Sample set construction, seismic monitoring method and system based on generalized neural network

Abstract

The application discloses a kind of sample set construction, generalization neural network-based earthquake monitoring method and system, the monitoring method includes: collecting the seismic wave data of each region and grouping according to epicentral distance and depth to obtain basic database;Randomly generate station distribution in target monitoring range, synthetic earthquake location, calculate the epicentral distance of synthetic earthquake to each station;According to the epicentral distance of station and the depth of synthetic earthquake, select the seismic wave data of the same epicentral distance and depth from the basic database for each station, then rotate to azimuth according to rotation formula;Using the seismic wave data of each station after rotation obtains simulated synthetic seismic wave data as training sample;Using training sample set trains neural network to obtain generalization neural network monitoring model, the application constructs generalization neural network monitoring model based on and realizes the earthquake monitoring of the same range in target monitoring range in any region, when a few stations trigger, can also detect earthquake signal.

Description

Technical Field

[0001] This invention belongs to the field of earthquake monitoring and early warning technology, specifically relating to a sample set construction, an earthquake monitoring method and system based on a generalized neural network. Background Technology

[0002] Earthquake monitoring is a primary task of seismology, and real-time reporting of seismic parameters has always been a crucial aspect of earthquake early warning (EEW) (Hsiao et al., 2009; Nakamura et al., 2009; Allen and Melgar, 2019; Allen and Stogaitis, 2022). The effectiveness of a monitoring system depends not only on its response time but also on its ability to infer seismic parameters from limited waveform data in the initial stages of an earthquake (Horiuchi et al., 2005; Satriano et al., 2008; Kuyuk and Allen, 2013). Current EEW systems typically include multiple modules, such as data processing, source parameter estimation, and alarm filtering (Serdar Kuyuk et al., 2013). Each step requires significant research in seismology to improve the efficiency and accuracy of earthquake detection, phase picking, phase correlation, earthquake location, and magnitude assessment (Li et al., 2018; Ross et al., 2018; Grigoli et al., 2018; Satriano et al., 2011; Zhang et al., 2019; Mousavi and Beroza, 2020; Mousavi and Beroza, 2020; Lomax et al., 2012; Baillard et al., 2013; Zhu and Beroza, 2019; Ross et al., 2019; Liu et al., 2020; Mousavi et al., 2020; Zhang et al., 2022; Zhu et al., 2022). An alarm is triggered if specific conditions are met, such as distinguishing teleseismic events, triggering a certain number of stations, reaching a certain magnitude, and meeting a threshold percentage of triggered stations (e.g., 40%) (Serdar Kuyuk et al., 2013). The complex empirical threshold settings involved in each processing step make defining optimal alarm criteria in an EEW system a challenge. Implementing overly stringent criteria, such as requiring the triggering of too many or too large numbers of stations, can negatively impact the real-time efficiency of the EEW system, while lenient criteria may lead to false alarms (Chung et al., 2019). Therefore, an efficient real-time monitoring algorithm should not only be computationally efficient without requiring extensive empirical configuration, but also capable of solving for seismic parameters under limited data conditions in the early stages of an earthquake.

[0003] Another approach to earthquake monitoring without human intervention involves using deep learning techniques to directly mine data from waveforms for earthquake detection and localization (Perol et al., 2018; Kriegerowski et al., 2018; Mousavi and Beroza, 2020a; Mousavi and Beroza, 2020b; Shen and Shen, 2021; van den Ende et al., 2020; Vinard et al., 2021; Zhang et al., 2020; Zhang et al., 2021; Münchmeyer et al., 2021). These methods resolve earthquake parameters by directly extracting features from seismic waveforms, bypassing many intermediate steps requiring complex empirical settings, such as various thresholds and alert criteria. However, due to the diverse distribution of stations and the diversity of geological structures, the generalization problem of monitoring neural networks remains a challenge. Currently, there is no generalized neural network model that can be applied to different regions globally and to multiple stations for real-time monitoring and early warning. These methods typically require transfer learning when applied to new areas.

[0004] Furthermore, in order to conduct early earthquake warning, neural networks need to process both triggered and untriggered stations from the beginning of the earthquake event (Horiuchi et al., 2005; Satriano et al., 2008). This increases the complexity of neural network learning. Currently, only the method of Zhang et al. (2021) can achieve real-time monitoring of earthquake magnitude and location starting from a small number of stations. However, this method can only be applied to a 100km*50km area in central Italy. The distribution of stations in a specific region is generally fixed, and the stations used in the training and test sets are the same. When applied to other regions, facing new station distributions, it cannot be directly applied and retraining is required. To address the issue of fixed stations, van den Ende et al. (2020) developed a method using graph neural networks to extract features from full waveform data and estimate the source location and magnitude. In this method, each station first extracts features independently, and then uses some network layers to synthesize the features of each station to finally obtain the earthquake location and magnitude. Although this method avoids the problem of fixed stations, it cannot achieve real-time earthquake monitoring starting from a very small number of stations, and assumes that the order of station input has no impact on the results. Moreover, the model cannot be applied to new areas outside the training data.

[0005] This shows that the different regions and different station distributions constrain data compatibility and create technical obstacles for the design of generalized neural networks. Summary of the Invention

[0006] To address the technical problems mentioned in the prior art, such as the difficulty in data sharing due to different regions and station distributions, and the lack of a generalized neural network model applicable to different regions globally and capable of real-time monitoring and early warning using multiple stations, this invention provides a sample set construction method and system for earthquake monitoring based on a generalized neural network. Specifically, this invention fully utilizes the key characteristic that earthquake waveforms with the same epicentral distance and depth across different regions globally share the same P-wave or S-wave travel times, demonstrating significant similarity in their phase travel times. It provides a sample set construction method, essentially a data recombination technique. Using seismic wave data from other regions, earthquakes can be simulated under different locations and monitoring station distributions. These earthquakes possess generalization properties, meaning that the actual earthquake data generated under any monitoring scheme in any region is similar to these generalized earthquakes, thus overcoming the compatibility barriers caused by different regions and station distributions. The training sample set constructed using these generalized earthquakes also possesses generalization properties, thereby building a generalized neural network monitoring model based on a generalized neural network. This model is adaptable to different regions and station distributions, eliminating the need for retraining when deployed to new regions as in traditional methods.

[0007] Therefore, the present invention provides the following technical solution:

[0008] On the one hand, the present invention provides a sample set construction method applied to earthquake monitoring, comprising the following steps:

[0009] Seismic wave data from various regions were collected and grouped according to epicentral distance and depth to obtain a basic database. Seismic wave data with the same epicentral distance and depth were grouped together.

[0010] Obtain the distribution of stations within the target monitoring range, randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station;

[0011] Based on the epicentral distance of the stations and the depth of the composite earthquake, seismic wave data with the same epicentral distance and depth are selected from the basic database for each station, and then the magnitude is rotated and normalized according to the azimuth rotation formula.

[0012] By using the seismic wave data from each station after rotation and normalization, simulated and synthesized seismic wave data within the target monitoring range are obtained, and these data are used as samples to expand the seismic wave sample set within the target monitoring range.

[0013] Further optionally, the azimuth rotation formula is as follows:

[0014]

[0015]

[0016] In the formula, d′ x and d′ y The waveforms of the x and y components in the rotated seismic wave data, d x and d y The waveforms of the x and y components in the original seismic wave data. For the azimuth difference, satisfying Let be the azimuth angle of the i-th station corresponding to the synthetic earthquake. The azimuth angle of the original actual seismic station corresponding to the seismic wave data retrieved from the basic database is given by acosine function.

[0017] The synthetic seismic location within the target monitoring range is represented as (s) x ,s y ,s z The coordinates of the i-th station are represented as follows: s x ,s y Let s be the x and y coordinates of the synthesized earthquake. z The depth of the synthesized earthquake. Let x and y be the x and y coordinates of the i-th station. The three-component waveform of the seismic wave data retrieved from the basic database is represented as (d x ,d y ,d z The three-component waveform of the seismic wave data of the i-th station after azimuth rotation is represented as (d′). x ,d′ y ,d z ).

[0018] The normalized magnitude refers to the process of obtaining the rotated waveform and then adjusting the amplitude of each station according to the magnitude formula so that the waveforms of all stations are normalized to the given magnitude M. In other words, since the waveforms of these stations are all derived from actual earthquakes of different magnitudes, these waveforms need to be transformed into the waveform of a synthetic earthquake of a specific magnitude M before being synthesized.

[0019] Secondly, the present invention provides an earthquake monitoring method based on a generalized neural network, comprising:

[0020] Step 1: Collect seismic wave data from various regions and group them according to epicentral distance and depth to obtain a basic database. Seismic wave data with the same epicentral distance and depth are grouped together.

[0021] Step 2: Randomly generate the distribution of stations within the target monitoring range, and randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station;

[0022] Step 3: Generate a training sample set for the target monitoring range using the method of reconstructing seismic maps from stations;

[0023] Specifically, based on the epicentral distance of the stations and the depth of the synthesized earthquake, seismic wave data with the same epicentral distance and depth are selected from the basic database for each station, and then rotated and normalized according to the azimuth rotation formula; using the rotated and normalized seismic wave data of each station, simulated and synthesized seismic wave data within the target monitoring range is obtained as training samples.

[0024] Step 4: Build a neural network architecture for earthquake monitoring, train the neural network using the training sample set to obtain a generalized neural network monitoring model, and realize earthquake monitoring in any region with the same range as the target monitoring range based on the generalized neural network monitoring model.

[0025] To achieve real-time earthquake monitoring, this invention constructs a generalized neural network monitoring model. Unlike traditional methods that require retraining upon deployment to new areas, this approach proposes a data recombination method to generate earthquakes that simulate different locations and monitoring station distributions. These earthquakes possess generalization properties, meaning that they are similar to actual earthquake data under any monitoring scheme in any region. For example, the arrival times of earthquake phases at different epicentral distances (distances from the station to the earthquake) and depths are almost identical, which are key characteristics for earthquake location. The key reason why existing methods cannot train models with high generalization performance is that they cannot simulate samples under arbitrary monitoring station distributions similar to actual earthquakes, nor can they collect such a large amount of actual earthquake data. Therefore, the technical solution of this invention randomly deploys stations and randomly generates synthetic earthquake locations within a set target monitoring range, enabling the simulated synthetic seismic wave data to have generalization properties. According to this principle, training samples can be generated under multiple station deployments and multiple synthetic earthquake locations within the target monitoring range. On the one hand, this ensures the generalization of the training samples, and on the other hand, it enriches the training sample set, that is, in addition to real seismic wave data, it also includes simulated synthetic seismic waves. This effectively improves the accuracy and reliability of the generalized neural network monitoring model, which can then be directly applied to earthquake monitoring in any area of ​​the same range, truly realizing the generalization application of the model.

[0026] The technical approach provided by the present invention is not limited to earthquake detection as its only application. That is, the generalized neural network monitoring model in step 4 can also be applied to earthquake location estimation, earthquake magnitude estimation, etc., and can be directly applied to earthquake monitoring and early warning systems.

[0027] Further optionally, the constructed neural network includes at least a neural network for earthquake detection, and the generalized neural network monitoring model includes at least an earthquake detection model;

[0028] The input to the earthquake detection neural network is the three-component waveform of seismic wave data from each station in the training sample set. The input size is represented as: N1*T1*3, where N1 is the maximum number of stations set for the target monitoring range, and zero padding is used when the number of stations in the training sample set is insufficient; T1 is the maximum length of the time sample; and 3 represents 3 channels, which corresponds to the three-component waveform.

[0029] The output of the earthquake detection neural network is labeled with a Gaussian distribution, and the peak value of the output corresponds to the arrival time of the first triggered P phase. If the input of the earthquake detection neural network contains only noise, the output is labeled with zero.

[0030] Further optionally, the constructed neural network includes at least a neural network for earthquake location, and the generalized neural network monitoring model includes at least an earthquake location model;

[0031] The input to the earthquake location neural network consists of the x and y coordinates of the stations and the three-component waveforms of the stations. The stations are input sequentially according to a preset sorting rule. The input size is represented as N1*T1*Q, where N1 is the maximum number of stations set within the target monitoring range, and zero padding is used when the number of stations in the training sample set is insufficient; T1 is the maximum length of the time sample; and Q represents the number of channels. For example, out of 10 channels, 5 channels contain the three-component waveforms and the station x and y coordinates sorted according to the station x coordinates, and the other 5 channels contain the three-component waveforms and the station x and y coordinates sorted according to the station y coordinates.

[0032] The output of the neural network for earthquake location is labeled with a 3D Gaussian distribution, and the peak value of the output indicates the predicted location.

[0033] During training, the technical solution of this invention sorts the stations according to their x or y coordinates or other fixed rules. Traditional methods consider the arbitrary station order as their advantage, while this invention sorts each station according to fixed rules in the input of the neural network, which can reduce the training difficulty of the model. If the station order is arbitrary, it means that the data from multiple stations of an earthquake will have many different inputs to the neural network depending on the different station orders, while the output of the neural network (earthquake location or magnitude) is only one. Therefore, it increases the model space in the training process and makes the training more difficult.

[0034] It should be understood that the above example uses the x or y coordinate arrangement of the stations. In other feasible embodiments, if other fixed rules are used for sorting, the input channels will be adjusted accordingly.

[0035] Further optionally, the constructed neural network includes at least a neural network for earthquake magnitude estimation, and the generalized neural network monitoring model includes at least an earthquake magnitude estimation model;

[0036] The input to the neural network for earthquake magnitude estimation includes the three-component waveforms of the station and the epicentral distance. The input size is represented as T1*4, where T1 is the maximum length of the time sample; 4 represents 4 channels, corresponding to the three-component waveforms and the epicentral distance.

[0037] The output of the neural network for earthquake magnitude estimation is labeled with a 1D Gaussian distribution. The peak value of the output represents the normalized magnitude Mr of the station. The normalized magnitude Mr satisfies: Mr = M - log(A), where M is the true magnitude and A is the amplitude of the station waveform.

[0038] The magnitude of an earthquake is determined based on the magnitude of each monitoring station.

[0039] The technical solution of this invention can extract data features from continuous seismic data in real time, enabling continuous earthquake monitoring. Even when only a portion of the stations trigger the seismic activity, the neural network can initiate monitoring, detecting earthquakes and estimating their location and magnitude from the limited waveform information available. Figure 1 and 2 (As shown).

[0040] This invention proposes to use single-station data training for magnitude estimation and defines a standardized magnitude Mr, which is the magnitude after removing the influence of the maximum amplitude of the input time window. For n stations of an earthquake, n magnitudes can be output, and the average or median of the magnitude of the earthquake can be obtained. While standardized magnitude was mentioned in the paper by Zhang et al. (2021), their method used data from multiple stations as input, with the maximum amplitude referring to the maximum amplitude of the waveforms from multiple stations, and the output being the magnitude of an earthquake. The drawback of this method is that if a station has abnormally large amplitude noise, the magnitude (Mr) will be completely incorrect, even if the data from other stations is normal, resulting in a very large estimation error. On the other hand, unlike Zhang et al. (2021), this invention considers epicentral distance information in magnitude estimation, while their method directly uses waveforms from multiple stations as input without any station information. Magnitude and epicentral distance are directly related; a large epicentral distance results in a weak amplitude, and a small epicentral distance results in a strong amplitude. Therefore, magnitude is a quantity related to both amplitude and epicentral distance. This invention's technical solution considers the epicentral distance as an input channel for the neural network, which can greatly improve prediction accuracy.

[0041] Furthermore, the technical solution of this invention proposes to utilize multiple seismic data inputs and label the fastest P-wave arrival times of multiple stations using a one-dimensional Gaussian distribution. Traditional seismic detection neural networks mainly use single-seismic data as input, while the method using multiple inputs outputs a binary classification scheme labeled 0 and 1, where 0 and 1 respectively represent whether the input time window contains a seismic event. This invention uses a Gaussian distribution to represent the fastest P-wave arrival time, which has the advantage of simultaneously detecting earthquakes and outputting the time information of the fastest P-wave within the time window. This solution is crucial for solving the problem of continuous earthquake monitoring starting from a small number of triggering stations, because at the triggering time of a small number of stations, the current time window may be mostly noise. Traditional methods might consider the time window as noise, while the monitoring solution of this invention requires identifying the time window as a seismic signal and outputting the fastest P-wave arrival time. Using the fastest P-wave arrival time and the location result of the localization neural network, the travel time of the station can be calculated. Subtracting the travel time from the fastest P-wave arrival time allows for the estimation of the seismic source's origin time, enabling earthquake monitoring to begin from a very small number of triggering stations.

[0042] Further optionally, the normalized magnitude is: before obtaining the simulated and synthesized seismic wave data within the target monitoring range using the seismic wave data from each station after rotation, it is also necessary to adjust the amplitude of seismic waves from different sources to the same magnitude.

[0043] Since the seismic waveforms from different stations A, B, C, and D originate from different seismic sources, it is necessary to adjust the amplitude of these waveforms to a given magnitude M. The adjustment method is as follows: First, adjust the amplitude of the rotated waveform (d′) to a value of M. x ,d′ y ,d′ z At the same time, divide by their maximum value to normalize, and then use the following formula to obtain the earthquake waveform with magnitude M:

[0044]

[0045] δ=M-1.110log(r / 100)-0.00189(r-100)-3

[0046] Where δ is the ML magnitude formula, but it can also be other magnitude formulas. This is the composite waveform with magnitude M, where r is defined as the epicentral distance.

[0047] In three aspects, the present invention also provides a system based on the above-mentioned earthquake monitoring method, comprising:

[0048] The basic database construction module is used to collect seismic wave data from various regions and group them according to epicentral distance and depth to build the basic database. Seismic wave data with the same epicentral distance and depth are grouped together.

[0049] The preprocessing module is used to randomly generate the distribution of stations within the target monitoring range, randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station.

[0050] The sample set construction module is used to generate a training sample set for the target monitoring range using a reconstruction method of station seismic maps;

[0051] Specifically, based on the epicentral distance of the stations and the depth of the synthesized earthquake, seismic wave data with the same epicentral distance and depth are selected from the basic database for each station, and then rotated to the azimuth and normalized magnitude according to the rotation formula; the simulated synthesized seismic wave data within the target monitoring range is obtained using the rotated and normalized seismic wave data of each station, as training samples.

[0052] The monitoring model construction module is used to build a neural network architecture for earthquake monitoring. The neural network is trained using the training sample set to obtain a generalized neural network monitoring model. Based on the generalized neural network monitoring model, earthquake monitoring is carried out in any region with the same range as the target monitoring range.

[0053] In four aspects, the present invention also provides an electronic terminal, comprising at least:

[0054] One or more processors;

[0055] And a memory that stores one or more computer programs;

[0056] The processor invokes the computer program to execute:

[0057] The steps of the sample set construction method or the steps of the earthquake monitoring method based on generalized neural networks.

[0058] Fifthly, the present invention also provides a computer-readable storage medium, characterized in that: it stores a computer program, which is invoked by a processor to execute: the steps of the sample set construction method or the steps of the earthquake monitoring method based on a generalized neural network.

[0059] Beneficial effects

[0060] Compared with existing methods, the sample set construction method provided by the present invention is a method for recombining station seismic maps, generating a large number of simulated seismic events that may occur at any location within the monitoring area, with arbitrary station distribution. Therefore, when constructing the sample set using this principle, seismic wave data from different regions and with different station distributions can be fully utilized. On the one hand, this achieves data compatibility and ensures sample generalization, especially providing feasible technical support for constructing a generalized neural network monitoring model. On the other hand, it expands the sample set and contributes to improving model accuracy.

[0061] Based on the provided method for reconstructing seismic maps from existing stations, this invention provides a seismic monitoring method based on a generalized neural network. This method constructs a generalized neural network monitoring model, which can then be directly applied to seismic monitoring in any area within the same region, unlike traditional methods that require retraining when deployed to new areas. Specifically, this invention randomly deploys stations and randomly generates synthetic earthquake locations within a set target monitoring range, enabling the simulated seismic wave data to have generalizability. The generalized neural network monitoring model constructed according to this invention can achieve multiple applications, including earthquake detection, earthquake location estimation, and earthquake magnitude estimation, making it more versatile.

[0062] The present invention further optimizes the technical solution by proposing a magnitude estimation method that uses single-station data training. This means that each station obtains its corresponding magnitude, and for n stations recording an earthquake, n magnitudes can be output. The average or median of these magnitudes is then used to determine the earthquake's magnitude. Compared to existing technologies, the magnitude determined by this invention has higher accuracy and reliability. Furthermore, this invention incorporates the epicentral distance into the magnitude estimation model as one of the input quantities. This fully considers the direct relationship between magnitude and epicentral distance; a large epicentral distance results in a weak amplitude, and a small epicentral distance results in a strong amplitude. Therefore, magnitude is a quantity related to both amplitude and epicentral distance. The present invention's technical solution considers the epicentral distance as an input channel for the neural network, which can greatly improve prediction accuracy.

[0063] The technical solution of this invention is further optimized by setting the output of the generalized neural network monitoring model and using Gaussian distribution labeling. In particular, for the earthquake detection model, the arrival time of the fastest P-wave at multiple stations is labeled using a one-dimensional Gaussian distribution. While detecting earthquakes, the time information of the fastest P-wave in the time window can be output, laying the foundation for monitoring earthquakes from the triggering of a very small number of stations. That is, by combining the earthquake detection model and the earthquake location model, the travel time of the station can be calculated using the arrival time of the fastest P-wave and the position results of the location neural network. The origin time of the earthquake source can be estimated by subtracting the travel time from the arrival time of the fastest P-wave. This enables the detection of earthquake signals when triggered by a few stations and the real-time and continuous output of earthquake location, magnitude, and origin time. Attached Figure Description

[0064] Figure 1 This invention provides a generalized neural network model for the earthquake monitoring method and a schematic diagram of the sample generation principle.

[0065] Figure 2 This is a schematic diagram of data processing for the earthquake detection and location model provided by the present invention. Detailed Implementation

[0066] This invention provides a sample set construction method and system for earthquake monitoring based on a generalized neural network. Its core principle is to leverage the key characteristic that earthquake waveforms with the same epicentral distance and depth across various regions globally share the same P-wave or S-wave travel times, demonstrating significant similarity in phase travel. This leads to a method for recombining station seismograms. Using seismic wave data from various regions worldwide, it simulates earthquakes at different locations and with different monitoring station distributions. These earthquakes exhibit generalization, meaning that the actual earthquake data generated under any monitoring scheme in any region will show similarities to these generalized earthquakes. Therefore, the sample set construction method provided by this invention can offer generalized simulated earthquake seismic wave data for various regions, enriching the sample library. Furthermore, for a defined target monitoring range, generalized training samples can be generated to construct a generalized neural network model applicable to earthquake monitoring within the same target monitoring range in any region, unconstrained by station distribution. For example, the basic database includes 104,808 single-station seismograms from real earthquakes occurring in Italy, Oklahoma, and Southern California. These seismic maps were recombined to generate 355,001 simulated seismic events ranging from 50 to 100 kilometers for training. The trained generalized neural network can be applied to different regions, and seismic parameters can be determined within 4 seconds of the first station triggering the earthquake. The results demonstrate that the generalized neural network is capable of real-time earthquake monitoring and can be deployed in early warning systems without additional training and complex empirical setup.

[0067] The present invention will be further described below with reference to embodiments.

[0068] Example 1:

[0069] This embodiment provides a sample set construction method applied to earthquake monitoring, which includes the following steps:

[0070] S1: Collect seismic wave data from various regions and group them according to epicentral distance and depth to obtain a basic database. Seismic wave data with the same epicentral distance and depth are grouped into the same group.

[0071] S2: Obtain the distribution of stations within the target monitoring range, randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station;

[0072] S3: Based on the epicentral distance of the station and the depth of the composite earthquake, select seismic wave data with the same epicentral distance and depth for each station from the basic database, and then rotate and normalize the magnitude according to the azimuth rotation formula.

[0073] S4: Using the seismic wave data from each station after rotation and normalization, simulated and synthesized seismic wave data within the target monitoring range are obtained, and used as samples to expand the seismic wave sample set within the target monitoring range.

[0074] Example 2:

[0075] This embodiment provides an earthquake monitoring method based on a generalized neural network, including:

[0076] Step 1: Collect seismic wave data from various regions and group them according to epicentral distance and depth to obtain a basic database. Seismic wave data with the same epicentral distance and depth are grouped together.

[0077] Step 2: Randomly generate the distribution of stations within the target monitoring range, and randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station;

[0078] Step 3: Generate a training sample set for the target monitoring range using the method of reconstructing seismic maps from stations;

[0079] Specifically, based on the epicentral distance of the stations and the depth of the synthesized earthquake, seismic wave data with the same epicentral distance and depth are selected from the basic database for each station, and then rotated and normalized according to the azimuth rotation formula; using the rotated and normalized seismic wave data of each station, simulated and synthesized seismic wave data within the target monitoring range is obtained as training samples.

[0080] Step 4: Build a neural network architecture for earthquake monitoring, train the neural network using the training sample set to obtain a generalized neural network monitoring model, and realize earthquake monitoring in any region with the same range as the target monitoring range based on the generalized neural network monitoring model.

[0081] Assume that step S1 of this embodiment and the basic database in step 1 include 104,808 single-station seismic maps of real earthquakes occurring in Italy, Oklahoma, and Southern California. These seismic maps are recombined to generate 355,001 simulated earthquake events for training, ranging from 50 km to 100 km.

[0082] Then, in steps S2 and 2, the distribution of stations within the target monitoring range is randomly generated, and the locations of synthetic earthquakes are randomly generated. The epicentral distance *r* from the synthetic earthquakes to each station is calculated. This embodiment assumes that the synthetic station distribution is within the range of 0-82 km in the x-direction and 0-100 km in the y-direction, while the earthquake occurs within the monitoring area within the range of 16-66 km in the x-direction, 0-100 km in the y-direction, and 0-23 km in the z-direction. To generate training samples, this invention randomly selects earthquake locations (sx, sy, sz) within the earthquake range and randomly selects 4-12 station locations within the station range.

[0083] Steps S3 and 3 both involve randomly selecting, from the grouped basic dataset, the three-component waveforms of the synthetic station corresponding to the epicentral distance r and depth sz (the three-component seismogram represents the vibrations in the x, y, and z directions, i.e., the three components of the particle vibration velocity vector, used to simultaneously record P-waves, S-waves, and converted waves) for each station. Figure 1 A shows a typical simulated training earthquake monitored by four stations, with waveforms derived from two different real earthquakes in the base dataset. The E and N components of the waveforms are rotated to a new azimuth direction centered on the simulated earthquake location, enabling the invention to use phase azimuth to constrain earthquake locations.

[0084] Among them, such as Figure 1 As shown in Figure A, assuming two earthquakes in arbitrary regions are received by four arbitrarily distributed seismic stations, this invention extracts the waveforms corresponding to stations A, B, C, and D from these seismic records. These waveforms can be combined to form a new synthetic seismic dataset. The epicentral distances of the four stations in the synthetic seismic dataset are the corresponding epicentral distances of stations A, B, C, and D. The waveforms only need to be rotated to the corresponding azimuth angles.

[0085] Randomly generate a synthetic earthquake location coordinate (s) x ,s y ,s z Then, randomly generate the coordinates of n monitoring stations. Calculate the epicentral distance from the composite earthquake to n stations, based on the value of each epicentral distance and depth s. z The value is obtained by retrieving a three-component waveform (d) for that station from the basic database. x ,d y ,d z The three-component waveform of the seismic wave data of the i-th station after azimuth rotation is represented as (d′). x ,d′ y ,d z By performing the same processing on each station, waveform data from n stations can be obtained. After normalizing the magnitude, a simulated synthetic seismic wave data is obtained, which is used for training the neural network.

[0086] The formula for azimuth rotation is as follows:

[0087]

[0088]

[0089] In the formula, d′ x and d′ y The waveforms of the x and y components in the rotated seismic wave data, d x and d y The waveforms of the x and y components in the original seismic wave data. For the azimuth difference, satisfying The azimuth angle of the i-th station corresponding to the synthetic earthquake is the azimuth angle of the original actual earthquake station corresponding to the seismic wave data retrieved from the basic database.

[0090] To further address the issue of waveform amplitude from different earthquake stations, this invention rescales the waveforms from all stations to a specific magnitude by multiplying the waveforms by an amplitude factor, ensuring that the waveform amplitude is the same as that of a magnitude 3 earthquake. Of course, scaling to other magnitudes is also possible. Details are as follows:

[0091] Since waveforms A, B, C, and D originate from different seismic sources, their amplitudes need to be adjusted to a given magnitude M. The adjustment method is as follows: First, adjust the amplitude of the rotated waveform (d′) to a given magnitude M. x ,d′ y ,d′ z At the same time, divide by their maximum value to normalize, and then use the following formula to obtain the earthquake waveform with magnitude M:

[0092]

[0093] δ=M-1.110log(r / 100)-0.00189(r-100)-3

[0094] Where δ is the ML magnitude formula, but it can also be other magnitude formulas. This is the composite waveform with magnitude M, where r is the epicentral distance.

[0095] Therefore, it should be understood that, following the above approach, simulated synthetic seismic wave data can be constructed as training set samples. By randomly generating station distributions and seismic event locations within the target monitoring range, and constructing training samples according to the above approach, the training set samples can possess generalization capabilities. It should be understood that, in addition to simulated synthetic seismic wave data, the training set samples can also include actual seismic wave data.

[0096] It should also be noted that the above example is based on the construction of a generalized training set sample, so the station distribution and the selection of earthquake time are random. According to the above technical idea, the technical solution of Example 1 is also applicable to generating simulated and synthesized seismic wave data for a specific monitoring range and a specific station distribution, in order to expand the sample library.

[0097] Regarding the neural network architecture for earthquake monitoring constructed in step S4, this embodiment does not impose constraints on its architecture or content. It should be understood that, following the technical ideas of this invention, after constructing the training set samples, the neural network for earthquake monitoring can meet the requirements of this invention and fall within the protection scope of this invention.

[0098] Example 3:

[0099] Based on Embodiments 1 and 2, this embodiment preferentially constructs a generalized neural network model that includes earthquake detection, earthquake location, and earthquake magnitude estimation models. In other feasible embodiments, implementing one or more of these functions also falls within the scope of protection of this invention.

[0100] This embodiment designs three neural networks, taking a fully convolutional neural network as an example. Detailed configuration is as follows: Figure 1 As shown, earthquake data is continuously fed into the network model for earthquake detection, location, and magnitude estimation. Figure 1 Earthquakes are detected and located simultaneously using waveforms from multiple stations (12 stations), and then a neural network estimates the magnitude of the earthquake at each station when an earthquake is detected.

[0101] Earthquake detection model:

[0102] like Figure 1 As shown, a 2D convolutional layer is used. The neural network input of the seismic detection model includes three components of waveform data from multiple stations. The input size (12*1024*3) in this embodiment is determined by the maximum number of stations, the maximum length of the time samples, and the number of components. Specifically, the total waveform length is 30 seconds, the time interval is 0.05 seconds, and there are a total of 600 time sampling points. The remaining 424 time sampling points are padded with zeros. In this application, the number of stations can vary (less than 12), and any points exceeding this number are also padded with zeros. The output of the neural network is labeled with a Gaussian distribution, with the peak value corresponding to the arrival time of the first triggered P-phase (if a seismic event is detected); if the input contains only noise, the output is labeled with zeros.

[0103] Therefore, following the above logic, for the neural network model of earthquake detection, the input consists of three-component waveforms from multiple stations. This invention first obtains the arrival time of the fastest P-wave among these station waveforms, and then uses a Gaussian probability distribution to label the arrival time of the P-wave. The x-coordinate corresponding to the maximum value of the Gaussian probability distribution is the arrival time of the P-wave. Using the labeled samples, the neural network model can be trained to obtain an earthquake detection model with generalization capabilities.

[0104] Earthquake location model: such as Figure 1 As shown, a 2D convolutional layer is used. The neural network input of the seismic location model includes the x and y coordinates of the stations and the corresponding waveform data, with a total size of 12*1024*10, where 12 represents the number of stations, 1024 represents the number of waveform sampling points, and 10 represents the number of channels. The station x and y coordinates are normalized to the range of 0 to 1 according to the station range (0-82 km and 0-100 km). These normalized coordinates are represented as two vectors, each with a length of 1024 samples, occupying two channels, and are used as input along with the corresponding waveform data. Unlike graph neural network methods, this invention utilizes the in-phase axis characteristics of the waveforms and organizes the input data by sorting the stations in ascending order of their x and y coordinates. Therefore, the total input size is as follows: the first 5 channels contain three waveform components and station locations sorted by x-coordinate, resulting in a sorted set of 12*1024*5 data. The remaining 5 channels contain the same data sorted by y-coordinate, also resulting in a set of 12*1024*5 data. These two sets of data are combined to form a 12*1024*10 set, which serves as the input to the neural network. This arrangement ensures that the neural network input is unique (12*1024*10) for any earthquake monitored by a given station. The output of the location neural network is labeled using a 3D Gaussian distribution, with peak values ​​representing the predicted location. The grid size corresponds to a monitoring range of 16-66 km in the x-direction, 0-100 km in the y-direction, and a depth of -6-22.8 km.

[0105] It should be noted that this embodiment sorts based on x and y coordinates. In other feasible embodiments, while ensuring that the stations of the present invention are input in a sorted manner, other sorting rules can also be set, such as sorting by the size of x*x+y*y.

[0106] Earthquake magnitude estimation model:

[0107] This invention calculates the magnitude for each station separately. The neural network input size for the earthquake magnitude estimation model is 1024*4. The first three channels contain waveform data, and the fourth channel represents the epicentral distance. Similar to the input of the detection and localization neural network, a 30-second earthquake waveform occupies 600 time sampling points, with any remaining 424 time sampling points padded with zeros. Furthermore, as... Figure 1B. The epicentral distance is normalized to the range of 0-1, corresponding to a distance of 0 to 110 kilometers; that is, if the epicentral distance range is 0-110 km, the normalized epicentral distance is obtained by dividing by 110, and together with the three-component waveform of the station, it is used as the four channels input to the neural network model. During the training process, this invention randomly truncates the waveform starting from the arrival time of the P wave to ensure that the final model can predict the magnitude using only a small amount of seismic signal within the input time window. The output of the magnitude model is labeled with a 1D Gaussian distribution, where the peak value represents the normalized magnitude Mr. The normalized magnitude Mr satisfies: Mr = M - log(A), where M is the true magnitude and A is the amplitude of the station waveform. The normalized magnitude Mr is labeled using a Gaussian probability distribution, and the position corresponding to the maximum value of the Gaussian distribution is the magnitude of Mr (e.g., ...). Figure 1 (As shown in B). When inputting a waveform, it is first normalized to obtain its amplitude A. The normalized waveform is then input into a neural network model to predict its standardized magnitude Mr. After obtaining Mr, the true magnitude M can be obtained using Mr + log(A). The advantage of defining the magnitude in this way is that during neural network training, the waveform can be normalized, and the neural network can learn magnitude characteristics other than amplitude influence.

[0108] This standardization is defined by subtracting the logarithm of the maximum amplitude from the Richter magnitude (ML), thus excluding the contribution of waveform amplitude (Zhang et al., 2021). The relationship between waveform amplitude and magnitude is well-defined: a tenfold increase in amplitude corresponds to a one-unit increase in magnitude. The final prediction, Mr magnitude, from the neural network must be supplemented with the logarithm of the waveform amplitude to obtain the final ML magnitude.

[0109] Regarding the neural network architecture selected in this embodiment, as follows: Figure 1 As shown, the neural network architecture mainly employs convolutional layers, max-pooling layers, and upsampling layers. The detection and localization neural network design includes 2D convolutional layers to process input data from multiple stations. The magnitude network consists of 1D convolutional layers to process data from a single station. The kernel sizes of both the 2D and 1D convolutional layers are uniformly set to 3*3 and 3, respectively. Zero padding is applied to the output of each convolutional layer to maintain a constant output size. Max-pooling layers are used to extract key features constraining the seismic parameters, while upsampling layers are used to adjust the final output size of the network model. Compared to localization and magnitude estimation, earthquake detection is a relatively simple task. Therefore, a fully convolutional neural network is sufficient to achieve the goal of earthquake detection. To alleviate the gradient vanishing problem, this invention introduces multiple replication layers in the localization and magnitude neural networks, such as... Figure 1 As shown.

[0110] It should be noted that, as Figure 1The network architecture shown is an example of this embodiment. Without departing from the technical idea of ​​constructing training samples using the recombination method and setting the input and output of the neural network, any neural network architecture used to implement earthquake monitoring is in line with the requirements of this invention and falls within the protection scope of this invention.

[0111] In this embodiment, the three models described above are trained using the Adam algorithm with a learning rate of 10 to the power of -4. Furthermore, 2, 4, and 2 Dropout layers are respectively incorporated into the detection, localization, and magnitude networks (Abadi et al., 2016). The trained detection and localization network models are merged into a single file, enabling simultaneous detection and localization. When the merged neural network detects an earthquake event, the corresponding theoretical time of arrival (P) is calculated. The triggering station is determined based on whether the P arrival time falls within the monitoring time window. The waveform data from the triggering station is then fed into the magnitude network to estimate the magnitude; the final result is the average of all triggering stations. Figure 2 The diagram shows the data processing of the earthquake detection and location model. (a) is an example of the input waveform, showing only the Z component waveform. The input includes the waveform and station location information. Detection and location can start at the 2nd second after triggering, and the neural network runs once at each moment.

[0112] In summary, the technical solution of this invention utilizes station data from anywhere in the world, and generates simulated earthquake samples monitored by any station at any location by combining data from different stations. For example... Figure 1As shown, the collected station waveforms are first organized into a basic database. Given an epicentral distance and focal depth, a set of corresponding station waveforms can be obtained from the basic database. The figure shows an example of earthquake sample generation. An earthquake location (pentagram) and four station locations (triangles) are randomly generated in the study area. Based on this, the epicentral distances from the four stations to the epicenter can be calculated. Based on each epicentral distance and focal depth value, a corresponding station waveform can be found from the basic database. This waveform can be used as the composite station waveform. The waveforms of the four stations in the figure are from two stations of two different earthquakes. Three neural networks were designed as network models for earthquake detection, earthquake location, and magnitude estimation, respectively. The first is the earthquake detection network model. The input is the waveforms monitored by multiple stations corresponding to a certain earthquake. This set of waveforms has the arrival time of the fastest P-wave phase, which is the P-wave corresponding to the first triggered station. The output of the neural network is labeled as a Gaussian probability density distribution. The location of the highest point of the Gaussian probability density distribution is the arrival time of the P-wave. If the input is completely noise, the output probability density distribution is zero. The second network model is for earthquake location. The input includes the x and y coordinates of the stations and their corresponding three-component waveforms. These stations are sorted by their x or y magnitudes. This sorting has the advantage that, given an earthquake, the waveform input is arranged in a specific way, making neural network training easier. The output is a Gaussian probability density distribution, with the point of maximum probability density indicating the earthquake location. The third network model is for earthquake magnitude calculation. The input includes the three-component waveforms and the epicentral distance. The output is a Gaussian probability density distribution, with the location of the maximum probability density value corresponding to the standardized magnitude. This standardized magnitude is defined as the true magnitude minus the logarithm of the waveform amplitude.

[0113] Example 4:

[0114] In some embodiments, the present invention also provides a system based on a sample set construction method, comprising:

[0115] The basic database construction module is used to collect seismic wave data from various regions and group them according to epicentral distance and depth to obtain the basic database. Seismic wave data with the same epicentral distance and depth are grouped together.

[0116] The preprocessing module is used to obtain the distribution of stations within the target monitoring range, randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station.

[0117] The data extraction module is used to select seismic wave data with the same epicentral distance and depth for each station from the basic database based on the epicentral distance of the station and the depth of the synthetic earthquake, and then rotate and normalize the magnitude according to the azimuth rotation formula.

[0118] The recombination module is used to obtain simulated and synthesized seismic wave data within the target monitoring range using the rotated and normalized seismic wave data of each station, and to expand the seismic wave sample set within the target monitoring range as a sample.

[0119] Example 5:

[0120] In some embodiments, the present invention also provides a monitoring system based on an earthquake monitoring method, comprising:

[0121] The basic database construction module is used to collect seismic wave data from various regions and group them according to epicentral distance and depth to build the basic database. Seismic wave data with the same epicentral distance and depth are grouped together.

[0122] The preprocessing module is used to randomly generate the distribution of stations within the target monitoring range, randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station.

[0123] The sample set construction module is used to generate a training sample set for the target monitoring range using a reconstruction method of station seismic maps;

[0124] Specifically, based on the epicentral distance of the stations and the depth of the synthesized earthquake, seismic wave data with the same epicentral distance and depth are selected from the basic database for each station, and then rotated to the azimuth and normalized magnitude according to the rotation formula; the simulated synthesized seismic wave data within the target monitoring range is obtained using the rotated and normalized seismic wave data of each station, as training samples.

[0125] The monitoring model construction module is used to build a neural network architecture for earthquake monitoring. The neural network is trained using the training sample set to obtain a generalized neural network monitoring model. Based on the generalized neural network monitoring model, earthquake monitoring is carried out in any region with the same range as the target monitoring range.

[0126] It should be understood that the implementation process of each module in Embodiments 4 and 5 can be described with reference to the content of the aforementioned method. The above division of functional modules is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. At the same time, the above-mentioned integrated units can be implemented in hardware or in the form of software functional units.

[0127] Example 6:

[0128] In some embodiments, the present invention provides an electronic terminal comprising at least: one or more processors; and a memory storing one or more computer programs; wherein the processor invokes the computer programs to execute:

[0129] The steps of the sample set construction method or the steps of the earthquake monitoring method based on generalized neural networks are described above. For specific implementation details, please refer to the detailed descriptions in Embodiments 1 and 2.

[0130] The memory may include high-speed RAM, and may also include a non-volatile defibrillator, such as at least one disk storage device.

[0131] If the memory and processor are implemented independently, they can be interconnected via a bus to communicate with each other. This bus can be an industry-standard architecture bus, an external device interconnect bus, or an extended industry-standard architecture bus, etc. The bus can be categorized as an address bus, data bus, control bus, etc.

[0132] Optionally, in a specific implementation, if the memory and processor are integrated on a single chip, the memory and processor can communicate with each other through an internal interface.

[0133] It should be understood that, in the embodiments of the present invention, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

[0134] Example 7:

[0135] In some embodiments, the present invention provides a computer-readable storage medium storing a computer program that is invoked by a processor to execute: the steps of the sample set construction method or the steps of the earthquake monitoring method based on a generalized neural network. Specific implementation processes can be found in the detailed descriptions of Embodiments 1 and 2.

[0136] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the controller described in any of the foregoing embodiments, such as the controller's hard drive or memory. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both the controller's internal storage unit and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0137] Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.

Claims

1. A method for constructing a sample set, applied to earthquake monitoring, characterized in that: Includes the following steps: Seismic wave data from various regions were collected and grouped according to epicentral distance and depth to obtain a basic database. Seismic wave data with the same epicentral distance and depth were grouped together. Obtain the distribution of stations within the target monitoring range, randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station; Based on the epicentral distance of the stations and the depth of the composite earthquake, seismic wave data with the same epicentral distance and depth are selected from the basic database for each station, and then the magnitude is rotated and normalized according to the azimuth rotation formula. By using the seismic wave data from each station after rotation and normalization, simulated and synthesized seismic wave data within the target monitoring range are obtained, and these data are used as samples to expand the seismic wave sample set within the target monitoring range.

2. The sample set construction method according to claim 1, characterized in that: The azimuth rotation formula is as follows: ； ； In the formula, and The waveforms of the x and y components in the rotated seismic wave data. and The waveforms of the x and y components in the original seismic wave data. For the azimuth difference, satisfying , For the corresponding synthetic earthquake i The azimuth of each station The azimuth angle of the original actual seismic station corresponding to the seismic wave data retrieved from the basic database is given by acosine function. The synthetic seismic locations within the target monitoring range are represented as follows: , No. i The coordinates of each station are represented as follows: , Let x and y be the coordinates of the synthesized earthquake. The depth of the synthesized earthquake. For the first i The x and y coordinates of each station, and the three-component waveform of the seismic wave data retrieved from the basic database, are represented as (d x ,d y ,d z After rotating the azimuth angle, the first... i The three-component waveform of the seismic wave data from each station is represented as follows: .

3. An earthquake monitoring method based on a generalized neural network, characterized in that: include: Step 1: Collect seismic wave data from various regions and group them according to epicentral distance and depth to obtain a basic database. Seismic wave data with the same epicentral distance and depth are grouped together. Step 2: Randomly generate the distribution of stations within the target monitoring range, and randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station; Step 3: Generate a training sample set for the target monitoring range using the method of reconstructing seismic maps from stations; Specifically, based on the epicentral distance of the stations and the depth of the synthesized earthquake, seismic wave data with the same epicentral distance and depth are selected from the basic database for each station, and then rotated to the azimuth and normalized magnitude according to the rotation formula; the simulated synthesized seismic wave data within the target monitoring range is obtained using the rotated and normalized seismic wave data of each station, as training samples. Step 4: Build a neural network architecture for earthquake monitoring, train the neural network using the training sample set to obtain a generalized neural network monitoring model, and realize earthquake monitoring in any region with the same range as the target monitoring range based on the generalized neural network monitoring model.

4. The earthquake monitoring method based on a generalized neural network according to claim 3, characterized in that: The constructed neural network includes at least a neural network for earthquake detection, and the generalized neural network monitoring model includes at least an earthquake monitoring model. The input to the earthquake detection neural network is the three-component waveform of seismic wave data from each station in the training sample set. The input size is represented as: N1*T1*3, where N1 is the maximum number of stations set for the target monitoring range, and zero padding is used when the number of stations in the training sample set is insufficient; T1 is the maximum length of the time sample; and 3 represents 3 channels, which corresponds to the three-component waveform. The output of the earthquake detection neural network is labeled with a Gaussian distribution, and the peak value of the output corresponds to the arrival time of the first triggered P phase. If the input of the earthquake detection neural network contains only noise, the output is labeled with zero.

5. The earthquake monitoring method based on a generalized neural network according to claim 3, characterized in that: The constructed neural network includes at least a neural network for earthquake location, and the generalized neural network monitoring model includes at least an earthquake location model; The input to the earthquake location neural network consists of the xy coordinates of the stations and the three-component waveforms of the stations. The stations are sorted and input sequentially according to a preset sorting rule. The input size is represented as N1*T1*Q, where N1 is the maximum number of stations set for the target monitoring range, and zero padding is used when the number of stations in the training sample set is insufficient; T1 is the maximum length of the time sample; and Q represents the number of channels. The output of the neural network for earthquake location is labeled with a 3D Gaussian distribution, and the peak value of the output represents the predicted earthquake location.

6. The earthquake monitoring method based on a generalized neural network according to claim 3, characterized in that: The constructed neural network includes at least a neural network for earthquake magnitude estimation, and the generalized neural network monitoring model includes at least an earthquake magnitude estimation model. The input to the neural network for earthquake magnitude estimation includes the three-component waveforms of the station and the epicentral distance. The input size is represented as T1*4, where T1 is the maximum length of the time sample; 4 represents 4 channels, corresponding to the three-component waveforms and the epicentral distance. The output of the neural network for earthquake magnitude estimation is labeled with a 1D Gaussian distribution. The peak value of the output represents the normalized magnitude Mr of the station. The normalized magnitude Mr satisfies: Mr = M - log(A), where M is the true magnitude and A is the amplitude of the station waveform. The magnitude of an earthquake is determined based on the magnitude of each monitoring station.

7. The earthquake monitoring method based on a generalized neural network according to claim 3, characterized in that: The normalized magnitude is: before obtaining the simulated and synthesized seismic wave data within the target monitoring range using the seismic wave data from each station after rotation, it is also necessary to adjust the amplitude of the seismic waves from different sources to the same magnitude. First, the rotated waveform At the same time, normalization is performed by dividing by their maximum value. After normalization, the earthquake waveform with magnitude M is obtained using the following formula: ； ； in, For the ML magnitude formula, This is the composite waveform with magnitude M, where r is defined as the epicentral distance.

8. A system for implementing the earthquake monitoring method based on a generalized neural network as described in any one of claims 3-7, characterized in that: include: The basic database construction module is used to collect seismic wave data from various regions and group them according to epicentral distance and depth to build the basic database. Seismic wave data with the same epicentral distance and depth are grouped together. The preprocessing module is used to randomly generate the distribution of stations within the target monitoring range, randomly generate the location of synthetic earthquakes, and calculate the epicentral distance of the synthetic earthquakes to each station. The sample set construction module is used to generate a training sample set for the target monitoring range using a reconstruction method of station seismic maps; Specifically, based on the epicentral distance of the stations and the depth of the synthesized earthquake, seismic wave data with the same epicentral distance and depth are selected from the basic database for each station, and then rotated to the azimuth and normalized magnitude according to the rotation formula; the simulated synthesized seismic wave data within the target monitoring range is obtained using the rotated and normalized seismic wave data of each station, as training samples. The monitoring model construction module is used to build a neural network architecture for earthquake monitoring. The neural network is trained using the training sample set to obtain a generalized neural network monitoring model. Based on the generalized neural network monitoring model, earthquake monitoring is carried out in any region with the same range as the target monitoring range.

9. An electronic terminal, characterized in that: At least includes: One or more processors; And a memory that stores one or more computer programs; The processor invokes the computer program to execute: The steps of the sample set construction method according to any one of claims 1-2 or the steps of the earthquake monitoring method based on generalized neural networks according to any one of claims 3-7.

10. A computer-readable storage medium, characterized in that: A computer program is stored, which is invoked by a processor to execute: the steps of the sample set construction method according to any one of claims 1-2 or the steps of the earthquake monitoring method based on a generalized neural network according to any one of claims 3-7.

Images