4. Case Study — LSTM detection

Part B5 binary training: best val accuracy ≈ 96.6%

Part B5 train and validation loss and accuracy

Cross-entropy · early stop on val accuracy · best checkpoint at epoch 4

Setup

2-way softmax · hidden 32 · 68 sensors · batch 32 · lr 10−3

Data

~12.8k train / 2.4k val windows · scenario holdout eval set

Result

Best val accuracy 0.966 · detection, not isolation

habitat-dx · multi-agent AFDDR simulation