Module sensor_graph

Expand description

Spatio-temporal GNN for distributed sonobuoy / sensor arrays (KG-013 / WEFT-354).

K-STEMIT-style dual-branch architecture adapted from ice-penetrating radar (arXiv:2604.09922) to underwater acoustic sensing:

Spatial: GraphSAGE on a buoy-position graph with inverse-distance (and optional propagation-delay) edge weights — learned beamforming over irregular arrays.
Temporal: Gated linear unit (GLU) convolution over per-buoy acoustic time series — learned matched filtering for transients.
Physics prior: ocean acoustic node features (SSP, thermocline, sea state, current, bottom type) projected into the same hidden space.
Adaptive fusion: learnable α ∈ [0,1] balances spatial vs temporal; physics residual is added with a fixed scale (or learnable γ).
Classification head: linear → logits for detection / species-style labels.

Pure Rust, no GPU / no ndarray. Behind the sensor feature flag.

Structs§

EvalMetrics: Classification metrics.
ForwardOutput: Outputs of a single-node forward pass.
GraphSageLayer: One GraphSAGE hop: h = σ(W_self x + W_neigh AGG(N(x))).
KstemitConfig: Hyperparameters for KstemitModel.
KstemitModel: K-STEMIT-style spatio-temporal GNN for sonobuoy arrays.
Linear: Dense affine map y = Wx + b with analytical gradients for SGD.
SensorEdge: An edge connecting two buoys.
SensorGraph: Graph-based sensor fusion substrate for distributed buoy arrays.
SensorNode: A buoy node in the sensor graph.
SyntheticConfig: Parameters for generate_synthetic_array.
SyntheticDataset: Full synthetic dataset: shared array graph + per-node labels.
SyntheticSample: One labelled observation: which node, and ground-truth class.
TemporalGlu: Gated linear unit: y = a ⊙ σ(g) where [a;g] = W x + b.
TrainConfig: SGD hyper-parameters.
TrainResult: Result of a training run.

class_name: Human-readable class name.
evaluate: Evaluate mean CE and accuracy on all samples.
generate_synthetic_array: Generate a synthetic sonobuoy array with class-conditional temporal events.
train_sgd: Train a new model on a synthetic dataset.