Innovation Sequence Monitoring
The statistics of a filter's innovations reveal whether the twin is healthy, overconfident, or missing physics.
The filter watching itself
Every time a filter processes a measurement it computes an innovation, the difference between the observed value and what the filter predicted. Over time these innovations form a sequence whose statistics are a sensitive diagnostic of the filter's health. A well-tuned filter with a correct model produces innovations that are small, unbiased, and uncorrelated in time. Departures from this are informative.
What healthy innovations look like
- Zero mean: no persistent bias between model and reality
- Magnitude consistent with the predicted innovation covariance
- No time correlation: each innovation is unpredictable from the last
- No structure tied to operating conditions
Reading the failures
Innovations larger than predicted mean the filter is overconfident, its covariance too small, so it under-weights new data; inflation or retuning is needed. A nonzero mean signals a bias, a miscalibrated sensor or a systematic model error. Time correlation in the innovations means the model is missing dynamics the data contains, a cue to raise fidelity or update the model structure. The innovation sequence thus tells the operator not just that something is wrong but what kind of wrong.
A bridge to anomaly detection
Innovation monitoring and anomaly detection are the same idea seen from two sides: anomaly detection asks whether the residual signals a fault in the machine, while innovation monitoring asks whether it signals a fault in the twin. Distinguishing the two, a real machine event versus a model or sensor problem, is a core skill, and consistent innovations are what let the twin trust its own alarms. See anomaly detection.
In the Kronos twins
For the Hyperion breeder and the burner, innovation monitoring is a standing health check on every filter in the twin, run continuously so that miscalibration or missing physics is caught early. During the design stage it is exercised on synthetic data to confirm the filters are correctly tuned before real measurements exist. It underpins the honest reporting of uncertainty the program insists on. See uncertainty quantification and the Kalman filter.