ACHRAF EL HANI · INDUSTRIAL DATA LAB
Could we have seen the pump failures coming?
A water pump monitored by 52 sensors failed 7 times in 5 months. Pick a failure and a sensor to see what the data looked like beforehand, and whether a simple detector would have raised an alarm in time.
Five months of anomaly score
Number of sensors outside their normal band, hour by hour. The detector learns on April and May only, so June to August is a fair test. Click a failure to inspect it.
Sensors out of band
Alarm threshold
Training period (April–May)
Failure
Failure
Hours shown
Sensors furthest from normal in the 6 h before this failure
Sensor reading
Normal band (learned on April–May)
Out of band
Failure
Anomaly score: sensors outside their normal band
Sensors out of band
Alarm threshold
Confirmed alarm (30 min above threshold)
Failure
All seven failures
Warning time: hours in a row before the failure during which the alarm was on most of the time.
How it works, and what it does not prove
- For each sensor, the normal band is its median ± 3 robust sigmas (MAD), learned on normal operation in April and May.
- The anomaly score counts sensors outside their band. The threshold is the level exceeded 1 % of the time on normal training data.
- Seven failures in total, three in the test period, one detected. This describes one pump over five months. It is not a detection rate.
- Sensors are anonymous, so no finding can be tied to a physical cause.