Research On Seismic Precursor Anomaly Detection

Earthquake is a natural disaster that is extremely harmful to human society.Pregnant of the earthquake is an extremely complex process.The research on earthquakes is still in the preliminary exploration stage.Numerous studies have shown that the occurrence of seismic events is accompanied by a variety of precursor anomalies.Therefore,it is very important to explore the correlation between precursor anomalies and earthquakes for understanding and predicting earthquakes.Geomagnetic field and seismicity obtained from historical seismic data are two common precursors of earthquakes.We conducts seismic precursor anomaly detection research based on these two precursors.The correlation between the temporal and spatial anomalies of b value,a kind of seismicity feature,and earthquakes is analyzed.Firstly,the statistical and physical significance of b value is summarized based on Gutenberg-Richter law.Then the spatial and temporal anomalies of b-values before the 2017 Jiuzhaigou M _S7.0 earthquake are analyzed.It is found that the decreasing trend of b-value is strongly correlated with the future earthquakes.What’s more,future earthquakes are likely to occur in a region with a much lower b-value.The correlation between the spatial and temporal anomalies of the fractal dimension of the geomagnetic field and earthquakes is analyzed.Tukey box plots combined with sliding windows are used to extract the anomalies of the fractal dimension.The analysis reveals that several stations showed simultaneous anomalies within 45 days before the earthquake,which are likely to be affected by the same earthquake.The closer the stations are to the earthquake,the more sensitive they are to the earthquake.In addition,the spatial distribution of fractal dimensions before the 2019 Changning M_S6.0 earthquake and the 2017 Jiuzhaigou M _S7.0 earthquake are plotted based on the kriging interpolation method.The analysis shows that there is a correlation between the epicenter of an earthquake and the region with low value anomalies in fractal dimension.Different precursor data have different sensitivities to earthquakes.Since some geomagnetic data are missing,we propose a method based on incomplete view learning to fuse features from two types of precursor data,seismicity and geomagnetic field,respectively,to obtain precursor features with more comprehensive information.The information completeness and structural separability of the fused features are fully considered.It is found that the fused features contain more information related to the seismic incubation state than the original precursor features.To avoid the interference of non-seismic anomalies,a supervised algorithm is used for seismic precursor anomaly detection.Considering the problem of class imbalance in the dataset,an ensemble learning method based on a resampling mechanism is proposed,which is able to train the model adaptively according to the classification difficulty of the samples.The method can fully learn the features of difficult samples without overfitting.Compared with other methods,it has a large improvement in anomaly detection performance.In view of the one-sidedness of the precursor information of a single station,graph neural network is used to mine the spatial correlation between the precursor information of multiple stations.At the same time,considering the aggregation phenomenon of earthquake occurrence,an attention mechanism is introduced to give higher attention to key nodes.The results show that considering multiple precursor data sources and spatial correlations among multiple stations is a useful strategy to obtain more comprehensive earthquake precursor information,which helps to capture more anomalies associated with earthquakes.Finally,a visualization platform for precursor data analysis are designed and implemented.It facilitates researchers to perform precursor data analysis in a more intuitive way and has practical value.