Multi-Station Collaborative Analysis of Impending Seismic Precursor Based on Graph Neural Networks

Abstract

Multi-station collaborative analysis is an important part of impending seismic precursor analysis. However, most analysis methods rely on manual feature selection and visual observation, and data missing is another problem in analysis. This paper proposes a method based on graph neural networks (GNNs) to facilitate message passing of adjacent stations, which is helpful to perform collaborative analysis of geomagnetic signals in the region and reduce the impact of data missing problem. A vertex drop layer is introduced in model training process for data enhancement and attention mechanism is introduced in the graph readout layer to model the importance of each station. On AETA dataset containing missing data, anomalies are found before 79.41% earthquakes, and anomaly detection precision reached 69.09%. Synchronized anomalies between stations are found before two big earthquakes. Besides, attention analysis shows the model can estimate the importance of each station, and the difference of attention weights can be explained by the data quality of stations.