Sensor Fusion in Control
Combining multiple imperfect sensors yields an estimate more accurate and robust than any single one, and lets control proceed when a sensor fails.
Many Sensors, One Estimate
Sensor fusion combines readings from several sensors, often of different types, into a single estimate of the quantity of interest. No real sensor is perfect: each has noise, bias, limited range, and failure modes. Fusion exploits their independence so that the combined estimate is more accurate, more robust, and available even when individual sensors disagree or drop out.
Why Fuse
- Noise reduction: independent errors partly cancel when combined
- Complementary strengths: one sensor covers what another misses
- Redundancy: the estimate survives the loss of any single sensor
- Observing the unmeasurable: infer states no single sensor reads directly
The Estimation Framework
The workhorse of sensor fusion is the recursive estimator, most famously the Kalman filter. It maintains a model of the system's state and its uncertainty, predicts the state forward using a dynamic model, and then corrects the prediction with each new measurement, weighting model and measurement by their relative confidence. The result is an optimal estimate for linear systems with well-characterized noise, and variants extend it to nonlinear cases. This connects fusion to state-space control, where the estimated state feeds the control law.
Weighting and Validation
Good fusion weights each sensor by how much it should be trusted, and that trust is not static. A sensor drifting out of range, changing too fast to be physical, or disagreeing with the others should have its influence reduced or be rejected. Validating each input before fusing it prevents one faulty sensor from corrupting the combined estimate, which is essential when the estimate drives control or safety decisions.
Application to Plasma Diagnostics
Controlling a fusion plasma depends on reconstructing quantities, position, shape, and stored energy, that no single diagnostic measures directly. Many diagnostics, magnetic sensors, interferometry, and others, are fused through a model-based estimator to produce the state the controller acts on. In a design like the Hyperion breeder, at the design and simulation stage, this fused estimate is what the multivariable controller uses, and its robustness to a failed diagnostic is part of making the control loop dependable.