Earthquake Prediction Based On Stable Causal Feature Selection

The main task of earthquake prediction is to determine the time,location and magnitude of an earthquake,in order to give early warning before the earthquake occurs,providing time for people to avoid risks and escape,reduce property damage and casualties.However,the existing research techniques for earthquake prediction are not mature.First,there are often missing values in the earthquake-related data collected by sensors,which require data preprocessing before prediction.Traditional methods of dealing with missing values such as deletion,mean and median filling often cannot make full use of the original data,resulting in a waste of time and human cost in collecting the data.Second,for the numerous precursor feature data,existing feature selection algorithms mainly rely on the correlation between features and response variables,making these algorithms lack interpretability and stability,which lead to incorrect conclusions.To better address the current issues in earthquake prediction,this thesis conducts the following three pieces of research works:(1)Missing value filling.Missing value processing is one of the key steps in data analysis.To address the issue of missing values in the data,this work uses the Auto Impute algorithm based on the autoencoder for imputation.Due to equipment failure,improper collection and other reasons,collected seismic data often have missing values,resulting in discontinuous time series data,which affects subsequent analysis and processing.Compared with traditional missing value imputation methods,the Auto Impute algorithm has advantages in handling time series data with multiple features.Mainly based on the overcomplete autoencoder,learn the inherent distribution and pattern of the original data,and reconstruct the expression matrix by projecting the time series data into the high-dimensional latent space,so as to fill the data and restore the real data to the greatest extent,and improve the utilization of the original data.Compared with Random Forest(RF)and K-Nearest Neighbor(KNN)imputation algorithms,the superiority of the Auto Impute algorithm filling data is verified.(2)Feature selection.In many tasks of data mining and machine learning,selecting relevant and key features from high-dimensional data is an important task.To improve the interpretability and stability of the feature selection algorithm,this work proposes a Stable Causal Feature Selection algorithm(SCFS)with direct causal effects based on the Markov Blanket(MB),which achieves interpretable feature selection by considering the causal relationship between features and response variables.Specifically,this thesis proposes a theorem based on graphical causal models to estimate direct causal effects to support the proposed causal feature selection algorithm.In addition,the SCFS algorithm also has good applicability in real-world applications,combining it with a Bidirectional Long Short Term Memory neural network(Bi LSTM)to predict earthquakes(SCFS.Bi LSTM).The feasibility and effectiveness of the proposed algorithm and model are verified through simulation experiments on real and synthetic datasets.(3)Attention mechanism.This thesis introduces attention mechanism to address the problem of long-distance information weakening in time series prediction tasks.At the same time,the prediction ability of the model is further improved by automatically obtaining the important weights of precursor feature sequences at different times.Although the traditional Bi LSTM model performs well in time series prediction,it suffers from the problem of information loss in long-term prediction tasks,which makes accurate prediction difficult.In order to solve this problem,this thesis proposes a model called Earth-ANN,which improves the accuracy of earthquake prediction by introducing attention mechanism and giving greater weight to more important precursory features.