Sensor Fusion
Sensor fusion combines many imperfect, partial measurements into one estimate more accurate and complete than any single sensor.
The whole from partial views
No single sensor sees everything, and every sensor is noisy. Sensor fusion is the combination of readings from many sensors, of the same or different kinds, into a unified estimate that is more accurate, more complete, and more robust than any one of them. In a twin, fusion is performed inside the assimilation loop, weighting each sensor by its trustworthiness.
Complementary, competitive, cooperative
- Complementary: sensors cover different parts of the system, together giving a fuller picture
- Competitive: sensors measure the same quantity, so agreement builds confidence and disagreement flags faults
- Cooperative: sensors that individually reveal little combine to infer something neither sees alone
How the weighting works
Each measurement enters with a noise model describing its expected error. The estimator trusts precise, well-characterized sensors more and noisy ones less, and it accounts for correlations, so two sensors sharing a common error are not double-counted. The result is a state estimate whose uncertainty reflects the full sensor set.
Handling disagreement
When redundant sensors disagree beyond their noise, fusion must decide which to believe. Techniques range from voting among redundant channels to statistical outlier rejection to explicitly modeling the possibility that a sensor has failed. This gatekeeping is what keeps one bad channel from corrupting the whole estimate.
Fusion in the Kronos twins
A fusion machine carries diagnostics of many types on very different timescales: magnetic, thermal, optical, neutron, and structural. Fusing them lets the twin reconstruct states no single diagnostic reveals, for example combining magnetic and pressure measurements to infer the Hyperion plasma equilibrium. It also provides graceful degradation: if a diagnostic is lost, the estimate widens rather than fails. See diagnostic fusion in fusion machines and data assimilation.