Predictive maintenance for buildings: from sensor to alert with AI

Corrective maintenance (repairing after failure) costs three to five times more than preventive maintenance (scheduled servicing). Predictive maintenance goes one step further: anticipate failures before they happen, by reading the weak signals sensors emit continuously.

Historically reserved for heavy industry (turbines, production lines), it becomes accessible to buildings thanks to three converging forces: affordable IoT sensors, AI embedded in modern BMS platforms, and protocol standardisation. Sensia integrates four native predictive algorithms on every supervised equipment.

RUL (Remaining Useful Life) estimates time-to-failure by linear regression on 90 days of health score. Typical precision ±15% at 30 days.

MTBF (Mean Time Between Failures) computes the mean time between failures from the alarm history (triggered → resolved).

Unsupervised anomaly detection flags behaviours deviating from baseline without manual labelling — a drifting temperature sensor, an overnight electrical consumption spike.

Composite health score aggregates age, work orders, alarms, anomalies and utilisation into a 0-100 score per equipment, readable at a glance in the equipment panel.

Two pipelines coexist. The polled pipeline runs daily and goes through organisations one after another, one equipment at a time — each saved before the next. For each equipment: component computation (age, work orders 90d, alerts 30d, anomalies, utilisation), 0-100 score, history snapshot, RUL estimation, repair time measured on its own work orders, persist on entity_record.attributes, AI-enriched recommendation if at risk.

The event-driven pipeline triggers on the twin.{thing_id}.changed firehose (transitions, sync bursts, alerts), debounced 60s per equipment, sub-minute latency after change. The polled stays as a safety net.

Recommendations show up in the Performance tab of Operations, below the maintenance indicators: for each at-risk equipment, the retained option (« replace », « repair » or « monitor »), its cost and the downtime it implies, an AI-written explanation, and one-click work order creation. Costs are those of the equipment's type, and the screen says where they come from. The equipment panel shows its estimated remaining life and its health score.

No separate tab to learn: predictive maintenance integrates into existing pages.

The rul_days variable is automatically exposed to the rule engine. A typical « End of life approaching » rule configures in two clicks: rul_days < 7 fires a critical alert, rul_days >= 14 clears it (anti-flapping hysteresis). Same for health_score, mtbf_days, availability_pct.

You combine these variables with fleet conditions (only the AC units in the Marseille building, for instance) to target your alerts.

Full predictive maintenance (RUL, LLM recommendations) is included in Sensia Hypervision, priced per m² per year and published openly: €4,500/yr flat below 3,000 m², then €1.50/m²/yr, and €1.10/m²/yr above 25,000 m².

Sensia Compliance is limited to base reliability metrics (MTBF, MTTR, availability), bounded to the regulatory scope, with no predictive RUL and no AI layer.

Models trained on your historical data and a dedicated support SLA fall within the bespoke Enterprise tier (€2.50–4/m²).