Predictive maintenance
Predictive maintenance replaces the calendar with the machine's real condition: measure continuously, detect drift, act just before failure. The challenge is not the algorithm but the data — scarce, noisy, and rarely labelled with real failures.
The technical building blocks
Spectral & envelope analysis
The vibration signature reveals imbalance, misalignment, bearing and gear faults long before the human ear. Accelerometers and FFT are the foundation.
MCSA & thermography
Motor current signature analysis (MCSA) detects broken rotor bars and eccentricity; thermography spots hot points and loose connections.
Sensors & edge gateways
Wireless sensors, high-frequency sampling and edge-gateway preprocessing cut the data volume sent up and enable real-time alerting.
Anomaly detection & remaining useful life
Lacking failure examples, models learn normality and flag deviation (anomaly detection); remaining-useful-life (RUL) models estimate time to failure.
Condition-monitoring techniques
| Technique | Measured quantity | Faults detected | Reference |
|---|---|---|---|
| Vibration analysis | velocity / acceleration (mm/s) | imbalance, misalignment, bearings (BPFO/BPFI), gears | ISO 20816 / 13373 |
| Envelope analysis | HF shock demodulation | incipient bearing faults | — |
| IR thermography | surface temperature (°C) | hot spots, loose connections, friction | ISO 18434 |
| MCSA | stator current spectrum | broken rotor bars, eccentricity | — |
| Oil analysis | particles, viscosity, water | wear, contamination, lubricant degradation | ISO 4406 |
| Ultrasound | acoustic emission 20-100 kHz | leaks, electrical discharge, bearing lubrication | — |
No single technique covers everything: vibration, thermography and oil analysis are combined per machine. The rule is to detect as early as possible — the whole point of the P-F curve below.
From calendar to predictive
Preventive maintenance replaces healthy parts "to be safe"; corrective waits for breakage. Predictive aims at the middle: act at the right time, neither too early nor too late. Four maturity levels describe this journey.
| Level | Principle | Limit / cost |
|---|---|---|
| Reactive | repair after failure | unplanned downtime, breakage, safety risk |
| Preventive | replace on a schedule | healthy parts scrapped, over-maintenance |
| Condition-based | measure, act at threshold | fixed thresholds, limited anticipation |
| Predictive | model the drift, estimate RUL | needs data, models, CMMS integration |
Detecting early: the P-F curve
Between the potential-failure point (P), where a defect becomes detectable, and functional failure (F), an interval elapses: the P-F window. All condition-based maintenance rests on one rule — the inspection interval must be shorter than the P-F window. The earlier a technique detects (vibration before heat, heat before audible noise), the wider the window and the more schedulable the intervention.
ISO 13374 structures the processing into six blocks (OSA-CBM architecture): data acquisition → manipulation → state detection → health assessment → prognostics → advisory generation. Classic preventive maintenance stops at state detection; the truly predictive value lives in the last two blocks — prognostics and advisory.
Key challenges
- Failure scarcity — good equipment rarely fails: few positive examples, heavily imbalanced classes. One often learns normality rather than failure.
- False alarms vs. misses — too many alerts and the operator ignores the system; too few and the failure slips through. Threshold tuning is a business trade-off, not just statistics.
- CMMS integration — a prediction is only valuable when turned into a work order: the link to the CMMS (SAP PM, Maximo) and planning is decisive.
- Prognostic uncertainty — a remaining-useful-life figure without a confidence interval is misleading; uncertainty must be shown to decide rationally.
Standards & references
- ISO 17359 — General guidelines for condition monitoring and diagnostics of machines.
- ISO 13374 — Data processing, communication and presentation for condition monitoring (OSA-CBM architecture).
- ISO 13379 / 13381 — Data interpretation and prognostics techniques (remaining-useful-life estimation).
- NAMUR NE 107 — Self-monitoring status signalling of field devices: maintenance required, out of specification, failure, function check.
Related standard pages
Major players
Dedicated platforms
Senseye (Siemens), Augury, Uptake, C3 AI, TwinThread.
Sensors & monitoring
SKF, Bently Nevada (Baker Hughes), ifm, Fluke, Banner.
Industrial suites
AVEVA Predictive Analytics, GE Vernova, Aspen Mtell, Seeq.