Training Method for Magnitude Estimation Model, Magnitude Estimation Method and Magnitude Estimation Model

Abstract

The present invention relates to the technical field of earthquake prediction and early warning, and specifically relates to a training method for a magnitude estimation model, a magnitude estimation method, and a magnitude estimation model. The method obtains a sample set, where the sample set includes seismic signals and the true results corresponding to the seismic signals; based on the sample set, a prediction result is obtained through the magnitude estimation model; the error between the prediction result and the true result is calculated, and the error gradient is calculated, where the calculation of the error gradient is obtained by using the backpropagation algorithm; a weight optimizer constructed based on the Adam algorithm is used to update the weights, where the weights are updated based on the negative direction of gradient descent, and the given weight learning rate and regularization parameter of the magnitude estimation model are set to update the weights of the model. By combining Transformer with ResNet and RNN structures and using arrival time and travel time information as a means of data augmentation, the present invention enables the magnitude estimation system to have better accuracy, a wider scope of application, and stronger robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake prediction and early warning, and particularly to a training method for a magnitude estimation model, a magnitude estimation method, and a magnitude estimation model. Background Art

[0002] Earthquake magnitude estimation is usually an important task in seismology. Magnitude is an indicator used to measure the relative size of an earthquake. In traditional seismological methods, magnitude is mainly estimated by the amplitude of seismic waves emitted from the earthquake source. However, since the locations of the earthquake sources of the vast majority of earthquakes are not the same as the locations of the stations responsible for receiving seismic waves, there are often huge differences between the received seismic signals and the actual seismic waves emitted from the earthquake source. To solve such problems, a common method is to construct a velocity model for the received seismic signals and make them match the seismic waves emitted from the earthquake source by appropriately amplifying or reducing the amplitude. However, considering the huge differences in the distances between stations and earthquake sources (i.e., epicentral distances), as well as various interference phenomena caused by the complex conditions of the internal media of the Earth, it is very difficult to design a universal velocity model to match most seismic signals in practical applications. On the other hand, manually setting for different types of earthquakes often requires huge human and time costs. Summary of the Invention

[0003] In order to solve the technical problems existing in the above-mentioned prior art, the present invention provides a training method for a magnitude estimation model, a magnitude estimation method, and a magnitude estimation model.

[0004] To achieve the above object, the embodiments of the present invention provide the following technical solutions:

[0005] In a first aspect, in an embodiment provided by the present invention, a training method for a magnitude estimation model is provided, and the method includes the following steps:

[0006] Obtain a sample set, where the sample set includes seismic signals and corresponding true results of the seismic signals;

[0007] Based on the sample set, obtain a prediction result through the magnitude estimation model;

[0008] Calculate the error between the prediction result and the true result, and calculate the error gradient, where the error gradient is calculated using the backpropagation algorithm;

[0009] Use a weight optimizer constructed based on the Adam algorithm to update the weights, where the weights are updated based on the negative direction of gradient descent, set the weight learning rate and regularization parameter given by the magnitude estimation model, and update the weights of the model.

[0010] As a further solution of the present invention, the error between the calculated prediction result and the true result is calculated by a loss function constructed by the mean square error MSE.

[0011] Second, in another embodiment provided by the present invention, a magnitude estimation method is provided, and the method includes:

[0012] Construct input features from the seismic signals and the corresponding arrival time information and travel time information of the seismic signals in the sample set, and preprocess the input features;

[0013] Based on the preprocessed input features, perform magnitude estimation to obtain a prediction result.

[0014] As a further solution of the present invention, the arrival time information includes the P-wave arrival time and the S-wave arrival time of the seismic signal.

[0015] As a further solution of the present invention, the travel time information includes the P-wave travel time.

[0016] As a further solution of the present invention, the preprocessing of the input features specifically includes: preprocessing the seismic signal X using a residual model, and preprocessing the arrival time information and the travel time information respectively using two linear models.

[0017] As a further solution of the present invention, the performing magnitude estimation based on the preprocessed input features to obtain a prediction result specifically includes:

[0018] Use a magnitude prediction module to perform magnitude estimation, where the magnitude prediction module includes UniMP and a residual model. Among them, a path graph is used as the topological structure of UniMP for information transmission;

[0019] For the input seismic signal d = [d 0 , d 1 , …, d N , regarding each moment as a node, all nodes can be connected to the nodes of the next moment to form a structure of a path graph, and its adjacency matrix can be expressed as

[0020]

[0021] Third, in another embodiment provided by the present invention, a magnitude estimation model is provided, and the model includes a preprocessing module and a magnitude prediction module;

[0022] The preprocessing module is used to construct input features from the seismic signals and the corresponding arrival time information and travel time information of the seismic signals in the sample set, and preprocess the input features;

[0023] The magnitude prediction module is used to perform magnitude estimation based on the preprocessed input features to obtain a prediction result.

[0024] As a further solution of the present invention, the magnitude prediction module is composed of UniM and a residual model.

[0025] The technical solution provided by the present invention has the following beneficial effects:

[0026] The training method, magnitude estimation method and magnitude estimation model of the magnitude estimation model provided by the present invention, by combining Transformer with ResNet and RNN structures, and using arrival time and travel time information as a means of data augmentation, can enable the magnitude estimation system to have better accuracy, a wider scope of application and stronger robustness. For seismic events with a large epicentral distance and far from the station, the present invention can maintain the original estimation performance, thus expanding the applicability. In addition, considering that there are always various natural noises in natural seismic signals, compared with the previous deep learning single-station estimation system, the present invention can show strong anti-interference ability under the interference of different input signal-to-noise ratios, so it has strong practical significance for future extensive seismic exploration.

[0027] These aspects or other aspects of the present invention will be more clearly understood in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other embodiments can be obtained based on these drawings.

[0029] Figure 1 It is a flowchart of the training method of the magnitude estimation model according to an embodiment of the present invention.

[0030] Figure 2 It is a structural diagram of the training device of the magnitude estimation model according to an embodiment of the present invention.

[0031] Figure 3 It is a flowchart of the magnitude estimation method according to an embodiment of the present invention.

[0032] Figure 4 It is a structural diagram of the magnitude estimation model according to an embodiment of the present invention.

[0033] Figure 5Structural diagram of the preprocessing module of the magnitude estimation model according to an embodiment of the present invention.

[0034] Figure 6 Structural diagram of the magnitude prediction module of the magnitude estimation model according to an embodiment of the present invention.

[0035] Figure 7 Position distribution diagram of seismic events and receiving stations of the dataset STEAD according to an embodiment of the present invention.

[0036] Figure 8 For M of the dataset STEAD according to an embodiment of the present invention L And M D Value distribution corresponding to the magnitude.

[0037] Figure 9 Estimation result according to an embodiment of the present invention, where the division ratio of the training set and the test set is 90% and 10%.

[0038] Figure 10 Estimation result according to an embodiment of the present invention, where the division ratio of the training set and the test set is 50% and 50%.

[0039] In the figure: input unit - 100, prediction result acquisition unit - 200, calculation unit - 300, weight adjustment unit - 400, preprocessing module - 500, magnitude prediction module - 600. Detailed implementation manners

[0040] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0041] The flowcharts shown in the accompanying drawings are only illustrative examples, and do not necessarily include all the contents and operations / steps, nor do they necessarily need to be executed in the described order. For example, some operations / steps can also be decomposed, combined or partially merged, so the actual execution order may be changed according to the actual situation.

[0042] It should be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0043] Specifically, the embodiments of the present invention will be further elaborated below in conjunction with the accompanying drawings.

[0044] Please refer to Figure 1 , Figure 1 which is a flowchart of a method for training a magnitude estimation model provided by an embodiment of the present invention. As Figure 1 shown, the method for training the magnitude estimation model includes steps S10 to S40. The magnitude estimation model is jointly built using the unified message passing model UniMP, the image classification network ResNet, and the recurrent neural network RNN.

[0045] S10. Obtain a sample set, where the sample set includes seismic signals and the corresponding true results of the seismic signals;

[0046] The sample set can be obtained at a seismic monitoring station.

[0047] S20. Based on the sample set, obtain a prediction result through the magnitude estimation model;

[0048] Specifically, obtaining the prediction result through the magnitude estimation model includes the following steps:

[0049] S211. Construct input features using the seismic signals in the sample set and the corresponding arrival time information and travel time information of the seismic signals, and preprocess the input features;

[0050] In the embodiment of the present invention, the arrival time information includes the P-wave arrival time and the S-wave arrival time of the seismic signal.

[0051] In the embodiment of the present invention, the travel time information includes the P-wave travel time.

[0052] Further, for an earthquake event E i , its seismic signal x i includes signals from three directions. When we define the P-wave arrival time and the S-wave arrival time of the seismic signal as pa i and sa i respectively, then the arrival time a i = [pa i , sa i can be defined. Considering that the seismic signal is a discrete time series, so pa i , sa i = 0, 1,..., L - 1, where L is the length of the seismic signal. Similarly, the P-wave travel time can be defined as t i . Assuming the number of earthquake events is N, then the seismic signal X = [x 0 , x 1 ,..., x N-1 , A = [a 0 , a 1 ,..., a N-1 and T = [t 0,t 1 ,…,t N-1 can be defined separately.

[0053] In an embodiment of the present invention, the preprocessing of the input features specifically includes: preprocessing the seismic signal X using a residual model, and preprocessing the arrival time information and travel time information respectively using two linear models.

[0054] Furthermore, since the shapes of X, A, and T are different, we constructed two linear models (LMs) and one residual model (RM) based on a neural network to jointly process them, as Figure 5 shown. The two LM models and one RM model together form the "preprocessing module". From the perspective of seismic analysis, the roles of the LM and RM in this architecture can be further explored. When the location of an earthquake source is detected, its magnitude can be calculated using the median of the peak amplitude ratio between this earthquake and some "template" earthquake samples (Peng et al., 2009). Considering the strong feature extraction ability of ResNet, the RM can perform a function similar to "template sample selection". Preprocessing X using the RM can effectively extract those samples whose hidden features best reflect the attributes of the seismic signal. In addition, since the arrival time is another important indicator commonly used in traditional magnitude estimation methods, when performing data augmentation, we applied another LM structure to A, hoping to further improve the expressive ability of MagInfoNet.

[0055] S212. Based on the preprocessed input features, perform magnitude estimation to obtain a prediction result;

[0056] The preprocessed input features are h = [h x , h a , h t . h is the concatenation of h x , h a and h t .

[0057] The performing magnitude estimation based on the preprocessed input features to obtain a prediction result specifically includes:

[0058] Use the Mag-Pred module to perform magnitude estimation, where the Mag-Pred module includes UniMP and a residual model. Among them, a path graph is used as the topological structure of UniMP for information transmission.

[0059] For the input seismic signal d = [d 0 , d 1 , …, d N, if each moment is regarded as a node, then all nodes can be connected to the nodes of the next moment to form a structure of a path graph, and its adjacency matrix can be expressed as

[0060]

[0061] The seismic signal d is the node signal in the seismic signal X.

[0062] When A m is used in the graph neural network UniMP, the value d t of the seismic signal d at time t will be affected by the values at the two timestamps before and after the current moment. Therefore, by adopting A m as the underlying topology structure of the network, the autocorrelation of the seismic signal can be extracted.

[0063] S30. Calculate the error between the predicted result and the true result, and calculate the error gradient;

[0064] In the embodiment of the present invention, the error between the predicted result and the true result is calculated by a loss function constructed by the mean square error MSE.

[0065] In the embodiment of the present invention, the error gradient is calculated using the backpropagation algorithm.

[0066] S40. Set the weight learning rate and regularization parameter given by the magnitude estimation model based on the negative direction of gradient descent, update the weights of the magnitude estimation model, and complete the training of the model.

[0067] By combining Transformer with ResNet and RNN structures and using arrival time and travel time information as a means of data augmentation, the present invention can make the magnitude estimation system have better accuracy, a wider applicable range, and stronger robustness. For seismic events with a large epicentral distance and far from the distance station, the present invention can maintain the original estimation performance, thus expanding the applicability. In addition, considering that there are always various natural noises in natural seismic signals, compared with the previous deep learning single-station estimation system, the present invention can show strong anti-interference ability under the interference of different input signal-to-noise ratios, so it has strong practical significance for future extensive seismic exploration.

[0068] It should be understood that although the above is described in a certain order, these steps are not necessarily executed in the above order successively. Unless there is a clear indication in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, a part of the steps in this embodiment may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0069] In one embodiment, referring to Figure 2 As shown, in the embodiment of the present invention, a training device for a magnitude estimation model is further provided. The device includes an input unit 100, a prediction result acquisition unit 200, a calculation unit 300, and a weight adjustment unit 400.

[0070] The input unit 100 is configured to obtain a sample set, where the sample set includes seismic signals and the corresponding true results of the seismic signals.

[0071] The prediction result acquisition unit 200 is configured to obtain a prediction result through the magnitude estimation model based on the sample set.

[0072] The calculation unit 300 is configured to calculate the error between the first prediction result and the true result, and calculate the error gradient.

[0073] The weight adjustment unit 400 sets the weight learning rate and regularization parameter given by the magnitude estimation model based on the negative direction of gradient descent, and updates the weights of the model.

[0074] In one embodiment, referring to Figure 3 As shown, a flowchart of a magnitude estimation method is as Figure 1 As shown, the magnitude estimation method includes steps S100 to step S200.

[0075] S100. Construct input features with the seismic signals in the sample set and the corresponding arrival time information and travel time information of the seismic signals, and preprocess the input features;

[0076] In the embodiment of the present invention, the arrival time information includes the P-wave arrival time and S-wave arrival time of the seismic signal.

[0077] In the embodiment of the present invention, the travel time information includes the P-wave travel time.

[0078] Further, for an earthquake event E i , its seismic signal x iIncluding signals from three directions. When we define the arrival times of the P-wave and S-wave of the seismic signal as pa i and sa i respectively, then the arrival time a i = [pa i , sa i can be defined. Considering that the seismic signal is a discrete time series, so pa i , sa i = 0, 1, …, L - 1, where L is the length of the seismic signal. Similarly, the P-wave travel time can be defined as t i . Assuming the number of seismic events is N, then the seismic signal X = [x 0 , x 1 , …, x N-1 , A = [a 0 , a 1 , …, a N-1 and T = [t 0 , t 1 , …, t N-1 can be defined respectively.

[0079] In an embodiment of the present invention, the preprocessing of the input features specifically includes: preprocessing the seismic signal X using a residual model, and preprocessing the arrival time information and travel time information using two linear models respectively.

[0080] Furthermore, since the shapes of X, A, and T are different, we construct two linear models (LMs) and a residual model (RM) based on a neural network to jointly process them, as Figure 5 shown. The two LM models and one RM model jointly form a "preprocessing module". From the perspective of seismic analysis, the roles of the LM and RM in this architecture can be further explored. When the location of the seismic source is detected, its magnitude can be calculated using the median of the peak amplitude ratio between this earthquake and some "template" earthquake samples (Peng et al., 2009). Considering the powerful feature extraction ability of ResNet, the RM can perform a function similar to "template sample selection". Preprocessing X using the RM can effectively extract those samples whose hidden features can best reflect the attributes of the seismic signal. In addition, since the arrival time is another important index commonly used in traditional magnitude estimation methods, when performing data augmentation, we apply another LM structure to A, so as to further improve the expressive ability of MagInfoNet.

[0081] S200. Based on the preprocessed input features, perform magnitude estimation to obtain a prediction result;

[0082] In an embodiment of the present invention, the preprocessed input features are h = [hx , h a , h t . h is the concatenation of h x , h a and h t .

[0083] In an embodiment of the present invention, the magnitude estimation is performed based on the preprocessed input features to obtain a prediction result, specifically including:

[0084] The magnitude prediction module is used for magnitude estimation, including UniMP and a convolutional residual structure. Among them, a path graph is used as the graph topology of UniMP for information transmission.

[0085] The path graph treats each moment t and its corresponding earthquake signal value d t as the graph node v t and the feature of this node respectively, and for all t, v t points to the next node v t+1 .

[0086] For the input earthquake signal d = [d 0 , d 1 , …, d N , regarding each moment as a node, all nodes can be connected to the nodes of the next moment to form a path graph structure, and its adjacency matrix can be expressed as

[0087]

[0088] The earthquake signal d is the node signal in the earthquake signal X.

[0089] When A m is used in the graph neural network UniMP, d t will be affected by the values at the two timestamps before and after the current moment at the same time. Therefore, by using A m as the underlying topology structure of the network, the autocorrelation of the earthquake signal can be extracted.

[0090] In one embodiment, as shown in Figure 4 , in the embodiment of the present invention, a magnitude estimation model is also provided. The magnitude estimation model uses UniMP, ResNet, and RNN to build a neural network structure; the system includes a preprocessing module 500 and a magnitude prediction module 600.

[0091] The preprocessing module 500 is used to construct input features with the earthquake signals in the sample set and the arrival time information and travel time information corresponding to the earthquake signals, and preprocess the input features;

[0092] In an embodiment of the present invention, the preprocessing module 500 includes a residual model and two linear models;

[0093] The residual model is used to preprocess the seismic signal X;

[0094] The two linear models preprocess the arrival time information and the travel time information respectively.

[0095] Furthermore, for an earthquake event E i , its seismic signal x i includes signals from three directions. When we define the P-wave arrival time and the S-wave arrival time of the seismic signal as pa i and sa i respectively, then the arrival time a i = [pa i , sa i can be defined. Considering that the seismic signal is a discrete time series, thus pa i , sa i = 0, 1, …, L - 1, where L is the length of the seismic signal. Similarly, the P-wave travel time can be defined as t i . Assuming the number of earthquake events is N, then the seismic signal X = [x 0 , x 1 , …, x N-1 , A = [a 0 , a 1 , …, a N-1 and T = [t 0 , t 1 , …, t N-1 can be defined respectively.

[0096] In an embodiment of the present invention, the preprocessing of the input features specifically includes: using the residual model to preprocess the seismic signal X, and using the two linear models to preprocess the arrival time information and the travel time information respectively.

[0097] Furthermore, since the shapes of X, A, and T are different, therefore, based on the neural network, we constructed two linear models (LMs) and one residual model (RM) to jointly process them, as Figure 5As shown. Two LM models and one RM model together constitute the "preprocessing module". From the perspective of seismic analysis, the roles of LM and RM in this architecture can be further explored. When the location of an earthquake hypocenter is detected, its magnitude can be calculated using the median of the peak amplitude ratios between this earthquake and certain "template" earthquake samples (Peng et al., 2009). Considering the powerful feature extraction ability of ResNet, RM can perform a function similar to "template sample selection". By preprocessing X using RM, samples whose hidden features can best reflect the attributes of seismic signals can be effectively extracted. In addition, since the arrival time is another important metric commonly used in traditional magnitude estimation methods, when performing data augmentation, we applied another LM structure to A, thus hoping to further improve the expressive ability of MagInfoNet.

[0098] The magnitude prediction module 600 is used to perform magnitude estimation based on the preprocessed input features to obtain a prediction result.

[0099] The magnitude prediction module 600 is the Mag-Pred module.

[0100] The magnitude prediction module 600 consists of a unified message passing model UniMP and another RM structure.

[0101] After preprocessing the seismic signals, arrival time information, and travel time information, another Mag-Pred module was then constructed to estimate the magnitude, which consists of a UniMP with a graph Transformer structure of the unified message passing model and another RM structure ( Figure 6 ). Note that h is the concatenation of h x 、h a and h t , so they need to be jointly integrated into subsequent analyses. Since ResNet has stronger expressive ability than simple CNNs (Allen Zhu et al., 2019), RM is used for information integration, so the output of this RM structure can be considered as a comprehensive result obtained based on the analysis of h.

[0102] The present invention improves an existing single-station magnitude estimation system by using deep learning models and data augmentation methods. We use a transfer structure of a Transformer of UniMP on graph neural networks and jointly build a deep learning model with the previously proposed residual convolutional neural network ResNet and recurrent neural network RNN, and use the arrival time and travel time information of seismic signals to perform data augmentation on the input features, thereby improving the prediction accuracy, applicable range, and robustness of the estimation system.

[0103] In one embodiment, an embodiment of the present invention further provides a computer device, including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus.

[0104] The memory is used to store computer programs.

[0105] The processor is used to execute the magnitude estimation method when executing the computer program stored on the memory. When the processor executes instructions, it implements the steps in the above method embodiments.

[0106] The communication bus mentioned in the above terminal may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, only a thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0107] The communication interface is used for communication between the above terminal and other devices.

[0108] The memory may include a Random Access Memory (RAM), or may also include a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0109] The above-mentioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0110] The computer device includes a user device and a network device. Among them, the user device includes, but is not limited to, a computer, a smart phone, a PDA, etc.; the network device includes, but is not limited to, a single network server, a server group composed of multiple network servers, or a cloud composed of a large number of computers or network servers based on Cloud Computing. Among them, Cloud Computing is a type of distributed computing, which consists of a super virtual computer composed of a group of loosely coupled computer sets. Among them, the computer device can run independently to implement the present invention, or can be connected to a network and implement the present invention through interactive operations with other computer devices in the network. Among them, the network where the computer device is located includes, but is not limited to, the Internet, a wide area network, a metropolitan area network, a local area network, a VPN network, etc.

[0111] It should also be understood that the term "and / or" used in the specification and appended claims of the present invention refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0112] In an embodiment of the present invention, a storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the embodiment of the above magnitude estimation method are implemented.

[0113] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided by the present invention can include at least one of non-volatile and volatile memories.

[0114] Experimental results

[0115] The present invention uses a publicly available earthquake signal dataset STEAD (Mousavi et al., 2019) to test the magnitude estimation performance of MagInfoNet. We randomly selected 200,000 samples from the dataset, 90% of which were used for model training, and the remaining 10% were used for model testing. Figure 7Shows the source locations of earthquake samples (a) and the corresponding locations of earthquake signal receiving stations (b), where the distances between the sources and stations of some samples (i.e., epicentral distances) are large. Since it is difficult to calculate the magnitudes of some earthquake events with large epicentral distances using traditional seismological methods (Jing et al., 2022), we hope that MagInfoNet can achieve a more accurate estimate of them. STEAD mainly includes two magnitudes: and. However, since the calculation methods and corresponding magnitude values of these two magnitudes are not the same, we need to analyze them separately from each other. The value distributions of M L and M D in the training set and test set are as Figure 8 shown. It can be found that the difference in value distribution between the training set and the test set is small, while the difference between the M L and M D samples is large.

[0116] Based on different models, we tested the estimation performance, including MagInfoNet, MagNet, RF, and MINLSTM (using the LSTM layer (Xue et al., 2022) instead of the UniMP layer to compare the performance between GNN and BRNN). We did not perform redundant selective steps and other filtering operations on the samples so that the experimental results could better reveal the estimation performance of MagInfoNet. The experimental results are as Figure 9 shown, where we used the coefficient of determination as a metric (McColl et al., 2014). As can be seen from the figure, the performance of MagInfoNet is indeed better than other existing estimation models. For and, from MagNet to MagInfoNet, there are increases of approximately 2.915% and 6.365% respectively. At the same time, the error distribution between the estimated values and the true values of MagInfoNet has a smaller standard deviation (std) than other models. We also assigned a larger proportion (50%) to the test set than the commonly used 10% to better represent the generalization ability of MagInfoNet. It can be considered that for more samples in the test set, more types of earthquake signals will be covered (the results are as Figure 10 ). Figure 9 and Figure 10 The results of both show that our model is superior to other models to a certain extent.

[0117] It should be understood that, as used herein, unless the context clearly supports an exception, the singular form "a" is intended to also include the plural form. It should also be understood that "and / or" as used herein refers to any and all possible combinations of one or more of the associated listed items. The serial numbers of the disclosed embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0118] Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope (including the claims) disclosed by the embodiments of the present invention is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other variations in different aspects of the embodiments of the present invention as above, which are not provided in detail for the sake of brevity. Therefore, any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present invention shall be included within the protection scope of the embodiments of the present invention.

Claims

1. A training method for a magnitude estimation model, characterized in that, the method comprises: obtaining a sample set, wherein the sample set includes seismic signals and corresponding true results of the seismic signals; based on the sample set, obtaining a prediction result through a magnitude estimation model; calculating the error between the prediction result and the true result, and calculating the error gradient; wherein, calculating the error gradient is obtained by using the backpropagation algorithm; updating the weights by using a weight optimizer constructed based on the Adam algorithm, wherein, the weights are updated by setting the weight learning rate and regularization parameter given by the magnitude estimation model based on the negative direction of gradient descent to update the weights of the model.

2. The training method for a magnitude estimation model according to claim 1, characterized in that, calculating the error between the prediction result and the true result is calculated through a loss function constructed by the mean square error MSE.

3. A magnitude estimation method, characterized in that, the method is applied to a magnitude estimation model that uses UniMP, ResNet, and RNN to build a neural network structure, and the magnitude estimation model is trained by using the method described in any one of claims 1-2. The magnitude estimation method comprises: constructing input features with the seismic signals in the sample set and the corresponding arrival time information and travel time information of the seismic signals, and preprocessing the input features; based on the preprocessed input features, performing magnitude estimation to obtain a prediction result, specifically including: using a magnitude prediction module to perform magnitude estimation, wherein the magnitude prediction module includes UniMP and a residual model, and wherein, a path graph is used as the topological structure of UniMP for information transmission; For the input seismic signal d = [d 0 , d 1 , …, d N , regarding each moment as a node, all nodes can be connected to the nodes at the next moment to form a structure of a path graph, and its adjacency matrix can be expressed as 4. The magnitude estimation method according to claim 3, characterized in that, the arrival time information includes the P-wave arrival time and S-wave arrival time of the seismic signal.

5. The magnitude estimation method according to claim 3, characterized in that, the travel time information includes the P-wave travel time.

6. The magnitude estimation method according to claim 3, characterized in that, preprocessing the input features specifically includes: preprocessing the seismic signals by using a residual model, and preprocessing the arrival time information and travel time information respectively by using two linear models.

7. A magnitude estimation model, characterized in that, it is trained by using the method described in any one of claims 1-2, and the magnitude estimation model uses UniMP, ResNet, and RNN to build a neural network structure.

8. The magnitude estimation model according to claim 7, characterized in that, the model includes a preprocessing module and a magnitude prediction module; the preprocessing module is used to construct input features with the seismic signals in the sample set and the corresponding arrival time information and travel time information of the seismic signals, and preprocess the input features; the magnitude prediction module is used to perform magnitude estimation based on the preprocessed input features to obtain a prediction result.

9. The magnitude estimation model according to claim 8, characterized in that, the magnitude prediction module is composed of UniMP and a residual model.

Images