Anomaly Detection in Distributed Climate Sensor Networks

Anomaly Detection in Distributed Climate Sensor Networks

The Pacific Climate Observation Network (PCON) operates 2,100 ocean buoys, atmospheric sensors, and subsurface probes across a 4,000km² monitoring area. Each sensor generates ~1.4MB of data per hour. Centralized anomaly detection was failing.

The problem is multivariate, non-stationary, and geospatially correlated. A temperature spike in sensor A7 is an anomaly on its own — unless sensors A6 and A8 show the same spike, in which case it's a weather event. Distinguishing the two requires local correlation across the mesh.

The Edge Approach

We deployed Edge Lattice across the PCON network. Each node runs a local autoencoder trained on the sensor's historical distribution, plus a mesh consensus layer that queries neighboring nodes before flagging an anomaly.

The consensus rule: if fewer than 20% of nodes within a 50km radius agree that a reading is anomalous, flag it as instrument noise. If ≥20% agree, escalate to the cloud for analysis.

Results (12-month deployment)