Anomaly Detection in Operations
Anomaly detection flags behavior that departs from the learned normal, catching problems that no one wrote an explicit rule for.
Beyond fixed thresholds
Simple alarms fire when a signal crosses a fixed limit, but many problems appear as unusual combinations of signals that are each individually within range. Anomaly detection learns what normal operation looks like across many signals at once and flags departures from it, catching novel or subtle issues.
How it learns normal
- Train on data from healthy operation to model the normal joint behavior.
- Score new data by how far it sits from that model.
- Raise a flag when the score exceeds a calibrated threshold.
- Present the contributing signals so an operator can interpret the flag.
Interpretability matters
An anomaly flag that cannot be explained is hard to act on. Good detection points to which signals drove the alarm, so operators and the digital twin can diagnose the cause rather than just noting that something is off.
Where it applies
Anomaly detection watches diagnostics for early signs of disruptions, watches component data for the drift that feeds predictive maintenance, and watches the data pipeline itself for sensor faults.
Model drift
Normal changes over time as components age, so the detector itself must be monitored and updated, or it will raise false alarms. Tracking the detector's own performance is part of running it responsibly.
For Kronos
On both the Hyperion breeder and the burner, anomaly detection is a safety-relevant layer that complements physics-based limits, and its flags are logged for safety review.