per machine and operating state; alarm on sustained deviation. Simple, explainable and often enough.
- Anomaly detection. Models such as isolation forests, one-class models or autoencoders learn healthy feature combinations and flag unusual ones, even for faults never seen before. See AI anomaly detection.
- Fault classification. Supervised models name the fault: bearing, imbalance, misalignment, cavitation. They need labeled examples of each fault, often from teardown records or test rigs.
- Remaining useful life (RUL). Regression or sequence models estimate time to failure from the degradation trend. They need run-to-failure history, which most plants have little of, so RUL usually comes last.
The data problem: failures are rare
A well-run plant has few failures to learn from, which is good for production and hard for supervised AI. Practical ways around it:
- Start with anomaly detection, which needs only healthy data
- Pool data across identical machines so one site’s failure teaches the others
- Log maintenance work orders with timestamps so findings can label the sensor data
- Use test rigs or seeded faults for critical components
- Keep raw snapshots around alarms so engineers can confirm what the model saw
Running a predictive maintenance pilot
- Pick a few critical, failure-prone assets with measurable failure modes.
- Instrument them: sensors, a DAQ system sampled for the fault frequencies, and speed and load from the PLC.
- Collect a baseline across normal operating states, typically weeks.
- Deploy anomaly detection with alarms routed to maintenance, and review every alarm.
- Feed confirmed findings back as labels; add fault classification once you have examples.
- Measure results: avoided downtime, planned vs unplanned work, parts saved.
Want to instrument a machine for predictive maintenance?
Sciotex builds industrial AI systems and custom DAQ systems: sensors, signal conditioning, edge feature extraction and the models on top, designed, programmed, built and supported in the US since 1997.
Frequently asked questions
It is the use of machine learning on sensor data such as vibration, current and temperature to detect early wear and forecast when equipment needs service, so maintenance is planned before a failure.
The most common are accelerometers for vibration, current sensors on motors, temperature sensors, pressure and flow sensors, and sometimes acoustic emission sensors, plus speed and load data from the machine controller.
It depends on the highest frequency you need to analyze. A common rule is 2.56 times the analysis bandwidth, so analyzing up to 10 kHz needs about 25,600 samples per second, with an anti-aliasing filter.
No. Anomaly detection can start from healthy-operation data alone. Labeled failure examples are needed later to classify fault types and estimate remaining useful life.
Raw high-rate vibration data is usually too large to stream continuously, so features are computed at the edge and only features, alarms and short raw snapshots are sent to a server or the cloud.
