The maintenance hierarchy
Three modes of maintenance, none of them universally correct:- Reactive — fix it when it breaks. Cheapest per intervention, most expensive per year for critical equipment.
- Preventive — fixed schedule based on hours / cycles / calendar. Cheap to design, often over-services equipment that doesn't need it and under-services equipment that does.
- Predictive (CBM, condition-based) — service based on actual condition (vibration, temperature, oil analysis, current signature). Most expensive to set up, lowest total cost on the right equipment.
A working programme is a deliberate mix of all three, applied per asset based on criticality and failure cost.
Asset criticality first, then maintenance mode
For each asset, score:- Operational criticality — does failure stop the line, or is there a workaround?
- Cost of failure — direct (parts, labour) plus indirect (downtime, scrap, customer impact)
- Failure mode predictability — does it degrade gradually (good for CBM) or fail catastrophically (CBM less useful)?
- Inspection cost — how cheap is the condition signal to acquire?
The matrix output:
- High-criticality, predictable degradation, cheap signal → predictive
- High-criticality, unpredictable degradation → preventive (with margin) + spare on shelf
- Low-criticality → reactive (or preventive on a long cycle)
- Anything safety-relevant → preventive, no exceptions
Predictive maintenance: what actually works
After several "predictive maintenance" pilots in industrial environments, the patterns that consistently produce results:- Vibration monitoring on rotating machinery (motors, pumps, fans). Mature technology, clear failure signatures (bearing defects, imbalance, misalignment).
- Motor current signature analysis — non-invasive, cheap, catches a lot of motor and load issues.
- Oil analysis on gearboxes / hydraulic systems — slow signal, but often the earliest indicator.
- Temperature trending on power electronics — drives, transformers. Signals capacitor degradation, cooling failure.
- Acoustic monitoring on compressors and steam systems — catches leaks early.
Patterns that we've seen oversold:
- Ultra-sophisticated ML on small datasets — vendor pitches that promise "AI predicts failure 30 days in advance" almost always fail to deliver. The simple statistical methods on good sensor data win.
- Cloud-only PdM platforms with monthly subscription per asset — economics rarely work for low-failure-cost equipment.
The CMMS — the boring half nobody discusses
The work order system (Computerised Maintenance Management System) is the real maintenance backbone. Predictive signals are useful only if they generate work orders that actually get done. The CMMS must support:- Work order generation from triggers (schedule, condition alarm, operator request)
- Parts inventory linked to work orders
- Time tracking linked to work orders
- Equipment history queryable by asset
- Reports by asset, by failure code, by technician
Most factory PdM projects we've seen succeed had a working CMMS first. The ones that failed tried to deploy condition monitoring without the work-order backbone.
The data your maintenance team owns
Track per-asset:- MTBF (mean time between failures) — trended, not just current
- MTTR (mean time to repair) — trended
- Cumulative maintenance cost (parts + labour)
- Failure mode codes — categorised, queryable
- PdM signal trends, where applicable
The asset that's failing more often this year than last year is the asset to investigate, before it surprises you.
One pattern that pays off
Reliability-Centered Maintenance (RCM) reviews on the top-10 failure-cost assets, every 2 years. Not the expensive consultancy version — a half-day workshop with the maintenance lead, the production lead, and an engineer. The conversation always reveals at least one asset that's being over-maintained and one that's being under-maintained.One pattern that doesn't pay off
Buying a "predictive maintenance platform" before having a working CMMS. The platform without the backbone produces alerts that nobody acts on. Solve the work-order side first, layer the predictive on top.What's your maintenance stack? And — for industrial folks here — has anyone successfully replaced a commercial CMMS with self-hosted (e.g. Snipe-IT + custom workflows)?
