Research On Time Series Recognition Of Earthquake Precursory Anomalies

Since the Xingtai earthquake in 1966,China has carried out extensive earthquake prediction research.Geophysical observation data is the basis of earthquake analysis and prediction.The Geophysical Observatory Network of Hebei Province provides various observation data services for earthquake prediction and earth science research through digital observation technology,network transmission and database storage.During earthquakes,it's generally believed that the physical and chemical properties of earth deformation,lithospheric geomagnetic field,underground fluid will change obviously.These changes are reflected in the observed data: over a period of time,the data recorded by the instrument will have abnormal reactions,such as the occurrence of a sudden jump in the value.Seismic precursor observation data are often presented in the form of time series.In this diserstation,the unstable time series of observation data in the process of earthquake is called seismic precursor abnormal time series.In the geophysical observation database of Hebei Province,this information can provide important data support for earthquake prediction.For example,the existence of similar time series,the number of occurrences in historical data,and the recent earthquake occurrence after occurrence.Therefore,it is necessary to study the similar series lookup and recognition method of seismic anomaly time series.Based on the observation data in the geophysical observation database of Hebei Province,this dissertation studies the data preprocessing technology through the main characteristics of seismic precursor observation data.The similarity searching method in time series digging is used to find equal length time series.Based on the Dynamic time warping algorithm of piecewise aggregate,a time series search with different lengths but similar shapes is proposed.Then,the object detection algorithm based on deep learning is studied to reduce the complexity of time series similarity finding.This method can effectively avoid the problems caused by sampling rate inconsistency,time series segmentation and data preprocessing.The main research work of this dissertation is as follows:(1)Experimental method of interpolation.In the data preprocessing stage,the applicable data interpolation method is mainly studied for the missing and broken numbers in thegeophysical observation database of Hebei Province.Using three different interpolation methods to perform interpolation operations with varying degrees of numerical absence.These methods are cubic spline interpolation,linear interpolation and piecewise cubic Hermite interpolation.By comparing the interpolation results with the standard error of the actual data,the practicability of the cubic spline interpolation method for the experimental data in this dissertation is verified.(2)Research on search method based on similarity.Correlation coefficient function corrcoef algorithm and time series similarity measure method are used to calculate the correlation coefficient and distance values between series.Measurement methods include euclidean algorithm,dynamic time warping algorithm and so on.The results are sorted to find similar series with the same length(i.e.the same number of data points)as the seismic precursor anomaly time series.Based on the comprehensive analysis of the experiment running time and running results,it is verified that the corrcoef function has a good effect on the search in this scenario.(3)Dynamic time warping algorithm based on piecewise aggregation.Using the piecewise cumulative data representation method to reduce the dimension of the original data series and generate new feature series.An improved dynamic time warping algorithm that introduces attenuation coefficients is used to find similar time series that are not equal to seismic anomaly time series.It is proved by experiments that the method can improve the search accuracy of unequal length similarity time series.(4)Target detection algorithm based on deep learning.An SSD_MobileNet target detection algorithm is proposed.The sample image is extracted for shape features,where the sample image is from a data set composed of seismic precursory anomaly time series.By using the sample data set,the model of SSD_MobileNet algorithm is trained.In this process,the model applicable to the experimental data series can be extracted and saved for detection and identification testing.The experimental results show that the proposed method can effectively identify similar time series,which greatly reduces the effect of sampling rate diversity on similar time series searching.The innovation of this dissertation is mainly reflected in: In the dynamic time warpingalgorithm based on piecewise aggregation,the dynamic time warping algorithm is improved by the common substring concept of series.The introduction of attenuation coefficient makes the algorithm more accurate in the measurement of distance value.The target detection algorithm based on depth learning does not need to preprocess and segment the time series.It can maintains the original characteristics of the time series.By effectively extracting the curve features of the original time series,it can intuitively and visually realize the similar time series recognition of various lengths.