4. Case Study — LSTM detection

Binary AnomalyLSTM: window → recurrent net → nominal / anomaly

Habitat sim 1 Hz telemetry Input window T=60 · F=68 · z-scored t−59 … t−1 t x ∈ ℝ⁶⁸ each second AnomalyLSTM (unrolled) same cell · shared weights · H = 32 xₜ₋₅₉ … xₜ₋₁ xₜ LSTM cell ··· LSTM cell LSTM cell h h h hₜ ∈ ℝ³² each step τ: (h_τ, c_τ) = LSTM(x_τ, h_{τ−1}, c_{τ−1}) only final h_t → binary head · gates: next slide Binary head two logits → softmax · true detection ⋮ hₜ (32) z₀ nominal z₁ anomaly P = softmax(z) · 2 classes if P(anomaly) ≥ 0.5 → observation In the habitat-dx loop Detection lstm-binary Hypothesis gated / off in open loop Evaluation P(nom), P(anom) Decision not in Study C open Execute disabled Detection publishes binary scores each tick anomaly observation gates diagnosis when hyp / eval are enabled solid = active in open-loop Study C dashed = disabled / gated
habitat-dx · multi-agent AFDDR simulation