The demo simulates a seawater reverse-osmosis (RO) train, introduces a fault, and shows whether AquaSentinel forecasts a future intervention-criterion crossing before the conventional comparator confirms it.
Controls
Fault type
- Scaling / Biofouling / Particulate — gradual fouling faults. These are where early warning shines.
- Integrity breach — a sudden membrane failure (a step). Both detectors fire at once, so the lead-time is ≈ 0 — that's the honest, expected result for an instantaneous fault.
- Sensor drift ⚠ — a trick case: the conductivity sensor drifts but nothing is actually fouling. A good detector should stay silent — try it.
Severity
Drag to make the fault milder or more aggressive. More severe faults usually cross the alarm sooner, so the early-warning lead can shrink.
↻ New sensor noise
Re-runs with a different random noise seed — same physics, different measurement noise. Shows the result isn't a fluke.
▷ Replay
Replays the reveal animation for the current scenario.
Advanced — plant & model parameters
Change feed salinity, temperature, recovery, clean permeability, fault onset day, and sensor-noise level. Tick Cleaning-in-place (CIP) to add periodic membrane cleaning — the detector re-baselines after each cleaning instead of being fooled by the permeability jump. "Reset to Gulf preset" restores defaults.
Reading the chart
- Dark line — true normalized permeability (1.0 = clean; lower = more fouled).
- Faint line — the same signal as actually measured (with noise).
- Red dashed line — the DuPont-aligned normalized-flow criterion (0.90); salt-passage and pressure-drop criteria are also evaluated.
- Gray marker — fault onset · green — forecast advisory · red — 24-hour-confirmed industry criterion.
Reading the fouling outlook
The outlook translates the selected synthetic run into an operator view. It shows the likely condition, expected criterion-crossing day, modeled intervention window, and a suggested review action. Green, yellow, and red use declared time-to-criterion bands. They are not an opaque AI score or a calibrated probability.
The numbers
- Modeled intervention window — how much time separates the advisory from the disclosed comparator's confirmed criterion crossing.
- Detection delay after onset — how long after the fault started the detector noticed.
- ROC-AUC — how cleanly the detector separates fault from no-fault (1.0 = perfect).
- Energy SEC — modeled specific energy (kWh/m³), clean → at the alarm.
- Diagnosis — the likely fault type inferred from which channels moved.
- Detectors compared — the same labeled run scored by the industry criterion, CUSUM, intervention forecast, and held-out logistic model.
Open data & cost
⤓ This scenario / Benchmark results download CSVs you can analyze yourself. The cost calculator lets you set your plant's capacity and energy price to see an illustrative, conservative estimate of what earlier action could be worth.