Predictive maintenance is the practice of servicing equipment based on its measured condition and a forecast of when it will fail, rather than on a fixed calendar or after it has broken. It is the most mature application of AIoT, because the ingredients line up: rotating machinery produces rich physical signals, cheap sensors capture them, and an unplanned stop costs enough to justify the effort.
The field has a long pre-AI history. Vibration analysts have diagnosed bearings by ear and by spectrum for decades. What IoT adds is continuous measurement on many assets instead of a monthly walk-around with a handheld meter; what AI adds is the ability to learn what normal looks like per machine and to flag the slow changes a threshold would miss.
This guide covers the three maintenance strategies, the signals and the processing behind them, the progression from monitoring to prediction, where the computation should run, and the pitfalls that sink most projects.
Reactive, preventive, predictive
Reactive maintenance runs equipment until it fails: cheap to organize, expensive in downtime and collateral damage.
Preventive maintenance replaces parts on a schedule derived from manufacturer data. It avoids most failures but wastes good components and still misses the early failures that ignore the schedule.
Predictive maintenance measures each asset and acts when the measurement says so. Done well, it extends intervals on healthy machines and catches the outliers. It is not free: it needs sensors, connectivity, models and a maintenance organization that trusts the alerts and can act within the warning window.
The underlying measurement layer is condition monitoring. Many deployments stop there, with trended values and simple thresholds, and that is a legitimate outcome. Prediction is the step beyond.
The signals
Four signals cover most rotating equipment, and the right choice depends on the failure mode you expect.
Vibration is the primary signal for bearings, gears, imbalance, misalignment and looseness, captured by an accelerometer on the bearing housing. MEMS accelerometers are cheap and adequate for many cases; piezoelectric sensors cover higher frequencies and are the traditional choice for critical assets.
Motor current is the signal for electrical faults and for mechanical faults that load the motor. A current clamp on one phase is non-invasive and lives in the switchgear rather than on the machine, which suits hard-to-reach or hazardous equipment.
Temperature tracks lubrication problems, overload and electrical faults. It is a lagging indicator, but cheap and easy to interpret.
Acoustic emission and ultrasound detect friction, leaks and early bearing damage before vibration shows it, at the cost of higher sampling rates and more specialist interpretation.
Reading the signatures
Vibration
A vibration signal is sampled at a rate set by the highest frequency of interest; the Nyquist criterion requires at least twice that rate, and practical systems sample higher. Bearing defect frequencies scale with shaft speed and depend on the bearing geometry, which is why speed must be measured or known.
The basic analysis is the Fourier transform, which turns a time window into a spectrum. Imbalance shows as a peak at shaft speed, misalignment at twice shaft speed, gear mesh faults at the number of teeth times shaft speed, and bearing faults at the characteristic frequencies computed from geometry. Overall RMS velocity trended over time is the simplest health indicator, and the ISO 20816 series gives severity zones for it by machine class.
Early bearing faults hide as tiny impacts riding on a large low-frequency signal. Envelope analysis, which band-pass filters around a resonance and demodulates, pulls the impact repetition rate out so it can be matched against the bearing frequencies. Time-domain features such as kurtosis and crest factor also rise sharply when impacts begin.
These spectra and features are the inputs to any model. Feeding raw samples into a neural network rarely beats well-chosen features on small data, and the features are cheap enough to compute on a microcontroller.
Motor current
Motor current signature analysis reads the same mechanical story through the electrical side. Load variations modulate the current, so faults appear as sidebands around the supply frequency. Broken rotor bars, eccentricity and some bearing and gearbox problems have recognizable signatures. The attraction is that the sensor lives in the electrical panel, the signal is immune to mounting problems, and one clamp per motor is cheap to deploy at scale. The limitation is lower sensitivity to early bearing damage than a well-placed accelerometer.
From monitoring to prediction
There is a ladder of increasing ambition, and each rung needs more data than the last.
Thresholds and trends: alert when RMS exceeds a limit or a trend crosses a line. No learning, easy to explain, prone to false alarms when conditions change.
Anomaly detection: learn the distribution of features during a baseline period and score how far a new window departs from it, with anything from statistical distance measures to autoencoders. This needs no failure data, which is why it is the usual first model, and it is small enough to run as TinyML on the sensor node. Its weakness is that it says “different”, not “why”.
Fault classification: train a model to name the fault from examples. This gives actionable diagnoses but needs labeled failure data, which is rare, expensive and often borrowed from test rigs that do not match the field.
Remaining useful life estimation: predict the time until failure from the degradation trajectory. It is the goal everyone states and the rung fewest deployments reach, because it needs multiple run-to-failure histories per asset class.
Most successful programs live on the second rung, with an analyst confirming anomalies and adding labels over time so that the third rung becomes possible.
Where to run it
The sensor node samples and extracts features, and may run anomaly detection itself so that it transmits a score rather than raw waveforms. This is what lets a wireless sensor last years on a battery.
A gateway aggregates many nodes, runs heavier models, and relays to the cloud. It is also where OPC UA or Modbus data from the machine controller, such as speed and load, is joined with sensor data to normalize for operating conditions.
The cloud stores history in a time-series database, retrains models, compares assets across sites, and hosts the dashboard planners use. The edge AI guide covers the trade-offs of that split.
Pitfalls, and when it pays off
Sensor mounting matters more than the model. A loose magnet mount filters out the high frequencies where early faults live.
Operating conditions change. A pump at a different speed or load produces a different spectrum, and a model that does not know the speed will call the change a fault. Capture context from the controller or the current signal.
Failures are rare, so evaluation is hard. A model that never alerts scores near-perfect accuracy. Judge on detected failures and false alarms per asset per month instead.
Alerts that nobody acts on are worse than none; they train the organization to ignore the system. Start with few assets, tune until the maintenance team trusts it, then scale.
With those caveats, predictive maintenance pays when downtime is costly, when failures are gradual rather than sudden, when assets are too numerous for walk-around inspection, and when the maintenance organization can act on a warning that arrives days or weeks ahead. It does not pay on cheap, redundant or fail-safe equipment, nor on failure modes that give no warning. The industrial IoT guide places it in the wider context of the connected plant.
Where to go next
- TinyML explained covers running anomaly detection on the sensor node itself.
- Edge AI explained covers the gateway tier and the placement decision.
- AI agents on the factory floor shows how maintenance alerts can feed an agent that drafts work orders.
- Industrial IoT: a practical guide is the wider context.
Topics: predictive maintenance, condition monitoring, vibration, anomaly detection