Anomaly Detection in Twins
A twin flags anomalies as persistent, unexplained gaps between what it predicts and what the sensors report.
The residual as an alarm
A twin continuously predicts what its sensors should read and compares that to what they do read. The gap, called the residual or innovation, should hover near zero within the known noise if all is well. A residual that grows or shows structure means something is happening the model does not expect: a developing fault, a sensor failure, or physics the twin is missing. Watching residuals is the most natural anomaly detector a twin has.
Why a twin detects better than a threshold
A simple alarm trips when a signal crosses a fixed limit, which misses subtle problems and cries wolf during normal transients. A twin knows what each signal should be given the current state and operation, so it can flag a signal that is abnormal for the circumstances even while still within its absolute limits. It detects the unexpected, not merely the extreme.
From detection to diagnosis
- Detect: a residual departs from its expected statistics
- Isolate: which sensors and states are involved narrows the location
- Diagnose: the physics model links the pattern to a candidate cause
- Assess: the twin estimates severity and how much time remains
Ambiguity and multiple hypotheses
One anomaly can have several explanations, a real fault or a failing sensor that mimics one. A twin can carry competing hypotheses with their probabilities until more data resolves them, rather than committing prematurely. Methods that represent multiple modes, such as particle filters, suit this directly. See particle filters.
In the Kronos twins
For the Hyperion breeder and the burner, residual-based detection watches for developing magnet, thermal, and structural problems, and for plasma behavior drifting from expectation. Because both machines are pre-construction, detectors are tuned today against simulated faults injected into the design model, an approach that establishes their behavior before real faults can occur. Cross-diagnostic consistency, from sensor fusion, adds a second detection layer. See diagnostic fusion and disruption prediction.