State Estimation for Plasma Control
The general problem of inferring the full plasma state from limited diagnostics, and the observers that solve it in real time.
Why estimation is unavoidable
The quantities controllers most want - internal profiles, exact position, stability margins - are not measured directly. Diagnostics see projections and integrals of the state, corrupted by noise. State estimation reconstructs the underlying state from these partial views, using a model to fill what the sensors cannot see. Every control loop rests on some form of estimation.
Observers
An observer is an algorithm that runs a model of the plant in parallel with the real system and corrects its internal state using the measurement error. If the system is observable, the observer's estimate converges to the true state. The Kalman filter is the optimal observer under Gaussian noise; simpler fixed-gain observers suffice where speed matters more than optimality.
Full versus reduced estimation
Full equilibrium reconstruction estimates the whole magnetic state each cycle and is the richest estimator, but it is expensive. Reduced observers estimate only the few quantities a given fast loop needs - vertical position for the vertical loop, current for the current loop - trading completeness for the speed that loop demands. A stack uses both, matched to each layer.
Handling missing and conflicting data
- Dropped sensors: the model carries the estimate until data return
- Conflicting sensors: the estimator weights by known reliability and flags disagreement
- Degraded diagnostics: uncertainty grows and downstream loops widen margins
- Calibration drift: cross-checks catch a sensor sliding out of trust
Estimation as the single source of truth
Centralizing estimation means every controller acts on the same state at the same timestamp, so loops never fight over contradictory pictures of the plasma. This consistency - one estimated state, shared by all - is as important to a stable stack as the accuracy of any single measurement.