Anomaly-Detection Ensembles
No single detector is reliable enough for safety-relevant precursors, so Kronos combines reconstruction, density, and forecasting detectors into a calibrated ensemble.
Why an ensemble
Sub-threshold precursors are rare, varied, and partly novel, exactly the regime where any single anomaly model has blind spots. Kronos runs an ensemble of complementary detectors so that a precursor invisible to one is likely visible to another, and combines their scores with a calibrated aggregator.
- Reconstruction detectors: autoencoders flag signals the model cannot reproduce
- Density detectors: estimate how improbable the current feature vector is
- Forecast-residual detectors: flag when observed evolution departs from the twin's prediction
- One-class detectors: learn the boundary of normal operation, score distance outside it
Combining scores
Raw detector scores are on different scales, so each is calibrated to a probability, then combined. Kronos weights detectors by their validated reliability on the relevant failure class and requires agreement for high-confidence alarms, which suppresses the single-detector false positives that would erode operator trust.
# calibrated ensemble score
p_k = calibrate_k(score_k(x)) # per-detector -> probability
w_k = validated_reliability[k] # from V&V on that failure class
p_ens = sum_k w_k * p_k / sum_k w_k
alarm = (p_ens > tau) and (n_agreeing >= m) # require agreement
The ensemble does not only watch magnets. It watches disruption precursors for the breeder (locked-mode phase patterns on the Mirnov graph), plug-density excursions for the burner, divertor thermal anomalies, and fuel-cycle isotope-balance drift. Each failure class has its own tuned detector mix, but all report into the same calibrated framework so operators see one coherent risk picture.
Relation to the twin
The forecast-residual detector is the tightest coupling to KRONOS-CTRL: it compares the live signals against the twin's 50-100 ms shadow prediction, so an anomaly is precisely a place where reality diverges from validated physics. That is often the earliest and most physically meaningful precursor of all.