Context
Sites that keep good maintenance records are better placed than most, but those records rarely get used for anything beyond reporting. Meanwhile the alarm point on a pump is usually set from a handbook figure or from experience, and it is hard to know whether it sits in the right place. Set it too tight and operators are sent to inspect a machine that turns out to be fine, until they stop responding at all. Set it too loose and the warning arrives with no time left to plan a repair. There is rarely any evidence for where the line should be drawn.
What we tested
The monitor learns the normal relationships between the signals the pump already produces. Each signal is estimated from the others, and the gap between estimate and measurement is condensed into a single anomaly indicator. It uses a class of lightweight machine-learning techniques (reservoir computing) suited to time-series signals and edge deployment. The alarm point is then calibrated against the recorded failures: we swept the threshold across its range and read off, at each setting, how much warning it would have given and how many alarms it would have raised during normal operation. Validated on a publicly available dataset of pump operation containing several recorded failures.
What it shows
Because the trade-off is measured rather than assumed, the alarm point can be placed deliberately: more warning where a repair takes time to arrange, fewer interruptions where it does not. The indicator also identifies which signal is drifting.
This may apply to your equipment
If you run pumps and keep reliable records of past failures, those records are worth more than they are usually given credit for. They let the alarm point be chosen from evidence rather than convention. A short conversation is usually enough to tell whether it fits.