Analytical & Predictive
Predictive Maintenance for Manufacturing
Predictive maintenance estimates equipment condition and future risk so maintenance can be planned around evidence instead of fixed intervals or unexpected failure. It begins with a critical asset and a defined failure mode, not with an algorithm.
Sense
Condition and context
Detect
Deviation and cause
Plan
Risk and intervention
Learn
Work order outcome
What is predictive maintenance?
Predictive maintenance uses condition and operating data to estimate whether an asset is degrading, what may fail and when intervention may be needed. It is part of a wider asset condition management system that connects sensing, diagnostics, prognostics and maintenance decisions.
The output should support a maintenance action: inspect, lubricate, align, replace, reschedule or continue operating. A dashboard that only displays vibration or temperature trends is condition monitoring, not a complete predictive maintenance workflow.
Which data is useful for predictive maintenance?
| Data | What it can reveal | Context required |
|---|---|---|
| Vibration | Imbalance, looseness, bearing or alignment changes | Speed, load, sensor position and sampling method. |
| Electrical current | Load changes, motor or process abnormalities | Operating state, setpoint and product. |
| Temperature and pressure | Friction, cooling, flow or process drift | Ambient conditions and operating mode. |
| Cycle and runtime counters | Usage, wear exposure and maintenance intervals | Asset identity and counter reset history. |
| Alarms and PLC states | Failure sequences and abnormal transitions | Event time, code definitions and machine mode. |
| Work orders and failure notes | Observed failure mode and intervention outcome | Consistent asset and component taxonomy. |
How is a predictive maintenance use case built?
- Rank assets by safety, quality, delivery and financial consequence.
- Choose a failure mode that can be detected early enough to act.
- Confirm that available signals represent the relevant physical behaviour.
- Establish normal operating modes and data quality rules.
- Build a baseline before introducing a more complex model.
- Define warning levels, response owners and permitted actions.
- Connect alerts to inspections and work orders.
- Use maintenance findings to relabel and improve the model.
Which predictive maintenance model should be used?
Use thresholds and trend rules when failure physics and tolerances are known. Use anomaly detection when normal behaviour is available but failures are rare. Use supervised classification when historical cases reliably identify failure states. Estimate remaining useful life only when degradation trajectories and maintenance endpoints are sufficiently consistent.
How should business value be evaluated?
- Warning lead time before the intervention window closes.
- False alarms and unnecessary inspections per period.
- Failures detected, missed and correctly rejected.
- Avoided downtime, scrap, expedited parts and secondary damage.
- Maintenance labour, sensor, integration and model operating cost.
- Performance across assets, products and operating conditions.
NIST research highlights that condition-monitoring evaluations are difficult to compare when engineering effects, uncertainty and economic analysis are reported inconsistently. A credible project therefore defines its baseline and evaluation method before the pilot begins.
Frequently asked questions
Can predictive maintenance work on older machines?
Yes, when useful condition or operating data can be obtained from existing controllers, external sensors, maintenance records or a combination of sources. The economic case must include the cost of instrumentation and integration.
How much failure history is needed?
Supervised failure models need representative labelled events. When failures are rare, teams can begin with physics-based limits, normal-behaviour models and structured inspection feedback.
Should the model create work orders automatically?
Only when permissions, confidence, failure consequence and operating policy allow it. Many deployments begin by drafting or recommending a work order for human approval.
Continue learning