Pilot first
A pilot tests the whole chain, from sensor to work order, on a manageable scale before you commit budget to the whole site.
- Scope: 20 to 50 assets in one area, mostly tier A and B, with a mix of types and some known bad actors.
- Duration: 3 to 6 months, long enough to see real faults develop and be repaired.
- Success criteria written up front: for example, at least three confirmed catches with recorded cost avoided, fewer than one false alarm per 10 assets per month after tuning, and every alert reviewed within the agreed time.
- Technical checks: radio coverage, battery drain against forecast, data completeness, integration with the maintenance system.
Baseline, then alarm
After installation, collect two to four weeks of data across the normal range of operating conditions before trusting alarms. Then build alarms in layers:
| Layer | How it works | Strength |
|---|---|---|
| Severity limits | Broad-band velocity compared with machine-class limits such as those in the ISO 20816 series | Simple, comparable across sites; a sensible absolute ceiling |
| Statistical baselines | Alarm when a value moves several standard deviations from the asset's own baseline, or rises by a set ratio | Adapts to each machine; catches change early |
| Band and fault-frequency alarms | Energy in bands around bearing, gear or vane frequencies | Points to a likely cause |
| Rate of change | Alarm when the trend slope exceeds a limit | Catches fast-developing faults |
| Machine learning | Anomaly scores on many features at once, conditioned on operating state | Fewer false alarms in variable operation; scales to many assets |
Use two or three alarm levels (for example alert and danger) with clear meanings and response times, and set the sensitivity by criticality: tighter on tier A, looser on tier C. Review false and missed alarms every month for the first six months.
From alert to work order
- Write recommendations that a planner can act on: "Replace motor non-drive-end bearing within 4 weeks; outer race defect, moderate severity; check alignment on reassembly."
- Give every recommendation a priority mapped to your work order system, and track it to closure.
- Record what was found when the machine was opened. That feedback confirms the diagnosis, labels data for machine learning, and is the evidence for the value of the program.
- Hold a short weekly review between condition monitoring, planning and operations for open findings.
Roles and skills
| Role | Responsibilities |
|---|---|
| Program owner | Goals, budget, KPIs, cross-department alignment |
| Reliability engineer | Criticality, failure modes, strategy, root cause analysis |
| Condition monitoring analyst | Data review, diagnosis, recommendations, alarm tuning |
| Technicians | Sensor installation and mounting quality, routes, corrective work, as-found feedback |
| Planner and scheduler | Turn recommendations into scheduled work with parts and permits |
| OT and IT | Network, gateways, security, integration, data retention |
| Data scientist (as the program grows) | Models, feature pipelines, monitoring of model performance |
Analyst competence is recognised internationally through ISO 18436-2, which defines four categories of vibration analyst from basic data collection (category I) to advanced diagnostics and program leadership (category IV). Other parts of ISO 18436 cover lubrication, thermography and ultrasound practitioners. Even with machine learning, someone with category II or III skills should review findings on critical assets.
KPIs that show whether it works
| KPI | Definition | Direction |
|---|---|---|
| Coverage | Share of tier A and B assets on the program | Up, towards 100% of tier A |
| Data completeness | Share of scheduled measurements actually received | Above 95% |
| Catch rate | Failures detected before functional failure ÷ all failures on monitored assets | Up |
| Mean warning time | Average time between first alert and functional failure or repair | Up |
| False alarm rate | Alerts not confirmed by review or inspection ÷ all alerts | Down |
| Recommendation closure | Recommendations completed by the due date | Up |
| Planned work share | Planned maintenance hours ÷ total maintenance hours | Up |
| Unplanned downtime | Hours lost to failures on monitored assets | Down |
| Cost avoided | Sum of estimated failure cost minus planned repair cost for each catch | Up |
A simple return on investment
A first-pass business case needs only a few numbers: how many failures happen per year on the assets in scope, what an average failure costs (repair, downtime, secondary damage), what share the program will catch, and what it costs.
Do not forget benefits that are harder to count: safety (fewer people near running machines on routes), energy (misalignment and unbalance waste power), spare parts inventory (order on condition instead of holding everything), and fewer secondary failures.
Scaling up
- Expand by area or asset class, reusing templates for sensor placement, data settings and alarms.
- Add techniques where the pilot showed blind spots: oil on gearboxes, ultrasound on slow bearings, motor current on remote motors.
- Introduce machine learning once you have consistent data and feedback labels from a few months of operation.
- Feed findings into proactive work: precision alignment, balancing and lubrication practice remove the causes that PdM keeps finding.
Key takeaways
- Pilot 20–50 assets for 3–6 months with written success criteria.
- Baseline before alarming; combine severity limits, statistical baselines and models.
- Every alert needs an owner, a recommendation and a closed work order with as-found feedback.
- Track catch rate, warning time, false alarms and cost avoided to prove value.
Plan your program
Use the PdM Program Setup app to turn your asset list into a monitoring plan with sensor counts and settings you can export.