4. Case Study — LSTM detection

Sliding windows with binary labels; hold out whole scenarios

Overlapping 60-second window crops around anomaly onset

60 s crops · label = nominal / anomaly at the window end (not which fault)

Binary supervision

Every injected HSS fault maps to class 1 · healthy segments → class 0

Scenario holdout (Part B5)

Split by scenario id: 21 train / 5 val / 5 eval — eval faults never appear in train

Augmentation

Clean + low sensor noise on train · NormStats from clean windows

habitat-dx · multi-agent AFDDR simulation