Exploration Of LSTM-based Assimilation Analysis Of Single-station Multi-path Observation Data

At this stage,accurate forecasting of earthquakes is a worldwide scientific challenge.During the gestation process of earthquakes,different precursor phenomena appear due to the increase of stress,and most of the methods of earthquake forecasting are based on the analysis of physical quantities in a single discipline,missing the models with physical significance and the methods of assimilation and analysis of earthquake precursor data.In the past 30 years,deep learning has made tremendous progress,and this technique has penetrated in several fields and has had a profound impact on people’s lives and scientific research.In particular,deep learning techniques are widely used in the field of seismology,for example,in earthquake precursor data prediction.Deep learning techniques have many advantages and provide new ideas for assimilation analysis of seismic data and new methods for more effective data prediction.In this paper,we first summarize and analyze the commonly used machine learning algorithms,and select from them the long short-term memory network(LSTM)machine learning algorithm to build a time series data prediction model,propose a data prediction model based on LSTM,and conduct an in-depth study on the preprocessing,activation function,optimization function,model framework,and parameter tuning of precursor data.Taking the 2021 Mado earthquake as an example,the two observations of deformation and water temperature are predicted and analyzed in a time-frequency manner,and the expanded diffusion model is combined to do assimilation analysis on the observed data.In this paper,the pre-processing process of earthquake precursor data is introduced in detail,and the pre-processing process of borehole strain and water temperature data is analyzed to provide reliable time series data for later studies.This paper also introduces the experimental environment of LSTM prediction algorithm and the data processing specification used.To improve the prediction accuracy of LSTM,the empirical mode decomposition method(EMD)is introduced,which can decompose the non-smooth signal into smooth signal,and the smooth signal is more suitable for prediction using LSTM,and the prediction accuracy of EMD-LSTM is more accurate than using a single LSTM model.Meanwhile,this paper introduces the time-frequency analysis and the dilation diffusion model.In this paper,we choose the Mado MS 7.4 earthquake as an earthquake example,combine the shallow water temperature observations and borehole strain observations from the Yushu seismic station with the dilation diffusion model,assimilate the observations,and draw the conclusions of this paper,which can correctly verify whether the pre-earthquake precursor changes are consistent with the trend changes proposed by the dilation diffusion model.