4. Case Study

Three studies in the same loop: diagnose, recover, then learn

Study A — diagnosis (RQ3)

31 HSS scenarios · 160 s · inject at 20 s

Four evaluation methods; everything else fixed

Decision = confidence gate → recovery signal (open loop)

Study B — recovery ordering

Same 31 · 3600 s · inject at 20 s

Diagnosis fixed: KG + cardinality · Decision = P/T priority

Execute = procedures with ETRs; verify on alarms

Study C — learned LSTM detection

AnomalyLSTM: 60 s × 68 sensors → nominal / anomaly

Offline: Part B5 scenario holdout · binary train

Online open loop: binary scores + threshold compare; anomaly obs can gate diagnosis

A asks which evaluator to trust. B asks which recovery to try next. C asks whether a learned detector beats thresholds and can arm the diagnosis stages.