These are common failures that take RNG plants off spec. Three of them are one chain: when the amine lets CO₂ through or the mole sieve lets water through, it freezes in the NRU. The numbers show where each failure starts. The dashed arrows show where it ends.
On the left is the channel the control room watches, with its H alarm. On the right is the residual: the same plant, with what normal operation explains taken out. The amber line is the moment the panel reacts. The green line is our warning. All numbers in the text come from the same run that draws the charts.
Many plants already have a steady-state process model, calibrated on real data. This demo does not run one. Where the twin would give the expected value, we use a simple physics curve fitted on the healthy stretch of each run. The twin would replace that curve. The detector stays the same.
Real inputs go into the model, and the model gives the expected value: CO₂ at the amine outlet, reboiler duty, cold box ΔP. The difference with the measured value is a feature for the detector.
The model says which combinations of values can happen. A reading outside that envelope is an instrument problem or a real upset. This matches the physics layer we already run.
Lower the amine strength in the model and see which signal moves first. Real failures are rare, so this is how we get enough cases to set a threshold.
A steady-state model does not handle starts and trips well. It is slow and needs a licence, so we run it over many operating points and fit a fast surrogate. Its calibration ages, for example with the O₂ catalyst.
| Failure | What the process model would give | Stand-in in this demo |
|---|
| Event | Sales valve closed (assumed) | Uptime points, one plant, one year | Where it comes from |
|---|
—
| Archiving | We warn | Lead time | Records/day |
|---|