What a PdM program is
Predictive maintenance (PdM), also called condition-based maintenance, means measuring the condition of equipment while it runs and acting when the measurements show that a failure is developing. The aim is simple: find problems early enough to fix them on your schedule, not the machine's.
A working program has five parts. Missing any one of them is the most common reason programs stall after the pilot.
| Part | Question it answers | Typical owner |
|---|---|---|
| Scope | Which assets are worth monitoring, and why? | Reliability engineer with operations |
| Technology | Which techniques, sensors, data settings and networks? | Condition monitoring lead, IT and OT |
| People | Who reviews the data, who decides, what skills do they need? | Maintenance manager |
| Workflow | How does a warning become a planned work order and a lesson learned? | Planning and scheduling |
| Measurement | Is the program catching failures and paying for itself? | Program manager |
The ten-step roadmap
The steps below follow the logic of international guidance on condition monitoring (ISO 17359 sets out the general procedure; ISO 55000 frames it within asset management). Each step links to the page in this track that covers it in detail.
- Set goals and a baseline. Write down what you want to improve: unplanned downtime hours, maintenance cost, safety incidents, energy. Record today's values so you can show progress later.
- Build the asset register. List every rotating and electrical asset in scope with its tag, type, driver, speed, bearings, power, duty and location. Data quality here decides the quality of every later step.
- Rank assets by criticality. Score the consequence of failure (safety, environment, production, cost, quality) and its likelihood. Split assets into tiers A, B and C. Asset selection and criticality →
- Understand how they fail. For each asset class, list the dominant failure modes and whether they give a warning. A failure mode effects and criticality analysis (FMECA) is the structured way to do this. Failure modes and techniques →
- Choose a maintenance strategy per failure mode. Condition-based where there is a detectable P-F interval, time-based where wear is age-related and cheap to fix, run-to-failure where failure is harmless. Strategies and the P-F curve →
- Select techniques and measurement intervals. Match each failure mode to the technique that detects it earliest, and measure at least twice within its P-F interval.
- Design the data collection. Online wired, online wireless or portable routes; sensor type, mounting, placement, sampling settings, network and gateways. Vibration sensors → and IIoT and wireless →
- Pilot, baseline and set alarms. Start with 20 to 50 assets, collect a few weeks of healthy data, then set statistical and model-based alarms on top of standard severity limits.
- Connect to the work management process. Every confirmed finding becomes a work order with a priority and a due date. Every repair is closed with what was found.
- Measure, learn and scale. Track catch rate, warning time, false alarms and cost avoided. Expand to more assets and techniques when the pilot shows results. Rollout, alarms and KPIs →
Do it with the planning tool
The PdM Program Setup app follows these steps: describe your assets, score their criticality, and get recommended techniques, sensor counts, sensor types, mounting, data settings and measurement intervals.
Why programs fail, and how to avoid it
| Common failure | What it looks like | Prevention |
|---|---|---|
| Technology first | Sensors bought before anyone decided which assets and failure modes matter | Do steps 1–6 before choosing hardware |
| Data without decisions | Dashboards full of alerts that nobody owns | Name an owner for every alert; route findings into work orders |
| Alarm flood | Hundreds of alerts in week one; people stop looking | Baseline first; tier alarms by criticality; review and tune monthly |
| Wrong sensor or mounting | Bearing faults missed because data above 2 kHz is not real | Match sensor bandwidth and mounting to the frequencies you need |
| No feedback loop | Repairs closed without recording what was found | Capture as-found condition and photos; use them to confirm or tune models |
| No proof of value | Program cut at the next budget review | Record each catch with the failure cost it avoided |
Where machine learning fits
Most of this track is about program design, because no model can rescue a program that monitors the wrong assets with the wrong sensors. Once the data is flowing, machine learning makes a program scale:
- Baselining and anomaly detection learn what normal looks like for each machine and operating state, so alarms adapt to the asset instead of using one fixed limit. See anomaly detection.
- Fault classification suggests a likely cause (unbalance, misalignment, bearing, gear) so analysts start from a hypothesis. See fault classification.
- Remaining useful life estimates how long the asset can keep running, which is what planners need. See prognosis.
Wireless sensors that report every hour create far more data than a monthly route. Without automated screening, analysts drown in it. That is why IIoT and machine learning are usually deployed together.
Key takeaways
- A PdM program has five parts: scope, technology, people, workflow and measurement.
- Decide which assets and failure modes matter before choosing sensors.
- Measurement intervals come from the P-F interval of the failure mode, not from habit.
- Value comes from planned work orders, not from alerts. Close the loop and record every catch.