An intelligent prediction model for short-term and impending aftershocks of moderate-strong earthquakes based on deep meta-learning

Abstract

The present invention discloses an intelligent prediction model for short-term aftershocks of moderate-strong earthquakes based on deep meta-learning, which is characterized by comprising a multi-source earthquake catalog preprocessing model, a dynamic framework model for predicting short-term aftershocks of moderate-strong earthquakes and a deep meta-learning network for aftershock prediction; the multi-source earthquake catalog preprocessing model comprises three parts, namely, removal of duplicate samples in the earthquake catalog, filling of missing fields such as earthquake energy and processing of earthquake sample abnormal values ​​by a box plot method; the dynamic framework model for predicting short-term aftershocks of moderate-strong earthquakes comprises generation of aftershock prediction model training samples, dynamic selection of aftershock prediction model parameters and flexible switching of multiple deep learning algorithms based on meta-learning. The present invention can break through the dependence on the geophysical relationship between main and aftershock parameters in aftershock prediction research, and use AI algorithms to mine and study the complex, abstract and deep correlation relationships in the earthquake catalog.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake prediction, and in particular to an intelligent prediction model for moderate-to-strong earthquakes and short-term aftershocks based on deep meta-learning. Background Art

[0002] Earthquake prediction is an internationally recognized global problem. After decades of development, the earthquake prediction industry around the world has accumulated a wealth of valuable experience and a large amount of historical data. In particular, with the further improvement and improvement of communication technology, network technology, observation technology, etc. in recent years, the data in the earthquake database has increased dramatically. Therefore, people hope to further explore the knowledge and information hidden behind the earthquake data. Most of them are implicit, unknown in advance, and have potential value. People are looking forward to using newer and more effective technologies to discover the important earthquake laws hidden in them and predict future earthquake development trends. The earthquake catalog is one of the most valuable and critical data in historical earthquake record data.

[0003] The existing multi-source earthquake catalogs in the world contain a large amount of information on earthquake energy radiation and focal mechanism solutions, and contain rich temporal and spatial correlations of earthquakes. Exploring these correlations has very important practical value for accurately determining earthquake trends, especially for predicting short-term aftershocks after moderate-strong earthquakes. It is generally believed that there is an abstract correlation between earthquake parameters such as the main shock magnitude, apparent stress, seismic moment, focal depth, and the maximum aftershock magnitude, the maximum aftershock occurrence time, and the maximum aftershock spatial position. Establishing a mathematical statistical model based on earthquake records can provide a strong theoretical basis for the rapid determination of post-earthquake trends.

[0004] The determination of post-earthquake trends is a very important part of earthquake prediction work and is highly valued by earthquake workers at home and abroad. At present, judging earthquake trends based on seismic activity is still the main determination method, such as b value, h value, K value, U value, seismic activity image, seismic activity enhancement and calm, focal mechanism, lunar characteristics of seismic activity, etc. These methods have played a good role in actual earthquake prediction work and will continue to play their role.

[0005] The existing intelligent prediction model for moderate-to-strong earthquakes and short-term aftershocks based on deep meta-learning has the following defects:

[0006] 1. The types and number of global earthquake catalogs used are relatively small, the sample representativeness is not high, and it needs to be continuously enriched, and the information cannot meet the needs;

[0007] 2. Existing technologies mostly use a single algorithm, which cannot accurately predict the model and easily leads to inaccurate predictions.

[0008] 3. After the existing technical forecasts are studied, their application is relatively complex and cannot achieve the purpose of notifying residents in a timely manner. Summary of the invention

[0009] The purpose of the present invention is to provide an intelligent prediction model for short-term aftershocks of moderate to strong earthquakes based on deep meta-learning.

[0010] To achieve the above-mentioned object, the present invention provides the following technical solutions: an intelligent prediction model for moderate-strong earthquakes and short-imminent aftershocks based on deep meta-learning, comprising a multi-source earthquake catalog preprocessing model, a dynamic framework model for prediction of moderate-strong earthquakes and short-imminent aftershocks, and a deep meta-learning network for aftershock prediction;

[0011] The multi-source earthquake catalog preprocessing model includes the elimination of duplicate samples in the earthquake catalog, the filling of missing fields such as earthquake energy, and the use of a box plot method to process abnormal values ​​of earthquake samples;

[0012] The dynamic framework model for predicting short-term aftershocks of moderate-strong earthquakes includes generation of aftershock prediction model training samples, dynamic selection of aftershock prediction model parameters, and flexible switching of multiple deep learning algorithms based on meta-learning;

[0013] The aftershock prediction deep meta-learning network includes two methods for simulation, a complex spatiotemporal relationship fitting method between main shock and aftershock sequences and a meta-learning optimization method for deep learning model network parameters.

[0014] Preferably, the multi-source earthquake catalog preprocessing model determines the fields that must be retained in the end and removes duplicate samples through in-depth analysis of multiple earthquake catalogs.

[0015] Preferably, the elimination of duplicate samples in the earthquake catalog means that the collected earthquake catalog data is very large, and during the collection process, it is inevitable that human or other unknown factors will cause individual data to be repeated during the collection and sorting process, and these data need to be eliminated as duplicate samples.

[0016] Preferably, the filling of missing fields such as earthquake energy mainly refers to fitting and filling missing fields such as earthquake energy using a support vector machine: ① According to the mechanical characteristics of earthquake occurrence, earthquakes in the CMT catalog are divided into four categories according to the corresponding slip angles: thrust type, normal fault type, strike-slip type and other types;

[0017] ② The support vector machine is used to construct a nonlinear regression prediction model to predict the energy fields of four types of earthquakes respectively, and then the energy fields are filled.

[0018] Preferably, the method of processing outliers in seismic samples by using the box plot method specifically involves finding outliers in seismic samples by using the box plot method.

[0019] Preferably, the training samples of the aftershock prediction model are generated. The three elements of an earthquake are the key information for identifying an earthquake and are also the target values ​​mainly predicted by the maximum aftershock prediction model. An in-depth analysis of various earthquake catalogs yields 13 characteristic parameters of the maximum aftershock prediction model and the target value to be predicted:

[0020] By finding all the main shock and aftershock sequences that meet the conditions and the corresponding maximum aftershocks, we can construct an aftershock prediction sample. Let the main shock occur at , and the aftershock occur at , the time difference between the main shock and aftershock; the distance between the main shock and aftershock is ; Main shock magnitude , aftershock magnitude , the conditions for judging whether an earthquake is an aftershock of a main shock include: ; ; , iteratively traverse all earthquakes from the processed earthquake samples, screen out all earthquake samples that meet the conditions, and finally form aftershock prediction sample information.

[0021] Preferably, the aftershock prediction model parameters are dynamically selected, and the characteristic parameters listed in all the collected earthquake information will be selected, and the individual characteristic parameters will be flexibly adjusted in combination with the information of the current earthquake focal mechanism solution.

[0022] Preferably, the multiple deep learning algorithms based on meta-learning are flexibly switched, and different deep learning algorithms can meet the prediction of earthquake aftershocks in different situations. An aftershock prediction AI model is constructed by using multiple deep learning algorithms based on meta-learning, and error comparison and analysis are performed based on the prediction results of different deep learning algorithms to achieve flexible switching between algorithms.

[0023] Preferably, the complex spatiotemporal relationship fitting method between the main shock and aftershock sequences is aimed at the abstract, deep and organic spatiotemporal relationship between the main shock (magnitude, radiation energy, apparent stress, seismic moment and other parameters) and the subsequent aftershocks (magnitude, main shock and aftershock time difference, main shock and aftershock spatial distance). The deep learning algorithm adopted by the patent is as follows:

[0024] 1) Nested LSTMs aftershock prediction AI model;

[0025] 2) Transformer network aftershock prediction AI model.

[0026] Preferably, the meta-learning optimizes the network parameters of the deep learning model, and uses meta-learning techniques such as MAML to optimize the parameters in the Nested LSTMs and Transformer network models. First, the processed earthquake catalog data is used as input data to construct the aftershock prediction AI models of the Nested LSTMs and Transformer networks respectively, and then the two deep learning algorithms of the Nested LSTMs and Transformer networks are embedded as operators in the MAML meta-learning framework respectively. Take any one of the algorithms as an example, and the other one is similar: ① The earthquake catalog is divided into four categories according to the corresponding sliding angles; ② Sample the sample space of these four categories of earthquake catalogs, meta-learn the operator model to improve the robustness of the operator, and obtain the generalization (preliminary) parameters of the operator model; ③ For a specific earthquake type sample, start from the generalization parameters obtained by meta-learning, continue to refine the training operator, and finally obtain a deep meta-learning model with a higher accuracy.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. This invention can break through the reliance on the geophysical relationship between main and aftershock parameters in aftershock prediction research, and use AI algorithms to mine and study the complex, abstract, and deep correlations in earthquake catalogs;

[0029] 2. This invention aims at the deep and abstract relationship between the main shock (magnitude, apparent seismic stress, moment, etc.) and the largest aftershock (magnitude, time and space difference with the main shock), improves a variety of the latest deep learning algorithms, such as Nested LSTMs and Transformer networks, and thus provides an AI prediction model for short-term aftershocks of moderate-strong earthquakes. It also uses meta-learning technologies such as MAML to optimize the corresponding model network parameters, significantly improving the prediction effect of the model;

[0030] 3. The present invention provides high-precision prediction software developed based on the AI ​​platform, which simplifies the research and application process of aftershock prediction and makes it more timely. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the structure of the intelligent prediction model for moderate-strong earthquakes and short-term aftershocks based on deep meta-learning of the present invention;

[0032] Figure 2 It is a strong earthquake short-term aftershock prediction flow chart of the present invention;

[0033] Figure 3 The structure diagram of the aftershock prediction AI model constructed for the Nested LSTMs neural network of the present invention

[0034] Figure 4 The internal structure diagram of the Transformer module of the present invention

[0035] Figure 5 The structure diagram of the aftershock prediction AI model based on the Transformer network of the present invention;

[0036] Figure 6 This is a meta-learning framework diagram of the aftershock prediction model of the present invention;

[0037] Figure 7 is the earthquake type table of the present invention;

[0038] Figure 8 These are the characteristic parameters and prediction target values ​​of the aftershock prediction model of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] See also Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 and Figure 8 , an intelligent prediction model for short-term and aftershocks of moderate-strong earthquakes based on deep meta-learning,

[0041] Embodiment 1: including a multi-source earthquake catalog preprocessing model, a dynamic framework model for medium-strong earthquake short-term aftershock prediction, and a deep meta-learning network for aftershock prediction;

[0042] The multi-source earthquake catalog preprocessing model includes three parts: elimination of duplicate samples in the earthquake catalog, filling in missing fields such as earthquake energy, and processing earthquake sample abnormal values ​​using a box plot method;

[0043] The dynamic framework model for predicting short-term aftershocks of moderate-strong earthquakes includes generation of aftershock prediction model training samples, dynamic selection of aftershock prediction model parameters, and flexible switching of multiple deep learning algorithms based on meta-learning;

[0044] The aftershock prediction deep meta-learning network includes two methods for simulation: a complex spatiotemporal relationship fitting method between main shock and aftershock sequences and a meta-learning optimization method for deep learning model network parameters.

[0045] Embodiment 2: The multi-source earthquake catalog preprocessing model determines the fields that must be retained in the end through in-depth analysis of multiple earthquake catalogs, removes duplicate samples, and uses algorithms such as support vector machines to fill in missing fields such as earthquake energy, and designs a box-and-line method or standard deviation method model to remove outliers in earthquake samples.

[0046] The elimination of duplicate samples in the earthquake catalog means that the collected earthquake catalog data is very large, and it is inevitable that human or other unknown factors will cause individual data to be repeated during the collection and sorting process. These data need to be eliminated as duplicate samples to prevent overfitting of the model and improve the generalization ability of the model.

[0047] The filling of missing fields such as earthquake energy mainly refers to the use of support vector machines to fit and fill missing fields such as earthquake energy: ① According to the mechanical characteristics of earthquake occurrence, earthquakes in the CMT catalog are divided into four categories according to the corresponding sliding angles:

[0048] ② The support vector machine is used to construct a nonlinear regression prediction model to predict the energy fields of four types of earthquakes respectively, and then the energy fields are filled.

[0049] The box plot method for processing outliers in earthquake samples is specifically to use the box plot method to find outliers in earthquake samples, and find out the data outside the upper and lower edges of the data according to the data distribution displayed by the box plot. The upper edge is the upper quartile plus 1.5 times the interquartile range, and the lower edge is the lower quartile minus 1.5 times the interquartile range. These data are deleted to eliminate outliers in earthquake samples.

[0050] Embodiment 3: The training samples of the aftershock prediction model are generated. The three elements of an earthquake are the key information for identifying an earthquake and are also the target values ​​mainly predicted by the maximum aftershock prediction model. Specifically, the training samples of the model are input into the aftershock prediction AI model based on deep learning, and three prediction target values ​​of aftershock magnitude, aftershock time and aftershock space are output. The characteristic parameters of the training sample set and the prediction sample set of the model are mainly based on the CMT catalog. At the same time, the earthquake energy field (from the NEIC earthquake catalog) is added as a characteristic parameter of the training data. In addition to the energy field, the earthquake apparent stress is also used as an input earthquake parameter. The earthquake apparent stress can be obtained by The earthquake efficiency is obtained by multiplying the average source stress. The strength of the earthquake source stress field can be reflected by the average earthquake apparent stress. The higher the earthquake apparent stress, the higher the corresponding stress level in the source area. On the contrary, the lower the stress level in the source area, many studies have shown that: at low stress levels, the possibility of a larger or larger earthquake occurring again after a strong earthquake is small, while at high stress levels, the possibility of a larger or larger earthquake occurring again after a strong earthquake is greater. Therefore, this paper also uses earthquake apparent stress as a characteristic value for aftershock prediction. Finally, an in-depth analysis of various earthquake catalogs yields 13 characteristic parameters of the maximum aftershock prediction model and the target value to be predicted:

[0051] After that, the aftershock prediction sample is constructed by finding all the main shock and aftershock sequences that meet the conditions and the corresponding maximum aftershocks. Assume that the main shock occurs at time , and the aftershock occurs at time , the time difference between the main shock and aftershock; the distance between the main shock and aftershock is ; Main shock magnitude , aftershock magnitude , the conditions for judging whether an earthquake is an aftershock of a main shock include: ; ; , iteratively traverse all earthquakes from the processed earthquake samples, screen out all earthquake samples that meet the conditions, and finally form aftershock prediction sample information.

[0052] The parameters of the aftershock prediction model are dynamically selected. We will select the characteristic parameters listed in all the collected earthquake information. However, in special circumstances (such as data missing, etc.), it is impossible to use all characteristic parameters. We can also flexibly adjust individual characteristic parameters according to the actual earthquake situation faced by earthquake experts and the information of the current earthquake focal mechanism solution.

[0053] The multiple deep learning algorithms based on meta-learning are flexibly switched. Different deep learning algorithms can meet the prediction of earthquake aftershocks in different situations. An aftershock prediction AI model is constructed by using multiple deep learning algorithms based on meta-learning, and an error comparison analysis is performed based on the prediction results of different deep learning algorithms to achieve flexible switching between algorithms. Under the meta-learning framework, according to the advantages of different deep learning algorithms, a variety of aftershock prediction AI models based on different deep learning are constructed using a variety of different deep learning algorithms, and an error comparison analysis is performed on the results predicted by these models. According to the analysis results and the actual judgment of earthquake experts, a suitable aftershock prediction AI model is flexibly selected to predict the maximum aftershock.

[0054] Embodiment 4: The complex spatiotemporal relationship fitting method between the main shock and aftershock sequences is aimed at the abstract, deep and organic spatiotemporal relationship between the main shock (magnitude, radiation energy, apparent stress, seismic moment and other parameters) and the subsequent aftershocks (magnitude, main shock and aftershock time difference, main shock and aftershock spatial distance). The deep learning algorithm used in the patent is as follows:

[0055] 1) Nested LSTMs aftershock prediction AI model

[0056] When accessing internal memory, Nested LSTMs have a higher degree of freedom than traditional stacked LSTMs, so they can handle internal memories of longer time scales. Intuitively, the output gate in LSTM will encode information that is still worth remembering, which may be irrelevant to the current time step. Nested LSTMs create a new memory time hierarchy based on this intuitive understanding. Access to internal memory is gated in the same way, so that long-term information can be selectively accessed only under context-related conditions. Experiments also show that Nested LSTMs surpass stacked LSTMs in many tasks. These characteristics of Nested LSTMs are consistent with the characteristics of earthquake sequences, that the time intervals vary in length, and that the correlations between sequences are highly abstract and ambiguous, making them difficult to directly explain with simple physical formulas. Figure 3 It is a structural diagram of an aftershock prediction AI model based on the Nested LSTMs neural network. It is a variant deep learning model of Nested LSTMs suitable for aftershock prediction. It takes the earthquake sequence as input and trains the aftershock prediction AI model based on the Tensorflow and Keras platforms to achieve intelligent prediction of aftershocks.

[0057] 2) Transformer network aftershock prediction AI model

[0058] The Transformer module is based only on the attention mechanism and no longer uses any RNN or CNN structure. Figure 4It is the internal structure of the Transformer module. The Transformer module consists of two parts: Encoders and Decoders. Each part is composed of 6 identical encoder or decoder modules stacked together. A single encoder and decoder module contains an attention mechanism. The advantage of using the attention mechanism is that the number of operations required to calculate the association between two positions does not increase with the distance, which solves the long-distance dependency problem in RNN and supports parallel computing, greatly improving the training efficiency of the model. Figure 5 It is a structural diagram of an aftershock prediction AI model based on the Transformer network. It is a Transformer network model suitable for aftershock prediction. It also inputs an earthquake sequence and trains the aftershock prediction AI model based on the Tensorflow and Keras platforms to achieve intelligent prediction of aftershocks.

[0059] The meta-learning optimizes the network parameters of the deep learning model. The quality of the parameters directly affects the training speed, overfitting, accuracy and other issues of a model. The present invention adopts meta-learning technologies such as MAML to optimize the parameters in the Nested LSTMs and Transformer network models. First, the processed earthquake catalog data is used as input data to respectively construct the aftershock prediction AI models of the Nested LSTMs and Transformer networks. Then, the two deep learning algorithms of the Nested LSTMs and Transformer networks are respectively embedded as operators in the MAML meta-learning framework. Take any one of the algorithms as an example, and the other one is similar: ① The earthquake catalog is divided into four categories according to the corresponding sliding angles; ② Sample the sample space of these four categories of earthquake catalogs, and meta-learn the operator model to improve the robustness of the operator and obtain the generalization (preliminary) parameters of the operator model; ③ For a specific earthquake type sample, start from the generalization parameters obtained by meta-learning, continue to refine the training operator, and finally obtain a deep meta-learning model with a higher accuracy rate. The meta-learning framework of the aftershock prediction model is as follows: Figure 6 As shown in the figure, this method can effectively improve the generalization ability of the moderate-strong earthquake short-term aftershock prediction model on earthquake samples from different sources, and make the model have higher accuracy and lower error when facing small sample data. Based on the Keras platform, a set of intelligent prediction software for moderate-strong earthquake short-term aftershocks was developed based on multi-source earthquake catalogs and focal mechanism solutions.

[0060] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0061] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent prediction model for short-term and aftershocks of moderate-strong earthquakes based on deep meta-learning, characterized in that: Including multi-source earthquake catalog preprocessing model, dynamic framework model for prediction of short-term aftershocks of moderate-strong earthquakes, and deep meta-learning network for aftershock prediction; The multi-source earthquake catalog preprocessing model includes the elimination of duplicate samples in the earthquake catalog, the filling of missing fields of earthquake energy, and the use of a box plot method to process abnormal values ​​of earthquake samples; The dynamic framework model for predicting short-term aftershocks of moderate-strong earthquakes includes generation of aftershock prediction model training samples, dynamic selection of aftershock prediction model parameters, and flexible switching of multiple deep learning algorithms based on meta-learning; The aftershock prediction deep meta-learning network includes two methods for simulation, namely, a method for fitting complex spatiotemporal relationships between main shock and aftershock sequences and a method for optimizing deep learning model network parameters through meta-learning; The filling of the missing field of earthquake energy mainly refers to fitting and filling the missing field of earthquake energy using a support vector machine: ① Based on the mechanical characteristics of earthquake occurrence, earthquakes in the CMT catalog are divided into four categories according to the corresponding slip angles: thrust type, normal fault type, strike-slip type and other types; ②Use support vector machine to construct nonlinear regression prediction model, predict the energy fields of four types of earthquakes respectively, and then fill the energy fields; The training samples of the aftershock prediction model are generated. The three elements of an earthquake are the key information for identifying an earthquake and are also the target values ​​mainly predicted by the maximum aftershock prediction model. An in-depth analysis of various earthquake catalogs yields 13 characteristic parameters of the maximum aftershock prediction model and the target values ​​to be predicted: By finding all the main shock and aftershock sequences that meet the conditions and the corresponding maximum aftershocks, we construct an aftershock prediction sample. Let the main shock occurrence time be data. i , the aftershock occurred at data j , the time difference between the main shock and aftershock t = data j -data i ; The distance between the main shock and aftershock is d; The magnitude of the main shock is Mw i , aftershock magnitude Mw j , the conditions for judging whether an earthquake is an aftershock of a main shock include: t≤n; d≤m; Mw i >Mw j , iterate through all earthquakes from the processed earthquake samples, filter out all earthquake samples that meet the conditions, and finally form aftershock prediction sample information; The meta-learning optimizes the network parameters of the deep learning model, and uses the MAML meta-learning technology to optimize the parameters in the Nested LSTMs and Transformer network models. First, the processed earthquake catalog data is used as input data to build the aftershock prediction AI models of the Nested LSTMs and Transformer networks respectively. Then, the two deep learning algorithms, Nested LSTMs and Transformer networks, are embedded as operators into the MAML meta-learning framework respectively: ① Divide the earthquake catalog into four categories according to the corresponding slip angles; ② Sampling in the sample space of these four types of earthquake catalogs, meta-learning the operator model to improve the robustness of the operator and obtain the generalization parameters of the operator model; ③ For a specific earthquake type sample, start from obtaining generalization parameters through meta-learning, continue to refine the training operator, and finally obtain a deep meta-learning model with high accuracy; The complex spatiotemporal relationship fitting method between the main shock and aftershock sequences adopts the following deep learning algorithm for the abstract, deep and organic spatiotemporal relationship between the main shock and subsequent aftershocks: 1) Nested LSTMs aftershock prediction AI model; 2) Transformer network aftershock prediction AI model.

2. According to claim 1, a deep meta-learning-based intelligent prediction model for moderate-strong earthquakes and short-term aftershocks is characterized by: The multi-source earthquake catalog preprocessing model determines the fields that must be retained in the end and removes duplicate samples through in-depth analysis of multiple earthquake catalogs.

3. The intelligent prediction model for moderate-strong earthquakes and short-term aftershocks based on deep meta-learning according to claim 1 is characterized by: The elimination of duplicate samples in the earthquake catalog means that the collected earthquake catalog data is very large, and during the collection process, it is inevitable that human or other unknown factors will cause individual data to be repeated during the collection and sorting process. These data need to be eliminated as duplicate samples.

4. According to claim 1, a deep meta-learning-based intelligent prediction model for moderate-strong earthquakes and short-term aftershocks is characterized by: The method of using the box plot method to process the outliers of the earthquake samples is to use the box plot method to find the outliers in the earthquake samples.

5. According to claim 1, a deep meta-learning-based intelligent prediction model for moderate-strong earthquakes and short-term aftershocks is characterized by: The dynamic selection of aftershock prediction model parameters is to select characteristic parameters listed in all earthquake information collected, and flexibly adjust individual characteristic parameters in combination with the information of the current earthquake focal mechanism solution.

6. According to claim 1, a deep meta-learning-based intelligent prediction model for moderate-strong earthquakes and short-term aftershocks is characterized by: The multiple deep learning algorithms based on meta-learning are flexibly switched, and different deep learning algorithms are used to meet the prediction of earthquake aftershocks in different situations. An aftershock prediction AI model is constructed using multiple deep learning algorithms based on meta-learning, and error comparison and analysis are performed based on the prediction results of different deep learning algorithms to achieve flexible switching between algorithms.

Images