Case: Early Detection on a Centrifugal Pump
How Rela AI detected degradation on a process pump and avoided a 12-hour shutdown
Case: Early Detection on a Centrifugal Pump
Context
- Industry: processed food
- Equipment: centrifugal pump, production line 3
- Sensors: vibration, bearing temperature, motor current
- Criticality: high — the whole line stops if this pump stops
The previous problem
The pump failed every 2-3 months without warning. Each failure meant:
- 8-12 hours of unplanned downtime.
- Direct cost: ~$15,000 USD per event (lost production + emergency repair).
- Technician team stress from urgent repairs.
What Rela AI did
Week 1-2: learning (Level 0)
Rela connected to the 3 sensors via MQTT gateway. For 2 weeks it collected data and established the baseline of normal behaviour.
Day 15: first signal
The anomaly model detected a 15% increase in front-bearing vibration. AHI dropped from 85 (B) to 72 (B).
Day 17: degradation confirmed
Vibration kept rising. AHI fell to 68 (C). Estimated RUL: 120 hours (~5 days). Confidence: 72%.
Day 17 — agent action
The WhatsApp agent automatically sent the maintenance manager:
PUMP-L3: degradation detected
- Vibration: +25% over baseline
- Health: 68/100 (C — acceptable, declining)
- Estimated time before failure: 5 days
- Recommendation: inspect within next 48 hours
Create a work order?
Manager replied: "Yes, assign to Carlos".
Day 17 — automatic coordination
- Work order MAI-089 auto-created.
- Assigned to Carlos (closest technician, pump specialist).
- Carlos received a WhatsApp with details, inspection checklist and LOTO procedure.
Day 18 — intervention
Carlos inspected the pump. Finding: front bearing with premature wear caused by mild misalignment.
Intervention: bearing replacement + alignment correction. Planned downtime: 2 hours.
Outcome logged
Carlos closed the work order and recorded:
- Finding: "Worn bearing, 0.3 mm misalignment".
- Action: "Bearing replaced, alignment corrected".
- Failure confirmed: Yes → model recalibrated.
Impact
| Metric | Without Rela | With Rela |
|---|---|---|
| Downtime type | Unplanned (emergency) | Planned (scheduled) |
| Duration | 8-12 hours | 2 hours |
| Downtime cost | ~$15,000 USD | ~$3,000 USD |
| Cost avoided | — | $12,000 USD |
| Detection → intervention | N/A (fails first) | 36 hours |
| Team stress | High (emergency) | Low (planned) |
Model learning
After this intervention:
- The model learned this pump's degradation pattern.
- The next prediction was more accurate (confidence rose to 85%).
- An earlier alert was configured to trigger on +10% vibration (instead of waiting until +25%).
This case illustrates the full flow: detection → prediction → coordination → intervention → feedback. Each step feeds the next, and the system gets more accurate with every cycle.