Name the fault. Name the cause. Price the consequence.
Detect turns residuals into a diagnosis. Not "something looks abnormal" — a named fault with a confidence, the root cause behind it, the indicators that made the call, and what it is costing you per hour while it runs.
MVPEvery figure above is counted in the minimum viable product: the amine train modelled for gas sweetening and post-combustion carbon capture.
What it does
Absorber and stripper foaming. Absorber and stripper flooding. Exchanger fouling and internal leak. Reboiler fouling, corrosion and tube leaks. Pump cavitation on all four pumps. Under-stripping. Filter plugging and bypass. Inlet conditioning failure. Each one is a signature with its own evidence, not a threshold on a tag.
Thermal degradation, heat-stable salts, hydrocarbon contamination, suspended solids, wrong antifoam, duty deficiency, internals damage, control-loop instability. Every attribution lists the indicators that witnessed it, the ones that corroborate, and the ones that would exonerate it.
Six consequence buckets: off-spec treated gas, amine losses, excess steam and energy, reliability and availability loss, and throughput given away to hold spec. Each active fault carries a severity and a nominal cost per hour, so the shift can rank what to fix first.
Transients and off-baseline load are gated out before anything is judged. A plant that is ramping is not a plant that is faulty, and a diagnostic engine that cannot tell the difference is noise.
How it works
Load envelope, baseline deviation and steadiness are checked first. Unsteady rows do not reach the inference stages.
Level, trend and oscillation flavours, drawn from raw tags and from the reference model residuals over a six-hour trailing window.
Each indicator gets a membership in absolute engineering-unit states. After expert review most indicators were moved off relative sigma anchoring — a plant that is slowly getting worse must not quietly redefine what "normal" means.
Primary evidence, corroborating evidence and exonerating context combine into a confidence and a tier: confirmed or suspected.
Witness indicators point at a cause. Latent propensities — degradation, salts — act as priors, so a slow background problem raises the odds of the fast fault it eventually produces.
Active faults are mapped through the causal graph onto the six cost buckets, with a nominal figure per hour.
Where it applies
The platform is focused on acid gas removal — gas sweetening and post-combustion carbon capture — because that is where the model is calibrated and where the fault library is real. The method underneath is not amine-specific: a physics reference model, residuals against expected-healthy, fuzzy diagnosis over graded indicators, and costed consequences. Extending it to another unit means new datasheets and a new calibration, not a new platform.
The platform runs on a real amine train with thirteen days of one-minute history behind it. Nothing on these pages is a mock-up.
Open the platform