How to Build a Predictive Maintenance Platform Using IoT and Machine Data
How to build a predictive maintenance platform: machine data, thresholds vs anomaly detection vs failure prediction, architecture, rollout and pitfalls.
The short answer
A predictive maintenance platform collects condition data from equipment, spots deterioration early, estimates how urgent it is, and turns that into planned maintenance work. To build one, choose a small set of critical assets, capture the right signals (often vibration, temperature, current and controller events), start with condition monitoring and anomaly detection, connect alerts to your maintenance workflow, and move to failure prediction only once you have reliable maintenance history to learn from.
This guide is for manufacturers, utilities, facilities operators and equipment makers. It explains the levels of maturity, the architecture, the data you need, and how to roll out without drowning technicians in false alarms. For the maintenance records side, see asset maintenance management software.
Four levels of maintenance, and where prediction fits
"Predictive maintenance" is used loosely. Being precise about which level you are building avoids disappointed expectations.
| Level | How it decides | Data needed | Typical result |
|---|---|---|---|
| Reactive | Fix it when it breaks | None | Unplanned downtime and secondary damage |
| Preventive (time- or usage-based) | Service every N hours, cycles or months | Run hours or calendar | Fewer failures, but parts replaced early |
| Condition monitoring | Act when a measurement crosses a limit | Live sensor or controller data | Earlier warning on known failure modes |
| Predictive | Detect patterns of deterioration and estimate time to failure | Condition data over time, plus maintenance and failure history | Maintenance planned close to need |
Worth noting
Threshold alerts are condition monitoring, not prediction. They are still valuable, and most successful programmes start there. Just call them what they are when setting expectations with the business.
How deterioration gives you time to act
Many mechanical failures develop gradually. Reliability engineers describe this with the P-F curve: the point at which a developing fault first becomes detectable (P, potential failure) comes some time before the point at which the asset functionally fails (F). The length of that interval depends on the failure mode and on what you measure. Vibration analysis on a bearing may give weeks of warning, while temperature might rise only shortly before failure.
Standards such as ISO 17359 (general guidance for condition monitoring programmes) and ISO 13374 (processing and presenting condition monitoring data) are useful references when designing the programme and the data pipeline.
This is why sensor choice matters more than algorithms. If you aren't measuring the signal that changes early for the failure you care about, no model can predict it.
| Asset type | Common failure modes | Signals that often give early warning |
|---|---|---|
| Motors and pumps | Bearing wear, misalignment, imbalance, cavitation | Vibration spectrum, temperature, motor current |
| Compressors | Valve and bearing wear, leaks | Vibration, discharge temperature and pressure, run cycles |
| Gearboxes | Gear tooth wear, lubrication failure | Vibration, oil temperature, oil analysis |
| Electrical panels | Loose connections, overloads | Thermal readings, current |
| CNC machines and presses | Spindle wear, tool wear, hydraulic issues | Controller alarms, spindle load, cycle times, hydraulic pressure |
From rules to remaining useful life: what each approach needs
| Approach | Question it answers | Data it needs | When to use |
|---|---|---|---|
| Rules and thresholds | Is this reading outside safe limits? | Limits from manufacturers, standards and engineers | Day one, for known limits |
| Anomaly detection | Is this machine behaving differently from its normal? | Weeks or months of normal operating data | Early, before you have failure history |
| Failure classification | Which fault is developing? | Labelled examples of each failure mode | Once technicians record confirmed causes consistently |
| Remaining useful life (RUL) | How long until it needs attention? | Run-to-failure or degradation histories for similar assets (public benchmarks such as NASA's simulated turbofan dataset show the kind of data needed) | Mature programmes with good history |
Platform architecture
Most predictive maintenance platforms follow the same layered pattern. Keeping the layers separate means you can add new machine types or swap models without rebuilding everything.
- Data acquisition. Wireless vibration and temperature sensors on older machines; controller and PLC data through OPC UA or gateway software on newer ones; existing SCADA or historian data where it's already being collected.
- Edge processing. Gateways that buffer data during network outages and compute features locally, such as vibration RMS and frequency bands, so you don't send raw high-frequency waveforms to the cloud continuously.
- Ingestion and storage. A message broker (often MQTT) feeding a time-series database for measurements, with asset metadata (hierarchy, make, model, criticality) in a relational database.
- Analytics. Feature calculation, rules, anomaly models and, later, failure and RUL models, each versioned so you know which model produced which alert.
- Alerting and workflow. Alerts grouped and prioritised by asset criticality, linked to work orders in your maintenance system or ERP, with a feedback step where the technician records what they found.
- Applications. Asset health dashboards for planners, mobile apps for technicians, and reports for plant managers.
The feedback loop is the product
The single biggest difference between programmes that improve and those that stall is technician feedback. Every alert should end with a recorded outcome: confirmed fault (and which one), no fault found, or not inspected. That record is what tells you which alerts to trust, which thresholds to adjust, and eventually provides the labelled data needed for failure prediction.
Make this quick on a phone, with structured choices for asset, component and failure mode rather than free text. If technicians have to type a paragraph, you won't get the data.
Planning a predictive maintenance pilot?
ZSpace Labs builds machine-data analytics and AI models, asset health dashboards and technician apps connected to your maintenance system.
Integration with maintenance and production systems
Prediction only saves money if it changes what the maintenance team does. Connect the platform to:
- Your maintenance system (CMMS) or ERP maintenance module, to create and close work orders and read maintenance history
- Your MES or production schedule, to plan work into scheduled downtime and to know whether a machine was running when readings changed; see MES development
- Spare parts inventory, so a predicted bearing replacement checks stock before it is scheduled
- Asset registers, so every sensor is linked to the right asset and component
A phased rollout
Treat this as an illustration of the order of work, not a fixed plan.
| Phase | Scope | Exit criteria |
|---|---|---|
| 1. Asset selection | Rank assets by criticality, failure history and detectability; pick 10–30 for a pilot | Agreed pilot list with failure modes for each |
| 2. Instrument and connect | Sensors or controller data, gateway, storage, asset hierarchy | Reliable data from every pilot asset for several weeks |
| 3. Monitor | Dashboards and engineering thresholds | Alerts reviewed daily; false alarm rate tracked |
| 4. Detect anomalies | Per-asset baselines and anomaly scores | Alerts that technicians agree are worth inspecting |
| 5. Close the loop | Work orders and outcome recording | Outcomes recorded for most alerts |
| 6. Predict | Failure classification or RUL for assets with enough history | Measured lead time and precision on confirmed faults |
Measuring whether it works
Agree how you will measure success before the pilot starts, using your own baseline rather than industry averages. Useful measures include unplanned downtime on pilot assets, the share of alerts confirmed as real faults, warning lead time before confirmed faults, maintenance hours spent on inspections that found nothing, and spare parts used. Compare against the same assets' history or a similar group of assets without monitoring.
Build, buy or extend
| Option | Good fit when | Watch out for |
|---|---|---|
| Sensor vendor's platform | Common rotating equipment; you want results fast | Data locked in one vendor's cloud; limited integration with your systems |
| Industrial analytics product | Several plants and asset types; in-house reliability engineers | Licence cost; model customisation limits |
| Custom platform | Mixed machines and controllers, tight ERP/CMMS integration, or predictive maintenance as part of your own product | You own data pipelines, models and support |
| Hybrid | Vendor sensors and analytics, with a custom layer for workflow, integration and reporting | Two systems to maintain; check the vendor's data export and API terms |
Pro tip
Whatever you choose, make sure you can export your raw and processed data. Managed services do get retired: Google Cloud IoT Core closed in August 2023, AWS IoT Analytics reached end of support in December 2025, and Amazon Lookout for Equipment, an AWS service built for exactly this use case, was discontinued in October 2026. Teams that kept their data portable moved far more easily.
Common mistakes
- Starting with machine learning before the data is reliable
- Instrumenting every machine instead of a pilot set of critical assets
- Alerts that don't create or link to maintenance work
- No technician feedback, so models never improve
- Too many false alarms in the first weeks, so people stop looking
- Ignoring operating context: a vibration change during a product changeover isn't a fault
Conclusion
Predictive maintenance is a programme, not a model. Choose critical assets, measure the signals that change early, start with monitoring and anomaly detection, connect alerts to real maintenance work, and record what technicians find. Prediction becomes possible once that loop is running. For collecting and displaying device data more broadly, see IoT monitoring platform development.
Want to reduce unplanned downtime?
Tell us about your critical assets, controllers and maintenance system. ZSpace Labs will suggest a realistic pilot. Discuss your predictive maintenance project.
Common questions.
Software that collects condition data from equipment (vibration, temperature, current, pressure, run hours and controller events), detects signs of deterioration, estimates how urgent each issue is, and turns that into maintenance work before the asset fails. It connects sensors and machine controllers on one side to maintenance planning and work orders on the other.
No. A threshold alert fires when a reading crosses a fixed limit, such as a bearing temperature above 80°C. That is condition monitoring, and it is often the right first step. Predictive maintenance uses patterns over time, such as rising vibration in a specific frequency band, to detect deterioration earlier and estimate how long remains before failure.
For anomaly detection, a few weeks to months of normal operating data per asset type is often enough to start. To predict failures or remaining useful life, you also need recorded failures and maintenance history, which many plants don't have in usable form. Most projects start with monitoring and anomaly detection while they build up labelled failure data.
Critical assets where an unplanned stop is expensive, failures develop gradually enough to detect, and you can install sensors or read controller data. Rotating equipment such as motors, pumps, fans, compressors and gearboxes is a common starting point because vibration and temperature give early warning.
Buy, or use a sensor vendor's platform, if you have common asset types and want results quickly. Build, or extend a platform, when you need to combine data from many machine types and controllers, integrate tightly with your ERP or maintenance system, apply your own engineering knowledge, or offer predictive maintenance as part of your own product or service.