When is “anomaly” declared? — limits vs learned binary score
Onset — p(anomaly) after inject
Speed — median / mean
Noise — pooled F1 (4×3×10)
Compare
Threshold limits vs LSTM-binary (60 s × 68) · open loop · inject 80 s · ready ≥ 60 s
med 7 s → 3 s
30/31 → 31/31
TPR 0.88 → 0.96
pre FPR 0
noise F1 0.92–0.99
High AR1/mixed: precision drops via FA — detection rate stays 31/31
Discuss
Faster onset vs noise false alarms — who gates diagnosis?
Limits are interpretable; LSTM is faster and covers the soft miss.
Noise mainly hurts precision (FA), not detection rate.