Predictive maintenance is often sold as a model that forecasts a breakdown. In a factory, the useful product is different: enough reliable warning for a competent person to inspect the right asset, choose a safe intervention and prevent an avoidable production or safety loss.
That distinction matters. A highly accurate anomaly score can still fail if it arrives after the planning window, floods engineers with alarms, ignores operating context or creates a maintenance job that cannot be performed safely.
The UK government’s June 2026 Advanced Manufacturing AI Adoption Plan recommends a practical “Scan–Pilot–Scale” pathway, supported by trusted operational data, workforce capability and factory validation. Predictive maintenance fits that pathway well because the decision, asset and outcome can be bounded. See the AI Adoption Plan for Advanced Manufacturing.
Know Which Maintenance Problem You Are Solving
| Approach | Trigger | Best fit | Limitation |
|---|---|---|---|
| Reactive | Equipment has failed | Cheap, non-critical and easily replaced assets | Downtime and secondary damage are accepted |
| Preventive | Calendar, hours or cycles | Known wear pattern and predictable service need | Work may happen too early or miss unusual deterioration |
| Condition-based | Measured condition crosses a rule | Failure has an observable physical precursor | Fixed threshold may not adapt to load or environment |
| Predictive | Model estimates failure risk or remaining useful life | Sufficient history, context and planning lead time exist | Sparse failures and changing operation can mislead |
| Prescriptive | System recommends an intervention | Mature process with trusted costs, constraints and approvals | Bad assumptions can optimise the wrong objective |
Do not start with the most complex method. A well-designed vibration threshold linked to a work order may outperform a remaining-useful-life model that nobody trusts.
ISO 17359 provides a general framework for setting up a machine condition-monitoring programme and remains current after its 2023 review. It is a useful reminder that sensing, diagnosis and maintenance decisions form one programme rather than isolated technology purchases. See ISO 17359:2018.
Select the Asset by Decision Value
The best pilot asset is not always the machine with the most data. Rank candidates using five questions:
- Consequence: what happens to safety, quality, delivery and adjacent equipment when it fails?
- Detectability: does deterioration create a measurable precursor such as vibration, temperature, current, pressure, acoustic change or lubricant condition?
- Lead time: can the signal arrive early enough to inspect, obtain parts and schedule isolation?
- Actionability: is there a defined inspection or maintenance action for the warning?
- Evidence: can historic faults, work orders and operating context be reconstructed?
Avoid a first pilot where failures are extremely rare, labels are unreliable, every unit is unique or an alert cannot change the maintenance decision.
Start with a failure-mode review led by maintenance and operations. For each mode, record the physical mechanism, precursor, safe response, planning horizon and cost of a missed or false warning. That becomes the specification for data and evaluation.
Build a Contextual Asset Record
A sensor stream without operating context confuses normal variation with damage. A motor running hot under exceptional load is different from the same temperature at normal load.
A useful asset record combines:
- asset hierarchy, make, model, installation and component relationships;
- timestamped sensor readings with units and calibration information;
- run state, load, speed, recipe, product and ambient conditions;
- alarms and control-system events;
- operator observations;
- inspection findings and replaced-component condition;
- work orders, fault codes, causes and actions;
- changes to firmware, lubrication, alignment, suppliers and operating practice.
Align clocks before modelling. Preserve missingness rather than silently filling every gap. A missing sensor, a stopped machine and a communications failure have different meanings.
The 2024 edition of ISO 55000 emphasises managing assets over their life cycles and aligning asset decisions with organisational value. Predictive maintenance should therefore connect to the asset-management plan, not sit as a data-science experiment. See ISO 55000:2024.
Use a Model Ladder
Move up the complexity ladder only when the lower step cannot support the decision.
1. Engineering limits
Use manufacturer limits and engineering rules. They are transparent, quick to validate and useful for known failure conditions.
2. Condition trends
Track rate of change, rolling baselines and relationships between signals. This often detects deterioration without needing labelled failures.
3. Anomaly detection
Learn normal operation for a defined operating regime and highlight deviations. Treat an anomaly as a request for investigation, not a diagnosis.
4. Failure classification
When labelled examples are credible, estimate which known failure mode is likely. Validate per failure mode because an overall accuracy score can hide poor detection of a critical fault.
5. Remaining useful life
Use only where degradation paths and maintenance horizons are stable enough to support the estimate. Provide a range and confidence, not a fictional exact date.
Large language models can help retrieve manuals or summarise work history, but they should not invent sensor values, diagnose from unsupported evidence or issue a safety-critical instruction.
For another bounded factory use case, see Computer Vision Quality Control.
An Illustrative Gearbox Workflow
This is a worked design example, not a claimed deployment.
A site selects a conveyor gearbox whose bearing deterioration produces vibration and temperature changes before seizure. The current process is time-based inspection plus operator reports.
The pilot collects vibration features, bearing temperature, motor current, line speed, product load and maintenance history. Engineers define operating regimes and exclude shutdown, washdown and known sensor-fault periods from the normal baseline.
The model produces three states:
- Normal: continue monitoring.
- Investigate: engineer reviews the trend and performs a defined inspection.
- Urgent: existing safety and maintenance procedures determine whether the equipment is stopped.
The model never bypasses the control system, changes guarding or authorises work. The engineer records whether a defect was found, the actual component condition and the action taken. That feedback improves both the maintenance record and future evaluation.
The business result is assessed through avoided disruption, planning lead time and unnecessary inspections—not the number of alerts generated.
Safety Duties Do Not Move to the Model
PUWER requires work equipment to be suitable, maintained in an efficient state and inspected where deterioration could create danger. Maintenance logs must be kept up to date where they exist. See the HSE PUWER overview.
HSE also states that maintenance should be planned, performed by competent people and normally carried out with plant isolated and made safe. An AI alert changes none of those requirements. Review the HSE maintenance guidance.
Write the operating boundary explicitly:
- AI may recommend inspection or prioritisation.
- Established controls and competent people determine safe shutdown.
- Lock-off, stored-energy release and safe access procedures remain mandatory.
- Protective functions do not depend on a cloud service or predictive model.
- Loss of the model returns the site to a documented maintenance method.
- A model or sensor change follows engineering change control.
Where the system affects machinery design or a safety function, obtain competent product-safety advice rather than assuming that “decision support” describes its real effect.
Protect the OT Boundary
Predictive maintenance increases connectivity between sensors, operational technology, historians, cloud services, suppliers and maintenance systems. That can improve visibility while expanding the attack surface.
The NCSC’s January 2026 OT guidance says organisations should make risk-informed connectivity decisions, document third-party exposure and preserve safety, reliability and availability. See its secure connectivity guidance for operational technology.
At minimum:
- maintain an authoritative asset and data-flow inventory;
- separate collection from control where possible;
- permit only required protocols, directions and destinations;
- authenticate devices, users and service accounts;
- protect engineering data and credentials;
- log configuration, model and access changes;
- define offline operation and recovery;
- test vendor exit and loss of cloud connectivity;
- prevent model output from becoming an unauthorised control command.
The NCSC’s secure AI development guidance also recommends security across design, development, deployment and ongoing operation.
Measure Decision Quality
| Measure | What it reveals | Common trap |
|---|---|---|
| Detection lead time | Whether warning arrives before the planning deadline | Celebrating an alert that arrived too late |
| Precision | Share of alerts that correspond to a useful finding | Ignoring inspection cost and alarm fatigue |
| Recall by failure mode | Share of relevant faults detected | Hiding a missed critical mode in an average |
| False alarms per asset-month | Operational burden | Reporting only a percentage |
| Warning-to-work-order conversion | Whether staff can act on the signal | Generating alerts outside the maintenance process |
| Unplanned downtime | End-to-end operational result | Claiming every reduction was caused by AI |
| Planned-to-reactive work ratio | Whether planning improved | Moving labels without changing work |
| Model coverage | Share of operation inside the validated range | Producing scores for unfamiliar regimes |
Use a time-based holdout so future records are not leaked into training. Test after maintenance strategy, product, supplier or operating changes. Include sensor outages and incorrect work-order labels.
Pilot Without Disrupting Production
Scan: select one asset class, map failure modes and establish a baseline. Confirm that the intervention is safe and valuable.
Instrument: repair timestamp, asset-ID and work-order problems before buying more sensors. Document units, calibration and expected operating ranges.
Shadow: run alerts without changing official maintenance. Engineers classify the alert and record evidence.
Assist: connect validated alerts to the CMMS as reviewable recommendations with ownership and due dates.
Scale: expand only to comparable assets after performance remains stable across shifts, loads and seasons. Revalidate rather than copying a threshold blindly.
Budget for integration, engineering review, sensor maintenance, security and change—not just model development. The Production AI Cost Stack provides a useful wider budget framework.
The Goal Is Better Maintenance Decisions
Predictive maintenance succeeds when it gives competent teams reliable time to act. It fails when anomaly scores become another unowned alarm feed.
Begin with the physical failure, the safe intervention and the planning window. Build the data around that decision, validate under real operating conditions and keep the model outside the machine’s protective controls. That approach produces a system engineers can challenge, operations can use and safety teams can defend.
