Backup — LSTM classifier (Stage 5)

AnomalyLSTM in the loop: window → recurrent net → class scores

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 hₜ₋₅₉ ··· h hₜ₋₁ h hₜ ∈ ℝ³² each step τ: (h_τ, c_τ) = LSTM(x_τ, h_{τ−1}, c_{τ−1}) only final h_t → class head · gates: next slide Class head one linear score per class, then softmax ⋮ hₜ (32) z₀ nominal z₁ anom. 1 z₂ anom. 2 ⋮ z₃₁ anom. 31 each z_k uses its own row of W on the same h_t z = W h_t + b , W ∈ ℝ³²ˣ³² P = softmax(z) → class probabilities In the habitat-dx loop Detection AnomalyLSTM Hypothesis skipped Evaluation = LSTM posteriors Decision confidence gate Execute procedures Detection publishes class scores as evaluation each tick hold while leader is nominal · decide on an anomaly · recover_nominal solid = active in Study C dashed = disabled
habitat-dx · multi-agent AFDDR simulation