Earthquake prediction method and device

Abstract

The present invention provides an earthquake prediction method and device, comprising: obtaining raw monitoring data of monitoring points in a monitored area for different time periods; preprocessing the raw monitoring data; extracting PCA outliers from the preprocessed data using an anomaly detection algorithm; fitting and training the outliers using a neural network to obtain a neural network model of training data and earthquake dates; and predicting earthquakes using the neural network model. The present invention does not require any damage to the surface and buildings near the measured points, and can perform long-term monitoring of the monitored area through fixed monitoring stations, and has good feasibility of implementation.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake prediction, and in particular to an earthquake prediction method and device. Background Art

[0002] Earthquakes are natural disasters that pose a great threat to human society and have caused huge socioeconomic losses to human society. China has frequent earthquake activities and is one of the countries most severely affected by earthquake disasters in the world. In an attempt to solve the problem of earthquake monitoring and prediction, the Key Laboratory of Integrated Microsystems of the Shenzhen Graduate School of Peking University has developed the AETA multi-component earthquake monitoring system, hoping to establish a complete monitoring network through low-cost, large-area, and high-density equipment deployment, capture more effective earthquake precursor signals, and serve earthquake prediction. However, most existing earthquake prediction methods will cause damage to the surface and buildings near the measured points, and their execution is not high. Summary of the invention

[0003] The main purpose of the present invention is to provide an earthquake prediction method and device to solve the problems existing in the prior art.

[0004] In order to achieve the above object, the present invention provides an earthquake prediction method, comprising:

[0005] Obtaining raw monitoring data of monitoring points in the area to be monitored for different time periods;

[0006] Preprocessing the raw monitoring data;

[0007] Use anomaly detection algorithm to extract PCA outliers from preprocessed data;

[0008] Use a neural network to perform fitting training on the abnormal value to obtain a neural network model of training data and earthquake date;

[0009] The neural network model is used to predict earthquakes.

[0010] Optionally, the preprocessing of the acquired raw data specifically includes:

[0011] Supplementing missing values ​​in the raw monitoring data;

[0012] The eighth-order Butterworth filter is used to filter the data after missing value supplementation to obtain low-frequency data.

[0013] Optionally, the extracting PCA outliers using an anomaly detection algorithm on the preprocessed data specifically includes:

[0014] Calculate the electromagnetic mean characteristic X of the low-frequency data true ;

[0015] Calculate the background reference value sequence X of the low-frequency data R ;

[0016] According to the electromagnetic mean characteristic X true and the background reference value sequence X R An outlier value of the low-frequency data is calculated.

[0017] Optionally, the calculation of the electromagnetic mean characteristic X of the low-frequency data true Specifically include:

[0018] For the low-frequency data, take the absolute value of all values, then intercept the sequence with a minute window, take the average value of the data in the window, and call the average value the ULF feature. Slide the window to obtain a ULF feature sequence with a minute granularity; then use the hour window to average the ULF feature sequence with a minute granularity to obtain a ULF feature sequence with an hour granularity, which is called the electromagnetic mean feature X. true .

[0019] Optionally, the calculation of the background reference value sequence X of the low-frequency data R Specifically include:

[0020] From the electromagnetic mean characteristic X true Get the time series matrix X from the sequence, calculate the covariance matrix C of X, and calculate the eigenvalues ​​λ1,λ2,…,λ of C n (The order is λ1≥λ2≥…≥λ n ) and the corresponding eigenvectors E1, E2, …, E n ;

[0021] Define the cumulative contribution rate α, let k increase continuously from 1 until α is greater than or equal to 85% for the first time, retain the first k eigenvalues ​​λ, and take the corresponding eigenvectors to form the eigenvector matrix E * =(E1,E2,…,E k ), we get the matrix P = XE * , and then get the reconstructed matrix X p =PE *T ;

[0022] Select X p The data in the last column is used as the background reference value X for the data on day n ref ;

[0023] The electromagnetic mean characteristic X in the time period to be analyzed true , using 27 days as the window length, using a sliding window to intercept the sequence with a step length of hours, and using the above method to calculate the background reference value X ref , and obtain a background reference value sequence X with hourly granularityR .

[0024] Optionally, the electromagnetic mean characteristic X true and the background reference value sequence X R Calculating the outlier value of the low-frequency data specifically includes:

[0025] Take the day as the window and true With X R The two sequences are intercepted, and the corresponding positions of the intercepted sequences are subtracted to obtain the difference between the 24 background reference values ​​and the corresponding electromagnetic mean characteristics for 24 hours in a day;

[0026] Sort all the differences from small to large, take the 23rd number as the tolerance Δ, define the upper limit up and the lower limit down, and calculate the abnormal value ΔX every hour abn ;

[0027] For the electromagnetic mean feature, if it is between the upper limit up and the lower limit down, it is judged as a normal point; otherwise, the electromagnetic mean feature exceeds the upper and lower thresholds, and the corresponding ΔX is calculated abn ; From 24 ΔX per day abn Select the largest one as the outlier of the day;

[0028] Get a daily granularity outlier sequence.

[0029] Optionally, the using a neural network to perform fitting training on the abnormal value to obtain a neural network model of training data and earthquake date specifically includes:

[0030] Determine earthquake days based on monitored seismic events;

[0031] The abnormal value sequence is intercepted with a window length of 27 days and a step length of one day to obtain a series of data segments, which are inputs for training the neural network;

[0032] Starting from the last day of each data segment, check the earthquake situation within a period of time. If there is an earthquake day, the label of the data segment is set to 1, otherwise it is set to 0.

[0033] A series of 27-day-long training data and their corresponding labels are obtained, and a neural network is used to fit and train them to obtain a neural network model of training data and earthquake days.

[0034] The present invention also provides an earthquake prediction device, comprising:

[0035] An acquisition unit, used for acquiring raw monitoring data of monitoring points in the area to be monitored for different time periods;

[0036] A preprocessing unit, used for preprocessing the raw monitoring data;

[0037] An outlier extraction unit, used for extracting PCA outliers from the preprocessed data using an anomaly detection algorithm;

[0038] A model building unit, used for fitting and training the abnormal value using a neural network to obtain a neural network model of training data and earthquake date;

[0039] An earthquake prediction unit is used to predict earthquakes using the neural network model.

[0040] Optionally, the preprocessing unit specifically includes:

[0041] A missing value supplement subunit is used to supplement the missing values ​​of the raw monitoring data;

[0042] The eighth-order Butterworth filter is used to filter the data after missing values ​​are supplemented to obtain low-frequency data.

[0043] Optionally, the outlier extraction unit specifically includes:

[0044] A first calculation subunit, used for calculating the electromagnetic mean feature Xtrue of the low-frequency data;

[0045] The second calculation subunit is used to calculate the background reference value sequence X of the low-frequency data. R ;

[0046] The third calculation subunit is used to calculate the electromagnetic mean value feature X true and the background reference value sequence X R An outlier value of the low-frequency data is calculated.

[0047] The beneficial effects of the present invention are: an innovative earthquake prediction method based on correlation analysis of earthquake events and precursor anomalies is proposed; non-destructive monitoring can be achieved without causing any damage to the surface and buildings near the measured points, and the deployment is convenient; and long-term monitoring of the monitored area can be carried out through fixed monitoring stations, and the implementation feasibility is good.

[0048] Other features and advantages of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present invention but do not constitute a limitation of the present invention. In the accompanying drawings:

[0050] Figure 1 A schematic diagram of a flow chart of an earthquake prediction method according to an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of setting monitoring points in a monitoring area to be monitored according to an embodiment of the present invention;

[0052] Figure 3 is a calculation formula of the time series matrix X of an embodiment of the present invention;

[0053] Figure 4 is a calculation formula for the cumulative contribution rate α of the embodiment of the present invention;

[0054] Figure 5 is a calculation formula for the upper limit up and the lower limit down of the embodiment of the present invention;

[0055] Figure 6 is the abnormal value ΔX per hour in the embodiment of the present invention abn Calculation formula of

[0056] Figure 7 is the calculation formula of the Es parameter of the embodiment of the present invention;

[0057] Figure 8 The present invention is a schematic structural diagram of an earthquake prediction device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0059] Figure 1 The present invention is a flowchart of an earthquake prediction method according to an embodiment of the present invention, wherein the steps include:

[0060] S101, deploying monitoring points in the area to be monitored;

[0061] First, select the hot spot area to be monitored and set up multiple monitoring points in the area. Figure 2 As shown, it is a schematic diagram of setting monitoring points for the monitored area in an embodiment. The range encircled by curve 202 is the monitored area, and marking line 203 and marking line 204 are the length and width of the positioning accuracy P (unit meter). According to the monitoring accuracy requirements, multiple monitoring points are set in the monitored area. The density of the monitoring points depends on the monitoring accuracy of the local earthquake. The higher the accuracy requirement, the higher the density of the monitoring points. In one embodiment, the monitoring points include fixed monitoring points 201 and temporary monitoring points 205. The fixed monitoring points 201 can realize long-term stable monitoring and provide rich raw data based on rich events to facilitate the screening of earthquake precursor characteristics.

[0062] S102, monitoring earthquake precursor signals at monitoring points in the monitored area;

[0063] The earthquake precursor signals at the monitoring points in the monitored area are monitored, and the monitoring raw data of the earthquake electromagnetic disturbance signals at different times at the monitoring points in the monitored area and the earthquake events occurring during the monitoring time are obtained. The earthquake events are all earthquake events occurring within a certain range near the monitoring point, such as within a range of 100km. The time when the earthquake event occurs during the monitoring time includes a certain period of time after the installation date of the earthquake monitoring point, and the preferred time length is greater than 1 month.

[0064] S103, pre-processing the data acquired by the monitoring device (hereinafter referred to as monitoring raw data);

[0065] During the monitoring process, due to weather, equipment failure and other possible influencing factors, the data obtained by the monitoring equipment may be interrupted or missing, so the missing values ​​of the original monitoring data are supplemented. Linear interpolation or other methods can be used to interpolate the missing data.

[0066] The frequency range of the raw monitoring data acquired by the AETA system monitoring terminal covers 0.1 Hz to 10 kHz. The embodiment of the present invention mainly focuses on the precursor characteristics of the ultra-low frequency components below 30 Hz in the low frequency range of 0.1 Hz to 200 Hz of the raw monitoring data, so a digital filter needs to be designed to filter out the components above 30 Hz.

[0067] Considering that the filtering step hopes to not change the relative influence of the signals of each frequency band on the filtering result as much as possible, and considering the balance between the filtering effect of the filter order and the computing power of the AETA terminal equipment, an eighth-order Butterworth filter is finally selected to obtain the filtered low-frequency data.

[0068] S104, extracting PCA outliers from the preprocessed data using an anomaly detection algorithm;

[0069] This step specifically includes:

[0070] Calculate the electromagnetic mean characteristic X of the low-frequency data true ;

[0071] Calculate the background reference value sequence X of the low-frequency data R ;

[0072] According to the electromagnetic mean characteristic X true and the background reference value sequence X R An outlier value of the low-frequency data is calculated.

[0073] Among them, the electromagnetic mean characteristic feature X of the low-frequency data is calculated true Specifically include:

[0074] For the low-frequency data, take the absolute value of all values, then intercept the sequence with a minute window, take the average value of the data in the window, and call the average value the ULF feature. Slide the window to obtain a ULF feature sequence with a minute granularity; then use the hour window to average the ULF feature sequence with a minute granularity to obtain a ULF feature sequence with an hour granularity, which is called the electromagnetic mean feature X. true .

[0075] Wherein, the background reference value sequence X of the low-frequency data is calculated R Specifically include:

[0076] The electromagnetic mean characteristic X obtained in the previous step true It is a one-dimensional sequence with hourly granularity, starting from the installation time to the current time in chronological order, first starting from the electromagnetic mean feature X true The time series matrix X is obtained from the sequence. For any continuous time period of n days (n is generally 27) within the monitored time period, the time series matrix X is as follows: Figure 3 The matrix consists of the AETA electromagnetic mean characteristics of n days, and each column represents the electromagnetic mean characteristics of an hourly time interval within a day, that is, x 1,1 represents the first hour of the first day, x 24,1 represents the 24th hour of the first day, x 1,n represents the first hour of the nth day, x 24,n represents the 24th hour of the nth day. Calculate the covariance matrix C of X. And calculate the eigenvalues ​​λ1,λ2,…,λ of C n (The order is λ1≥λ2≥…≥λ n ) and the corresponding eigenvectors E1, E2, …, E n ;

[0077] Define the cumulative contribution rate α as Figure 4 As shown, let k start from 1 and continue to increase until α is greater than or equal to 85% for the first time, retain the first k eigenvalues ​​λ, and take the corresponding eigenvectors to form the eigenvector matrix E * =(E1,E2,…,E k ), we get the matrix P = XE * , and then get the reconstructed matrix X p =PE *T ;

[0078] Select X p The data in the last column is used as the background reference value X for the data on day n ref ;

[0079] The electromagnetic mean characteristic X in the time period to be analyzed true, using 27 days as the window length, using a sliding window to intercept the sequence with a step length of hours, and using the above method to calculate the background reference value X ref , and obtain a background reference value sequence X with hourly granularity R .

[0080] According to the electromagnetic mean characteristic X true and the background reference value sequence X R Calculating the outlier value of the low-frequency data specifically includes:

[0081] X true With X R are all hourly granularity sequences, and their data are one-to-one corresponding. Taking the day (24 hours) as the window, true With X R The two sequences are intercepted, and the corresponding positions of the intercepted sequences are subtracted to obtain the difference between the 24 background reference values ​​and the corresponding electromagnetic mean characteristics for 24 hours in a day;

[0082] Sort all the differences from small to large, take the 23rd number as the tolerance Δ, and define the upper limit up and the lower limit down, such as Figure 5 As shown; calculate the abnormal value ΔX per hour abn ,like Figure 6 As shown;

[0083] For the electromagnetic mean feature, if it is between the upper limit up and the lower limit down, it is judged as a normal point; otherwise, the electromagnetic mean feature exceeds the upper and lower thresholds, and the corresponding ΔX is calculated abn ; From 24 ΔX per day abn Select the largest one as the outlier of the day;

[0084] Get a daily granularity outlier sequence.

[0085] S105. Model training and prediction based on neural network.

[0086] Determine the earthquake date based on the monitored earthquake events: The magnitude and epicenter distance of the earthquake are two important parameters that characterize the impact of the earthquake on a certain area. In addition, many earthquakes with higher magnitudes are accompanied by dense aftershock sequences. Taking the above factors into consideration, the embodiment of the present invention uses Figure 7 The Es parameter shown is used as the definition standard for earthquake days to quantify earthquake events. Among them, M is the Richter magnitude, r is the distance between the observation point and the epicenter, Es' can be regarded as the impact of a single earthquake on the observation point, and Es is the superimposed impact value of earthquakes occurring on the same day. The threshold of Es is taken as 10 8 , the date when Es is greater than this value is considered to be the earthquake day of the observation point.

[0087] The abnormal value sequence obtained above is intercepted with a window length of 27 days and a step length of one day to obtain a series of data segments, which are the inputs for neural network training;

[0088] Taking the last day of each data segment as the starting point (not including that day), check the earthquake situation for a period of time, such as a week, that is, whether there is an earthquake day within a week after the observation point. If there is an earthquake day, the label of the data segment is set to 1, otherwise it is set to 0; obtain a series of 27-day-long training data and its corresponding labels, use a neural network to fit and train them, and obtain a neural network model of training data and earthquake days;

[0089] When making predictions, the same processing method as above is used to extract a data segment of a time period (27 days long), and the trained neural network model can be used to predict the data segment, thereby obtaining a prediction of the earthquake situation within one week after the corresponding time period near the monitoring point.

[0090] Based on the above embodiments, multiple monitoring points can be set in the monitored area according to the actual monitoring accuracy requirements, and the monitoring raw data of earthquake precursor signals at different times at the monitoring points and the earthquake events occurring during the monitoring time can be monitored. The monitoring raw data of the preset time period before the earthquake event is statistically analyzed, and its abnormal features are extracted according to the embodiment method. The correlation between the monitoring data and the earthquake event is learned using a neural network model to predict the earthquake event. The embodiment of the present invention innovatively proposes an abnormality extraction and earthquake prediction method based on AETA monitoring data, which does not require any damage to the surface and buildings near the measured point, and can perform long-term monitoring of the monitored area through fixed monitoring stations, and has good feasibility.

[0091] like Figure 8 As shown, an embodiment of the present invention further provides an earthquake prediction device, comprising:

[0092] An acquisition unit, used for acquiring raw monitoring data of monitoring points in the area to be monitored for different time periods;

[0093] A preprocessing unit, used for preprocessing the raw monitoring data;

[0094] An outlier extraction unit, used for extracting PCA outliers from the preprocessed data using an anomaly detection algorithm;

[0095] A model building unit, used for fitting and training the abnormal value using a neural network to obtain a neural network model of training data and earthquake date;

[0096] An earthquake prediction unit is used to predict earthquakes using the neural network model.

[0097] Optionally, the preprocessing unit specifically includes:

[0098] A missing value supplement subunit is used to supplement the missing values ​​of the raw monitoring data;

[0099] The eighth-order Butterworth filter is used to filter the data after missing values ​​are supplemented to obtain low-frequency data.

[0100] Optionally, the outlier extraction unit specifically includes:

[0101] A first calculation subunit, used for calculating the electromagnetic mean feature Xtrue of the low-frequency data;

[0102] The second calculation subunit is used to calculate the background reference value sequence X of the low-frequency data. R ;

[0103] The third calculation subunit is used to calculate the electromagnetic mean value feature X true and the background reference value sequence X R An outlier value of the low-frequency data is calculated.

[0104] The first calculation subunit is specifically used to take the absolute value of all values ​​of the low-frequency data, then intercept the sequence with a minute window, take the average value of the data in the window, and call the average value the ULF feature. Slide the window to obtain a ULF feature sequence with a minute granularity; then take the average value of the ULF feature sequence with a minute granularity with an hour window to obtain a ULF feature sequence with an hour granularity, which is called the electromagnetic mean feature X. true .

[0105] The second calculation subunit is specifically used to calculate the electromagnetic mean feature X true Get the time series matrix X from the sequence;

[0106] For any continuous period of n days (n is usually 27) within the monitored period, the time series matrix X is as follows: Figure 3 The matrix consists of the AETA electromagnetic mean characteristics of n days, and each column represents the electromagnetic mean characteristics of an hourly time interval within a day, that is, x 1,1 represents the first hour of the first day, x 24,1 represents the 24th hour of the first day, x 1,n represents the first hour of the nth day, x 24,n represents the 24th hour of the nth day. The second computing subunit is also used to calculate the covariance matrix C of X. And calculate the eigenvalues ​​λ1,λ2,…,λ of C n (The order is λ1≥λ2≥…≥λ n ) and the corresponding eigenvectors E1, E2, …, E n ;

[0107] Define the cumulative contribution rate α as Figure 4 As shown, let k start from 1 and continue to increase until α is greater than or equal to 85% for the first time, retain the first k eigenvalues ​​λ, and take the corresponding eigenvectors to form the eigenvector matrix E * =(E1,E2,…,E k ), we get the matrix P = XE * , and then get the reconstructed matrix X p =PE *T ;

[0108] The second calculation subunit is also used to select X p The data in the last column is used as the background reference value X for the data on day n ref ;

[0109] The second calculation subunit is also used to analyze the electromagnetic mean characteristic X within the time period. true , using 27 days as the window length, using a sliding window to intercept the sequence with a step length of hours, and using the above method to calculate the background reference value X ref , and obtain a background reference value sequence X with hourly granularity R .

[0110] The third calculation subunit is specifically used to calculate X using the day (24 hours) as a window. true With X R The two sequences are intercepted, and the corresponding positions of the intercepted sequences are subtracted to obtain the differences between the 24 background reference values ​​and the corresponding electromagnetic mean characteristics within 24 hours of the day; all the differences are sorted in order from small to large, and the number ranked at the 23rd position is taken as the tolerance Δ, and the upper limit up and the lower limit down are defined, such as Figure 5 As shown; calculate the abnormal value ΔX per hour abn ,like Figure 6 As shown;

[0111] For the electromagnetic mean feature, if it is between the upper limit up and the lower limit down, it is judged as a normal point; otherwise, the electromagnetic mean feature exceeds the upper and lower thresholds, and the corresponding ΔX is calculated abn ; From 24 ΔX per day abn Select the largest one as the outlier of the day;

[0112] Get a daily granularity outlier sequence.

[0113] The model building unit specifically includes:

[0114] The earthquake date determination subunit is used to determine the earthquake date according to the monitored earthquake events: the magnitude and epicenter distance of the earthquake are two important parameters that characterize the impact of the earthquake on a certain area. In addition, many earthquakes with higher magnitudes are accompanied by dense aftershock sequences. Taking the above factors into consideration, the embodiment of the present invention uses Figure 7 The Es parameter shown is used as the definition standard for earthquake days to quantify earthquake events. Among them, M is the Richter magnitude, r is the distance between the observation point and the epicenter, Es' can be regarded as the impact of a single earthquake on the observation point, and Es is the superimposed impact value of earthquakes occurring on the same day. The threshold of Es is taken as 10 8 , the date when Es is greater than this value is considered to be the earthquake day of the observation point.

[0115] The neural network model establishment subunit is used to intercept the abnormal value sequence obtained above with 27 days as the window length and one day as the step length to obtain a series of data segments, which are the inputs of the neural network for training; starting from the last day of each data segment (and excluding that day), check back a period of time, such as the earthquake situation in a week, that is, whether there is an earthquake day within a week after the observation point, if there is an earthquake day, then the label of the data segment is set to 1, otherwise it is set to 0; obtain a series of training data with a length of 27 days and its corresponding labels, use the neural network to fit and train them, and obtain a neural network model of training data and earthquake days;

[0116] The earthquake prediction unit is used to use the above processing method to cut out the data segment of the time period (27 days long) when making predictions, and then use the trained neural network model to predict the data segment, so as to obtain the prediction of the earthquake situation within one week after the corresponding time period near the monitoring point.

[0117] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, a variety of simple modifications can be made to the technical solution of the present invention, and these simple modifications all belong to the protection scope of the present invention.

[0118] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0119] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. A method for earthquake prediction, characterized in that: include: Obtaining raw monitoring data of monitoring points in the area to be monitored for different time periods; Preprocessing the raw monitoring data; Use anomaly detection algorithm to extract PCA outliers from preprocessed data; Use a neural network to perform fitting training on the abnormal value to obtain a neural network model of training data and earthquake date; Using the neural network model to predict earthquakes; The preprocessing of the acquired raw data specifically includes: Supplement missing values ​​for the raw monitoring data; use an eighth-order Butterworth filter to filter the data after missing value supplementation to obtain low-frequency data; The method of extracting PCA outliers using an anomaly detection algorithm on the preprocessed data specifically includes: Calculate the electromagnetic mean characteristic X of the low frequency data true ; Calculate the background reference value sequence X of the low-frequency data R According to the electromagnetic mean characteristic X true and the background reference value sequence X R An outlier value of the low-frequency data is calculated.

2. The method according to claim 1, characterized in that The calculation of the electromagnetic mean characteristic X of the low-frequency data true Specifically include: For the low-frequency data, take the absolute value of all values, then intercept the sequence with a minute window, take the average value of the data in the window, and call the average value the ULF feature. Slide the window to obtain a ULF feature sequence with a minute granularity; then use the hour window to average the ULF feature sequence with a minute granularity to obtain a ULF feature sequence with an hour granularity, which is called the electromagnetic mean feature X. true .

3. The method according to claim 2, characterized in that The background reference value sequence X of the low-frequency data is calculated R Specifically include: From the electromagnetic mean characteristic X true Get the time series matrix X from the sequence, calculate the covariance matrix C of X, and calculate the eigenvalues ​​λ1,λ2…,λ n and the corresponding eigenvectors E1, E2, …, E n ; where the order is λ1≥λ2≥…≥λ n ; Define the cumulative contribution rate α, let k increase continuously from 1 until α is greater than or equal to 85% for the first time, retain the first k eigenvalues ​​λ, and take the corresponding eigenvectors to form the eigenvector matrix E * =(E1,E2,…,E k ), we get the matrix P = XE * , and then get the reconstructed matrix X p =PE *T ; Select X p The data in the last column is used as the background reference value X for the data on day n ref ; The electromagnetic mean characteristic X in the time period to be analyzed true , using 27 days as the window length, using a sliding window to intercept the sequence with a step length of hours, and using the above method to calculate the background reference value X ref , and obtain a background reference value sequence X with hourly granularity R ; Among them, the cumulative contribution rate α is defined as follows: In the formula, α is the cumulative contribution rate, k is the number of eigenvalues ​​selected when the cumulative contribution rate α is greater than or equal to 85% for the first time, n is the number of all eigenvalues ​​calculated, and λ i is the i-th eigenvalue.

4. The method according to claim 3, characterized in that According to the electromagnetic mean characteristic X true and the background reference value sequence X R Calculating the outlier value of the low-frequency data specifically includes: Take the day as the window and true With X R The two sequences are intercepted, and the corresponding positions of the intercepted sequences are subtracted to obtain the difference between the 24 background reference values ​​and the corresponding electromagnetic mean characteristics for 24 hours in a day; Sort all the differences from small to large, take the 23rd number as the tolerance Δ, define the upper limit up and the lower limit down, and calculate the abnormal value ΔX every hour abn ; For the electromagnetic mean feature, if it is between the upper limit up and the lower limit down, it is judged as a normal point; otherwise, the electromagnetic mean feature exceeds the upper and lower thresholds, and the corresponding ΔX is calculated. abn ; From 24 ΔX per day abn Select the largest one as the outlier of the day; Get a daily granularity outlier sequence; The upper limit up and the lower limit down are defined as follows: In the formula, up is the upper limit up, low is the lower limit down, Δ is the tolerance, and X ref is the background reference value.

5. The method according to claim 4, characterized in that The use of a neural network to perform fitting training on the abnormal value to obtain a neural network model of training data and earthquake date specifically includes: Determine earthquake days based on monitored seismic events; The abnormal value sequence is intercepted with a window length of 27 days and a step length of one day to obtain a series of data segments, which are inputs for training the neural network; Starting from the last day of each data segment, check the earthquake situation within a period of time. If there is an earthquake day, the label of the data segment is set to 1, otherwise it is set to 0. A series of 27-day-long training data and their corresponding labels are obtained, and a neural network is used to fit and train them to obtain a neural network model of training data and earthquake days.

6. An earthquake prediction device, characterized in that: include: An acquisition unit, used for acquiring raw monitoring data of monitoring points in the area to be monitored for different time periods; A preprocessing unit, used for preprocessing the raw monitoring data; An outlier extraction unit, used for extracting PCA outliers from the preprocessed data using an anomaly detection algorithm; A model building unit, used for fitting and training the abnormal value using a neural network to obtain a neural network model of training data and earthquake date; An earthquake prediction unit, used for performing earthquake prediction using the neural network model; Wherein, the pre-processing unit specifically includes: A missing value supplement subunit is used to supplement the missing values ​​of the original monitoring data; an eighth-order Butterworth filter is used to filter the data after the missing values ​​are supplemented to obtain low-frequency data; Wherein, the outlier extraction unit specifically includes: The first calculation subunit is used to calculate the electromagnetic mean feature X of the low-frequency data true ; The second calculation subunit is used to calculate the background reference value sequence X of the low-frequency data. R ; The third calculation subunit is used to calculate the electromagnetic mean value feature X true and the background reference value sequence X R An outlier value of the low-frequency data is calculated.

Images