Research On Earthquake Prediction Model Based On Multi-Source Deep Learning

Earthquakes have been proven to be one of the most destructive natural disasters in the world,and their direct and secondary disasters pose a significant threat to human life,infrastructure,and the national economy.Therefore,if future earthquakes can be predicted,then appropriate precautions can be taken in advance to minimize the impact caused by earthquakes.In order to collect signal characteristics with impending characteristics associated with earthquakes,the Shenzhen Earthquake Monitoring and Prediction Technology Research Center of Peking University has developed a multi-component seismic monitoring system AETA,which has accumulated a vibrant set of observation data so far.In this thesis,based on the AETA observation data and using deep learning theory,we investigate the problem of short proximity prediction of earthquakes from the perspective of time series prediction,and the main work is as follows:(1)A pre-processing method for AETA data is proposed.This includes detecting and deleting outliers due to power outages,interpolating missing values due to network outages,and the logarithm transformation of the original data to reduce the impact on the model performance due to the inconsistency of the feature data.Then a method for labeling station epicentral distances and magnitudes on the AETA dataset is proposed to assist in achieving the goals of the earthquake prediction task.Based on this,various machine learning algorithms are used for feature selection on the AETA data,and the effectiveness of the process is demonstrated in subsequent model experiments.(2)An earthquake prediction model is proposed.In this thesis,from the perspective of feature fusion,we first improve the original Informer model into a multi-encoder-singledecoder architecture and then introduces a multimodal cooperative attention mechanism and a low-rank multimodal feature fusion algorithm for this model.The former solves the problem that data from different sources cannot interact with each other in the early fusion process of the model,and the latter improves the problem that the data from multiple sources cannot be fully fused due to the simple splicing of features.Finally,a multiencoder-single-decoder model is proposed based on the multi-source cooperative attention mechanism.Based on this model,ablation and comparison experiments for different input sequence lengths and input features are conducted on the AETA dataset.The results of the ablation experiments demonstrate the effectiveness of the two main modules of the model.And the results of the comparison experiments show that the proposed model outperforms other algorithms in terms of MAE,MSE,and RMSE under all conditions.Also,this thesis proposes a method to convert the station epicenter distance output from the model into latitude and longitude coordinates and realize the prediction of the three elements of the earthquake.(3)An earthquake forecasting system is designed and implemented.Using the multiencoder-single-decoder model based on the multi-source cooperative attention mechanism proposed in this thesis as the core component,an earthquake forecasting system was designed and implemented,and the system was functionally tested by using black-box testing methods.