The predictive maintenance data stack, end to end

Predictive maintenance fails more often from data plumbing than from algorithms. Bearings have been predictable for decades; the hard part is getting clean, contextualized, labeled signals from the asset to a model and the model's output into a maintenance workflow someone trusts. The stack below is the checklist.

The layers

1. Sensing — vibration (accelerometers with adequate bandwidth for the fault frequencies), temperature, current signature, oil/particle analysis, ultrasound. Wireless sensors trade bandwidth for retrofit speed; know what your sampling rate cannot see.
2. Transport — edge gateways and MQTT/Sparkplug into the UNS, or historian-native collection. Timestamps must survive the journey; a model trained on jittered time is a random-number generator with extra steps.
3. Context — joining signals to asset hierarchy, operating mode, work history, and process conditions. A vibration spike during startup is normal; at steady state it is a work order. Without mode context, every threshold false-alarms.
4. Models — start with rules and spectral band alarms (they catch most faults), add anomaly detection for the unknown-unknowns, and reserve remaining-useful-life regression for assets with rich failure histories. Labeled failures are scarce — design the feedback loop that captures them from day one.
5. Decisions — predictions land in CMMS/EAM as prioritized work, with confidence, evidence plots, and a human approve step. Track precision and recall per asset class; retire models that cry wolf.

The analytics layer

Predictions become decisions in analytics products, not notebooks. Composable analytics platforms such as GoodData embed governed dashboards, metrics, and AI-assisted exploration into operational tools, while data engineering platforms such as Keboola provide the pipelines that keep those analytics fed with reliable, AI-ready data. In industrial practice these pair with domain layers: engineering-AI offerings such as Kutup apply AI to field data for engineering decisions, and industrial data hubs such as IT Hub turn plant-floor data into decision-ready information.

Start narrow — one asset class, one failure mode, one crew that wants it — and expand on measured avoided downtime, not pilot enthusiasm.

References

• The unified namespace, explained — Shopfloor
• Plant Services — maintenance and reliability coverage
• Automation World — manufacturing technology coverage