Failure Events

Every logged failure improves the system's predictions. Rela AI automatically captures sensor state at the moment of the failure and uses it to recalibrate the model.

What is it for?

• Turn every intervention into data the predictive model uses to improve.
• Separate predictions that came true from false alarms.
• Feed the "predict → intervene → confirm → retrain" loop that makes the system evolve.

How it works

1. Detection — a technician or the system detects a failure.
2. Record — it gets logged with type, description and damaged parts.
3. Automatic capture — Rela saves the last 24 h of sensor data.
4. Retraining — with 3+ failures of the same type, the model recalibrates automatically.
5. Improvement — the next prediction is more accurate.

Log a failure

POST /api/v1/failure-events/{asset_id}
{
  "failure_type": "mechanical",
  "description": "Front bearing worn",
  "root_cause": "0.3 mm misalignment",
  "parts_damaged": ["bearing_front", "seal_shaft"]
}

Failure types

• mechanical — Mechanical (bearings, gears, shafts)
• electrical — Electrical (windings, connections, insulation)
• hydraulic — Hydraulic (seals, pumps, valves)
• pneumatic — Pneumatic
• process — Process (SPC out of control)
• instrumentation — Instrumentation
• software — Software / control
• other — Other

Closing a work order

When a technician closes a work order they can confirm whether the failure was real:

POST /api/v1/tasks/{task_id}/close
{
  "findings": "Worn bearing, 0.3 mm misalignment",
  "actions_taken": "Replaced bearing, corrected alignment",
  "parts_replaced": ["bearing_front"],
  "failure_confirmed": true,
  "failure_type": "mechanical"
}

If failure_confirmed: true, the system automatically:

• Logs the failure to feed the model back.
• Updates the equipment's last-maintenance date.
• Checks whether to retrain the model (3+ failures).

Key benefits

• Every intervention feeds the tenant's predictive model — accuracy that grows with use.
• Canonical catalog of 8 failure types aligned with the rest of the predictive pipeline.
• Automatic capture of 24 h of telemetry around the event, with no manual work from the technician.
• Automatic retraining after 3+ failures of the same type — the system learns on its own.
• Closed feedback loop: prediction → intervention → confirmation → better future predictions.