Multi-Station Collaborative Analysis of Earthquake Precursors Considering Data Missing

To address the problems of multi-station analysis and missing station data in geomagnetic monitoring data, based on graph neural network, a regional seismic short prognostic anomaly detection method is proposed, which utilizes the vertex information exchange process of graph convolution to achieve overall multi-station analysis, and introduces a vertex random discard link in the model training process to enhance the model’s recognition of partially missing data. To facilitate the modeling of the importance of multiple stations, an attention mechanism is introduced in the graph readout layer. On the $A E T A$ dataset containing missing data, 85.29 % of the data were identified by the network before the earthquake, and the anomaly detection accuracy reached 73.68 %, and two earthquakes with Ms (magnitude) $\geq 5.7$ were found to be station synchronization anomalies before the earthquake.