4. Case Study — LSTM detection

Z-score sensors with train-set statistics, then classify the window

Z-score normalized sensors and LSTM anomaly probability

Illustration · top: channels on a common z-score scale · bottom: rising P(anomaly) after inject

Fit on train only

Per-sensor mean / std from clean training windows; frozen for val, holdout, and live inference

Not per-window

Absolute offsets can be diagnostic — we do not re-center each 60 s crop on its own

Effect

Mixed units share one scale so the binary LSTM can weigh cross-channel dynamics

habitat-dx · multi-agent AFDDR simulation