Fusing the Diagnostics Constellation

L3 turns 60+ heterogeneous diagnostic channels into one coherent physics state, weighting each by what it measures and how much it can be trusted.

Many instruments, one state

The diagnostics constellation spans very different physics: Thomson scattering (electron temperature and density), interferometry (line-integrated density), Mirnov coils and flux loops (magnetics), REBCO strain gauges (magnet mechanics), ECE (electron temperature), fast-ion trackers, and neutron flux. Each measures a different projection of the machine state, at a different rate, latency, and noise level. L3's job is to fuse them into one consistent twin state, not to display them side by side.

Fusion by measurement model

Fusion works because each diagnostic has a forward model H mapping the physics state to what that instrument reads, a line integral for interferometry, a local pitch-angle for motional Stark effect, a flux integral for a loop. State estimation inverts all of them at once: it finds the state whose predicted measurements best match every instrument, weighting each by its noise and current trust. A disagreement between instruments is resolved by the physics prior rather than by trusting one arbitrarily.

• Thomson/ECE: local temperature and density profiles
• Interferometry: line-integrated density constraint
• Mirnov/flux: equilibrium anchor and MHD mode structure
• REBCO strain: magnet mechanical state and structural health
• Neutron flux / fast-ion: source and confinement cross-checks

Cross-checks between instruments are a built-in health test: an interferometer chord and a Thomson profile that imply inconsistent densities flag a sensor problem or a genuine off-normal state, raising the reconstruction residual and lowering confidence. The GNN's topology model makes these cross-checks physically aware, comparing instruments that share a flux surface rather than arbitrary pairs.

The fused state, with per-quantity uncertainty and lineage back to each contributing channel, is what every L3 consumer sees, so control, anomaly detection, and operators all reason over one coherent, traceable picture of the machine rather than a wall of raw signals.