Research On Deep Learning Earthquake Prediction Model Based On Geoacoustic And Electromagnetic Data

Short-term earthquake prediction has been an important and difficult problem in geophysics that is difficult to solve.With the introduction of deep learning technology into seismology in recent years,earthquake short-term prediction has made some development,and some researchers have tried to use deep learning technology to carry out earthquake prediction research.In the field of seismic short-term prediction based on geoacoustic and electromagnetic data,there are few prediction models and low performance of model seismic short-term prediction at present.To solve the above problems,this thesis applies the existing deep learning techniques on geoacoustic and electromagnetic data,constructs a CNN-GRU earthquake prediction model,and introduces the Inception-Res Net module with attention mechanism to improve the prediction performance of the network.The main work of this thesis is as follows:1.To use the geoacoustic and electromagnetic data for deep learning model training,this thesis designs a method to construct a seismic sample set using the geoacoustic and electromagnetic monitoring data of AETA system.The method uses filtering,downsampling,and complementation to clean and preprocess the data? uses the sliding window method to construct the samples? and combines the historical seismic records to annotate the samples to transform the geoacoustic and electromagnetic data into a sample set for deep learning models.2.In this thesis,we first design a CNN-GRU multi-task prediction model and improve the Inception-Res Net-V2 module in the field of computer vision and embed it into the model to form a CNN-GRU seismic prediction model based on improved InceptionRes Net.The model integrates the use of CNN and GRU structures,extracts sample local features and global features on the time series using CNN network and GRU network respectively,performs feature fusion of the two types of features,and performs earthquake latitude and longitude and earthquake category prediction simultaneously.The introduced Inception-Res Net module improves the model prediction performance by increasing the depth of CNN layers to improve the extraction of local features.3.To further explore the methods to improve the prediction ability of the network,this paper introduces 2 attention modules in the CNN-GRU network and designs a way to improve the feature picking ability of the model using the attention module to obtain better prediction performance.The model uses channel attention to pick the local features extracted by the CNN network and uses a multi-head attention mechanism to pick the sequence global features extracted by the GRU network to improve the prediction performance by improving the feature picking ability of the model.In this thesis,an AETA geoacoustic and electromagnetic dataset is constructed,two CNN-GRU seismic prediction models are designed,and the constructed dataset is used to conduct comparison experiments and ablation experiments on the CNN-GRU seismic prediction models to verify the seismic prediction capability of the models.Finally,in order to reduce the threshold of using the models for researchers in seismology and geology,this thesis uses Spring Boot as the basic framework to design an auxiliary system for short-term earthquake prediction,and makes an attempt to put the earthquake prediction model into practical use.