Predictive maintenance is an approach that monitors equipment condition data such as vibration, temperature and current to anticipate when a failure will develop and schedules maintenance work based on that forecast. The goal is not to put sensors on every asset, but to catch failures at a detectable early stage on the assets whose downtime costs the most. This guide gives maintenance and production teams a roadmap for moving to predictive maintenance in a controlled, measurable way.
You can use this guide as a project control tool: tick the checklists in each section to track progress with your team.
Maintenance maturity levels
Predictive maintenance is not built in a single step. Most plants run different assets at different levels, and that is normal. What matters is choosing the right level for each asset deliberately. Running a low-criticality fan to failure can be a reasonable decision; running the main compressor to failure is not.
- L1ReactiveAct after failure
- L2PreventiveCalendar or run hours
- L3Condition-basedMeasured value crosses threshold
- L4PredictiveModel forecast and trend
| Level | Trigger | When does it fit? |
|---|---|---|
| Reactive | After failure occurs | Cheap, redundant, non-critical assets |
| Preventive | Calendar or run hours | Parts with clear age-related wear |
| Condition-based | Measured value crosses threshold | Failures with a measurable precursor |
| Predictive | Model forecast and trend | Critical assets with data and known failure modes |
Condition-based maintenance is the prerequisite for predictive maintenance. A team that first learns to collect the right signals reliably and manage threshold alarms moves to model-based forecasts on much firmer ground. Do not remove your preventive maintenance plans either; predictive maintenance helps you trim and focus those plans on the basis of evidence.
Critical asset analysis
The first deliverable of predictive maintenance is not a sensor list but a prioritised asset list. Criticality analysis weighs the impact of an asset's failure on production, safety, environment, quality and cost together with the likelihood of that failure. Run this analysis with maintenance, production, quality and health and safety representatives; a single department's view is misleading.
| Impact / Likelihood | Low likelihood | Medium likelihood | High likelihood |
|---|---|---|---|
| High impact | Medium priority | High priority | Highest priority |
| Medium impact | Low priority | Medium priority | High priority |
| Low impact | Can stay reactive | Low priority | Medium priority |
When scoring impact, write down factors such as redundancy, alternative routing, spare part lead time and knock-on effects explicitly. For likelihood, use maintenance history, operator observations and manufacturer recommendations. The team should agree on the scoring scale (for example three or five levels) up front, and the meaning of each level should be written down.
Failure modes and FMEA basics
You cannot decide which signal to monitor without knowing why and how an asset fails. Failure Mode and Effects Analysis (FMEA) systematically lists the possible failure mode, cause, effect and current detection method for each component. The FMEA method is defined in IEC 60812; in a maintenance context, the goal is to map failure modes to signals you can monitor.
- 01Define the function
Write down what the asset must do and its acceptable performance limits. A failure is the loss of that function.
- 02List the failure modes
Derive realistic modes such as bearing damage, imbalance, misalignment, insulation breakdown, cavitation and oil contamination from maintenance history and technician experience.
- 03Write the cause and effect
Describe each mode's root cause and its effect on production, safety and quality in short sentences.
- 04Find the precursor
What physical change can you observe while the failure develops? Specific frequencies in the vibration spectrum, rising temperature, current imbalance and so on.
- 05Choose the detection method and interval
Based on the time between the precursor and functional failure, decide whether you need continuous monitoring or periodic measurement.
The time between the precursor and functional failure is known in the literature as the P-F interval. If the interval is very short, periodic measurement does not help and you need continuous monitoring. If it is long, route measurements with handheld instruments may be enough and less investment is needed.
Choosing signals and sensors
Sensor selection follows the failure modes identified in the FMEA. Every sensor adds installation, calibration and data management load, so instead of “let's measure everything we can”, ask “which signal shows which failure mode earliest?”
| Signal | What it typically shows | Watch out for |
|---|---|---|
| Vibration | Bearings, imbalance, misalignment, looseness | Mounting point and sampling rate are critical |
| Temperature | Friction, lubrication, overload | Read together with ambient temperature |
| Motor current | Load changes, rotor and electrical issues | Normalise by operating regime |
| Pressure / flow | Pump wear, clogging, leaks | Relate to process conditions |
| Oil analysis | Wear particles, contamination, degradation | Sampling method must be consistent |
For vibration monitoring, the ISO 20816 series provides evaluation guidance by machine class; the general condition monitoring approach is described in ISO 17359. These standards are a starting reference; final alarm limits should be tuned to the asset's own operating history. Also inventory the current, pressure and temperature tags already in your PLC and SCADA systems; before buying new sensors, the data you already have is often enough for the first stage.
Data infrastructure
Many predictive maintenance projects stall on the data infrastructure, not the model. Signals must be collected at the right rate, timestamps must be consistent, the connection must not lose data during outages, and data must be stored with its context.
- 01Sensor and PLCRaw signal
- 02Edge devicePreprocessing, buffering
- 03HistorianTime series and context
- 04Analytics and modelThreshold, anomaly, forecast
- 05CMMS work orderPlanned intervention
Sampling and preprocessing
Vibration analysis needs high sampling rates, and the raw data volume is large. That is why the raw waveform is usually processed in the field on an edge device; summary features such as RMS, peak value and spectral bands, plus raw samples at set intervals, are sent to the central system. Slow-changing signals such as temperature, pressure and current need lower sampling rates.
Historian, connectivity and time sync
Time series data is kept in a historian or time series database. OPC UA and MQTT are common connection methods between OT and IT. All devices must be tied to the same time source (NTP or PTP) so that different signals can be compared around the same event. Plan for the edge device to buffer data during network outages and send it when the connection returns (store and forward).
A data management layer such as OREL DM helps bring sensor, historian, ERP and maintenance data together under the same asset ID. Whatever tool you choose, the core principle is the same: every data point must know which asset, which component and which operating condition it belongs to.
Maintenance history and CMMS data quality
For a model to learn failures, it needs to know what happened in the past. That information lives in the work orders kept in the computerised maintenance management system (CMMS). In many plants, however, work orders are recorded as free text, with missing failure codes or against the wrong asset. In that case even the best sensor data stays unlabelled.
- The asset hierarchy should use the same IDs in the CMMS, historian and ERP.
- Failure codes should be picked from a short, clear list: symptom, cause, action taken.
- Failure start time, response time and return-to-service time should be kept in separate fields.
- Replaced parts and findings (including photos) should be attached to the work order.
- Planned stops and breakdown stops should be recorded in separate categories.
Cleaning historical records takes time, so limit it to the pilot assets and, at the same time, put rules in place so new records are entered correctly. A maintenance management system such as OREL IMS makes this discipline easier by managing failure codes and the asset hierarchy as required fields.
Common mistakes: starting to build models before cleaning maintenance history; choosing sensors by ease of mounting instead of failure mode; interpreting signals without process context (mistaking vibration that rises with load for a fault); spreading the pilot over too many assets; going live without defining who receives the alarm.
First model: from thresholds to forecasts
Instead of targeting remaining useful life from day one, move in stages. Each stage builds on the previous one, and the team learns to trust the system's signal at every step.
- L1Threshold and trend
- L2Regime normalisation
- L3Anomaly detection
- L4Failure classification
- L5Remaining useful life
- 01Threshold and trend rules
Define warning and alarm thresholds using standard references and manufacturer recommendations. Track the rate of change as well as the absolute value. This stage delivers value quickly and tests data reliability.
- 02Normalisation by operating regime
Normal values change as speed, load and product type change. Group or normalise signals by regime; otherwise the false alarm rate stays high.
- 03Anomaly detection
If labelled failure data is scarce, use models that learn the asset's normal behaviour and flag deviations from it. An anomaly score alone is not a decision; it is assessed together with technician verification.
- 04Failure classification
As enough verified failure examples accumulate, move to models that suggest which failure mode an anomaly belongs to.
- 05Remaining useful life
Forecast remaining useful life (RUL) for failure modes with a consistent degradation curve and enough history. Always present the forecast with a confidence interval.
Assess model performance not only by accuracy but also by missed failures, false alarms and how far ahead of the failure the warning arrives. For the maintenance team, the most expensive error is often a missed failure, but constant false alarms also quickly erode the team's trust in the system.
Alarm handling and work order flow
A forecast creates value only when it reaches the right person and turns into a work order. Design the alarm flow as carefully as the model.
- 01Warning
- 02Review and verification
- 03Work order and planning
- 04Intervention
- 05Findings recorded
- 06Model and threshold update
- 01
- 1.A warning is raised and displayed with the asset, component, likely failure mode and the relevant chart.
- 2.The condition monitoring owner reviews the warning within the agreed time: confirms it, puts it on watch or rejects it.
- 3.For a confirmed warning, a maintenance request is opened with the recommended action and urgency.
- 4.Planning places the intervention in a suitable downtime window, considering spare parts and the production plan.
- 5.After the intervention, findings (was there really a fault, at what stage) are recorded on the work order.
- 6.Findings are fed back to the model; thresholds and the model are reviewed at regular intervals.
Pilot scope, roles and success criteria
The pilot should start with a small number of assets that scored high priority in the criticality analysis, have well-known failure modes and accessible data. Choosing a group of similar assets (for example pumps of the same type) makes comparison and generalising results easier.
| Role | Responsibility |
|---|---|
| Project owner | Goal, budget and decision authority |
| Maintenance lead | Failure modes, alarm response, work order flow |
| Condition monitoring specialist | Sensors, analysis, warning verification |
| OT / automation | PLC, network, edge device, data collection |
| IT / data | Infrastructure, security, data model, integration |
| Production representative | Downtime windows, process context |
Write the success criteria before the pilot starts and measure the baseline. Criteria can be qualitative and quantitative; what matters is that everyone uses the same definitions.
Scaling up and the ISO 55000 context
When the pilot succeeds, it is tempting to roll the same approach out to the whole plant at once. Instead, move in waves following the criticality list and reuse templates in each wave: a failure mode library per asset type, sensor package, feature set, threshold and alarm rules.
- Build asset type templates so that adding a new asset becomes a configuration task.
- Version models and thresholds so that who changed what and why is traceable.
- Plan the maintenance of the sensor and edge device fleet itself: batteries, calibration, software updates.
- Connect predictive maintenance outputs to spare parts planning and budgeting processes.
- Review preventive maintenance plans based on evidence and cut unnecessary interventions.
The ISO 55000 series defines asset management as a management system aligned with organisational objectives. Predictive maintenance is a tool within this framework: it helps you manage the balance of risk, performance and cost of assets based on data. ISO 55000 is a management system standard and does not mandate a specific technology or predictive maintenance method; if you are aiming for certification, assess the scope with a qualified consultant.
Conclusion
Moving to predictive maintenance is not a sensor purchasing project but a project about criticality, failure knowledge, data discipline and workflow. Choose the right assets, map failure modes to signals, fix your maintenance history, start with thresholds and progress in stages, and make sure every warning turns into a work order. ORELSIS supports you in applying this roadmap on the floor, bringing maintenance and data together in one flow with OREL IMS and OREL DM.