The whole chain on one page
Prognostics and health management (PHM) is the discipline of turning measurements into maintenance decisions. In an AI-driven programme, machine learning does the pattern recognition, but data plumbing, context and the feedback loop decide whether the result is trusted and used.
1 · Sense: what to measure
Choose sensors for the failure modes you want to catch, not the other way round. A failure modes and effects analysis (FMEA) on the asset tells you which physical symptoms appear first.
| Data type | Typical rate | What ML can find |
|---|---|---|
| Vibration (accelerometer) | 5–50 kHz | Bearing, gear, unbalance, misalignment, looseness, resonance |
| Motor current (CT) | 2–20 kHz | Broken rotor bars, eccentricity, load oscillation, some bearing faults |
| Temperature, pressure, flow | 1 s–1 min | Process deviations, fouling, cooling loss, efficiency drift |
| Acoustic emission, ultrasound | 100 kHz+ | Very early bearing distress, leaks, partial discharge |
| Images (visual, thermal) | per inspection | Corrosion, cracks, hot spots, misaligned couplings, leaks |
| Maintenance records | per event | Labels for training, failure history, recurring problems |
2 · Edge: reduce data close to the machine
A single accelerometer at 25.6 kHz produces about 2 GB a day. Sending everything to the cloud is expensive and rarely necessary. Edge devices typically capture short windows at set intervals or on events, then compute features or compact embeddings locally. Some run a small anomaly model on the device itself so that alerts do not depend on the network.
Classical baseline · data collector routes
A technician walks a route with a portable analyzer once a month. It is still useful for wide coverage, but a monthly snapshot misses faults that develop in days and gives ML very few points per asset.
3 · Models: detect, diagnose, prognose
Three model families sit side by side, often fed by the same features:
- Detection produces an anomaly score per window. It is trained on healthy data and answers "is this unusual?" Read more.
- Diagnosis produces a fault class with a confidence. It needs labelled examples and answers "what is it?" Read more.
- Prognosis produces a remaining useful life (RUL) with an interval. It needs degradation histories and answers "how long?" Read more.
Run them in sequence. Diagnosis on a window the detector called normal produces confident nonsense, so most systems only classify after detection fires.
4 · Decide and act
An alert is worthless until someone plans work around it. The decision layer turns model outputs into a ranked list that a planner can use:
- Explain each alert with the features that drove it and a chart of the trend. Technicians trust what they can check.
- Prioritise by risk: asset criticality × probability of failure × consequence. A small anomaly on a spare-less compressor outranks a large one on a redundant pump.
- Integrate with the CMMS (computerized maintenance management system) so an accepted alert becomes a work order with the right parts and skills.
5 · The feedback loop
When a technician opens the machine, what they find is the most valuable data in the whole programme. "Outer race spall, 3 mm" turns an unlabeled anomaly into a labelled training example. "Nothing found" is equally valuable: it tells you the alert was a false positive. Capture both in a structured form, not free text, and retrain on a schedule.
Key takeaways
- Pick sensors from failure modes, not from catalogues.
- Process at the edge to cut data volume and keep alerts working offline.
- Run detection first, then diagnosis and prognosis on flagged windows.
- Close the loop: technician findings are your labels, and your measure of model quality.