Finding (v1.114 full-loop rotor-out campaign)
The new run_fullloop_motor_out_campaign (200 dispersed full-production-loop trials) surfaced that the rotor-out FDI detection latency blows up under heavy sensor noise. At gps σ up to 0.3 m and gyro-white up to 0.02 rad/s, isolation latency reached ~1882 steps (7.5 s) and a few trials mis-isolated (wrong rotor) or never isolated. The v1.114 release scoped its guard to a modest sensor-noise envelope (gps σ ≤ 0.12, gyro ≤ 0.006) where detection is reliable (worst 5 steps); the heavier-noise regime is uncovered.
Why it matters
The rotor-out recovery allocation is solid (bounded tilt, no flip) — but the whole chain depends on the FDI first isolating the dead rotor. The detection gate is tilt_cos > 0.90 && rate2 < 1.0; noise pushes rate2 over the gate, delaying/blocking isolation.
Scope
Make the CUSUM FDI robust to realistic GNSS/gyro noise: filter the effectiveness residual, adapt the gate, or add hysteresis — measured against the dispersed full-loop campaign at the heavier envelope. Not a v1.114 regression (pre-existing); tracked for v1.115.
Rivet: SWREQ-FALCON-FAULT-P03.
Finding (v1.114 full-loop rotor-out campaign)
The new
run_fullloop_motor_out_campaign(200 dispersed full-production-loop trials) surfaced that the rotor-out FDI detection latency blows up under heavy sensor noise. At gps σ up to 0.3 m and gyro-white up to 0.02 rad/s, isolation latency reached ~1882 steps (7.5 s) and a few trials mis-isolated (wrong rotor) or never isolated. The v1.114 release scoped its guard to a modest sensor-noise envelope (gps σ ≤ 0.12, gyro ≤ 0.006) where detection is reliable (worst 5 steps); the heavier-noise regime is uncovered.Why it matters
The rotor-out recovery allocation is solid (bounded tilt, no flip) — but the whole chain depends on the FDI first isolating the dead rotor. The detection gate is
tilt_cos > 0.90 && rate2 < 1.0; noise pushesrate2over the gate, delaying/blocking isolation.Scope
Make the CUSUM FDI robust to realistic GNSS/gyro noise: filter the effectiveness residual, adapt the gate, or add hysteresis — measured against the dispersed full-loop campaign at the heavier envelope. Not a v1.114 regression (pre-existing); tracked for v1.115.
Rivet:
SWREQ-FALCON-FAULT-P03.