Score, training pumps
94.0
In one sentence
RetroSense tells the operator early what is wrong, what to do, and when not to trust a reading. It only reads data; people act, and the plant’s own alarms stay in charge.
Results · organizer’s own scorer
94.0
85.4
0.06
Every setting was chosen on training pumps only. Official organizer scorer. Settings chosen on training pumps only; validation scored once. Validation has 2 pumps and 9 events, so a few points either way is noise.
What a fixed alarm can’t do
01 · Pump or sensor
A reading moves but the signals that should move with it don’t: it’s the sensor. Check the transmitter, don’t stop the pump.
02 · Cause, not symptoms
A clogging strainer leads to cavitation. RetroSense links them, so the crew cleans the strainer instead of chasing alarms.
03 · Honest
With no healthy history, it compares the pump with the maker’s data sheet. Where sensors are missing, it says “can’t see”, never “healthy”.
The five clues · training pumps
Before = our first complete version (score 86.8). Now = the final version (score 94.0). The number is how many of those faults were named right.
6 → 7 of 7
Fixed: one clogging strainer was first called a drifting gauge, because only suction moved clearly. Now RetroSense checks the pressure in kPa: when discharge falls with suction, it is the pump, not the gauge. It also flags cavitation after 5 minutes below the safe margin, so short episodes are caught.
3 → 3 of 3
Already right: the bearing’s own defect frequency shows up in the vibration spectrum long before the overall vibration alarm. Nothing to fix; it stayed perfect while we changed the rest.
2 → 3 of 4
Added: a pump worn from the first minute looks normal against its own history. We now also compare it with the maker’s pump curve, so no healthy first day is needed. Still open: a closed cooling loop has no suction reference, so that check can’t run; the screen says “can’t see” instead of guessing.
8 → 8 of 8
Already right: a reading that moves while the signals that should follow it don’t is the sensor. It stayed perfect while every other check was added.
6 → 7 of 7
Added: cards now carry their cause. Cavitation soon after clogging points back to the clogging, and a fault right after a repair points back to the repair (bearing replacement → misalignment).
No path to the PLC, drive, valves or safety system.
A person approves every action; every alert prints as a work card that starts with the local gauge.
Designed with IEC 62443 principles in mind; to be assessed per site.
Faults injected into PUMP-01’s discharge pressure, 30 minutes.
No alarm
10 min
to flag it
5 min
to flag it
Hui’s controlled synthetic tests; not a measure of real-world accuracy.
What it could be worth
Illustrative0.9 yr
payback, one duty pump
€4,905
net per year, one duty pump
never
payback, standby pump (€−2,037 / year)
Organizer’s own business-case model, run with our measured numbers. Synthetic data and assumed costs: start with duty pumps.
What we don’t claim
Nothing has run on a real plant yet. Next step: a few weeks in shadow mode.
Official organizer scorer. Settings chosen on training pumps only; validation scored once. Validation has 2 pumps and 9 events, so a few points either way is noise.
Read-only installation and IEC 62443 alignment are proposed, to be assessed per site.