Plasma State Estimation

State estimation fuses many noisy diagnostics into a single best estimate of the plasma state for controllers to act on.

From measurements to state

Controllers act on the plasma state, but no diagnostic measures that state directly. Each measures something related, with its own noise, delay, and blind spots. State estimation combines these measurements with a model of how the plasma evolves to produce a coherent, filtered estimate of the quantities the controllers need.

Filtering

The classic tool is a recursive filter, such as a Kalman filter, that predicts the next state from the model and then corrects the prediction with new measurements, weighting each by its trustworthiness. Filtering smooths noise, fills gaps when a diagnostic drops out, and provides estimates of quantities that are only measured indirectly.

# One recursive estimator step (conceptual)
x_pred = A @ x_est + B @ u        # model prediction
P_pred = A @ P @ A.T + Q          # predicted uncertainty
K = P_pred @ H.T @ inv(H @ P_pred @ H.T + R)  # gain
x_est = x_pred + K @ (y_meas - H @ x_pred)    # correct with data
P = (I - K @ H) @ P_pred

Handling faults

A good estimator degrades gracefully when a sensor fails: it drops the bad input and relies more on the model and the remaining sensors. It can also flag when measurements and model disagree beyond expected noise, which is an early sign of an off-normal event. This makes estimation part of fault detection, not just control.

In the Kronos program

The Hyperion breeder's estimator fuses magnetics, interferometry, and other diagnostics to feed the vertical, shape, density, and profile loops. It is designed to tolerate the loss of individual sensors without losing the fast loops, supporting the graceful-degradation ladder. The estimator is tuned against the machine model in simulation ahead of first plasma.