4. Case Study

Three light examples in the same loop: detect, diagnose, recover

Detection — threshold vs learned

AnomalyLSTM: 60 s × 68 → nominal / anomaly

Offline holdout · online open-loop scores

Inject at 80 s · ready after 60 s of context

Noise stress: 4 modes × 3 intensities × 10 seeds

Ask: when is anomaly declared?

Diagnosis — evaluation methods

31 HSS · four evaluators · confidence gate

Everything else fixed · open-loop recovery signal

KG · TS single · TS multi · Bayesian

Same scenario timelines for side-by-side read

Ask: correct / wrong / abstain?

Recovery — procedure orderings

Same 31 · 3600 s · KG fixed · P/T vs baselines

ETR procedures · verify on alarms

P/T · Highest P · Shortest T · Random

First-try vs recovered-after-wrongs mix

Ask: which order finishes under time?

habitat-dx stages spotlighted in the next three slides

Detectionlimits vs LSTM
→
Hypothesisshared catalog
→
Evaluationfour methods
→
Prognosisoptional
→
DecisionP/T orderings
→
Executionverify alarms

Maroon = spotlighted today · dashed = optional · dense plots in backup