A seismic prediction method based on a graph neural network

Abstract

This invention discloses an earthquake prediction method based on graph neural networks. First, the target area is subdivided and preprocessed with an earthquake catalog to obtain frequency characteristics and G-R relationship curves for each sub-region, and a training set is constructed for training the graph neural network. The training set is then provided to the graph neural network for training. For each sub-region, a corresponding number of graph neural networks are trained to predict whether an earthquake will occur in each sub-region. After training, the earthquake occurrence data for the target area within one year is input into the graph neural network to predict whether an earthquake of magnitude 5 or higher will occur in each sub-region of the target area within the next month. This method can effectively predict the probability of an earthquake of magnitude 5 or higher occurring in the target area within the next month, thereby ensuring the safety of people's production and lives.

Description

Technical Field

[0001] This invention relates to the field of earthquake prediction technology, and in particular to an earthquake prediction method based on graph neural networks. Background Technology

[0002] Earthquake prediction can generally be divided into long-term (over 10 years), medium-term (1 to 10 years), and short-term (1 to several hundred days). The distribution of strong regional earthquakes over time is uneven and can be divided into relatively calm periods and significantly active periods that can transform into each other. Short-term prediction is the more difficult part. For example, earthquake precursor anomalies include gravity field anomalies, magnetic field anomalies, electric field anomalies, changes in underground resistivity, changes in groundwater level, changes in groundwater chemical composition, etc. However, these anomalies are also diverse and unreliable.

[0003] Machine learning, a combination of algorithms and statistical models, enables computers to directly extract feature information from large datasets. Unlike scientists who discover and understand physical laws, machine learning methods learn specific patterns directly from data without inferring the underlying physical mechanisms. Experienced signal analysts possess intuition and logic for analyzing data, but machine learning can surpass humans in discovering new features invisible to them. With the rapid pace of technological advancement, more and more types of signals are being detected, and the amount of signal data being detected is currently undergoing rapid changes in quantity, variety, and speed, with data volume growing exponentially, making traditional data processing methods increasingly difficult. Many machine learning algorithms are designed with this need in mind; often, more data yields better results. Therefore, the application of machine learning technology in signal processing and seismology has developed rapidly. However, current technologies lack effective solutions for short-term earthquake prediction. Summary of the Invention

[0004] The purpose of this invention is to provide an earthquake prediction method based on graph neural networks, which can effectively predict the probability of an earthquake of magnitude 5 or above occurring in a target area within the next month, thereby ensuring the safety of people's production and life.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] An earthquake prediction method based on graph neural networks, the method comprising:

[0007] Step 1: Perform sub-region division and earthquake catalog preprocessing on the target area to obtain frequency characteristics and GR relationship curves of sub-regions, and construct training set data for graph neural network training;

[0008] Step 2: Provide training set data to the graph neural network for training;

[0009] Step 3: For each sub-region divided in Step 1, train a corresponding number of graph neural networks to predict whether an earthquake will occur in each sub-region.

[0010] Step 4: After completing the graph neural network training, input the earthquake data of the target area within one year into the graph neural network to predict whether an earthquake of magnitude 5 or above will occur in each sub-region of the target area within the next month.

[0011] As can be seen from the technical solution provided by the present invention, the above method can effectively predict the probability of an earthquake of magnitude 5 or above occurring in the target area within the next month, thereby ensuring the safety of people's production and life. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic diagram of the earthquake prediction method based on graph neural networks provided in an embodiment of the present invention;

[0014] Figure 2 This is a schematic diagram illustrating the target region division according to an embodiment of the present invention;

[0015] Figure 3 This is a schematic diagram of the evaluation indicators provided in the embodiments of the present invention;

[0016] Figure 4 This is a schematic diagram of the ROC curve of the trained network on the validation set according to an embodiment of the present invention;

[0017] Figure 5 This is a schematic diagram of the test results of each sub-region on the validation set in an embodiment of the present invention;

[0018] Figure 6 This is a schematic diagram comparing the method described in the embodiments of the present invention with the convolutional neural network prediction scheme in the prior art. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments, and do not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0020] like Figure 1 The diagram shown is a schematic flowchart of an earthquake prediction method based on a graph neural network provided in an embodiment of the present invention. The method includes:

[0021] Step 1: Perform sub-region division and earthquake catalog preprocessing on the target area to obtain frequency characteristics and GR relationship curves of sub-regions, and construct training set data for graph neural network training;

[0022] In this step, such as Figure 2 The diagram shown is a schematic diagram of the target area division according to an embodiment of the present invention. In this example, the target area is divided into 11 sub-areas. For each sub-area, the frequency of earthquakes occurring within 1 month, 3 months, 6 months and 1 year is counted. The horizontal axis is the magnitude of the earthquake and the vertical axis is the number of earthquakes of different magnitudes.

[0023] The frequencies were normalized by dividing the earthquake frequency within different time intervals by the corresponding time range. For example, the frequency of earthquakes occurring within 3 months was divided by 3 to obtain the final normalized result for that time period. In addition, the Gutenberg-Richter (GR) relationship curves for each sub-region over 1 month, 3 months, 6 months, and 1 year were calculated and used together with the frequencies as training set data.

[0024] The preprocessing process for the earthquake catalog is as follows: earthquakes of magnitude 3 or above are selected as input to the dataset, and earthquakes of magnitude 5 or above are selected as labels for the dataset.

[0025] In practice, the graph neural network is trained on a monthly basis, with the total data spanning 620 months from January 1971 to August 2022.

[0026] For earthquake prediction based on machine learning, the network training process requires a dataset containing both positive samples (those where earthquakes will occur in certain areas) and negative samples (those where earthquakes will not occur in certain areas) to enable the network to predict whether earthquakes will occur. Therefore, in creating the training set, samples where no magnitude 5 earthquake occurred in the preceding and following two months were used as negative samples. After removing data that was neither positive nor negative, all samples were divided into the following time periods:

[0027] A total of 300 samples from January 1971 to August 2013 were used as the training set data;

[0028] A total of 50 samples from September 2013 to April 2020 were used as the validation set data;

[0029] A total of 28 samples from May 2020 to August 2022 were used as the test set data.

[0030] Step 2: Provide training set data to the graph neural network for training;

[0031] In this step, during the training of the graph neural network in step 2, the graph neural network is based on the concept of a graph in graph theory, that is, a graph contains nodes, node features, edge connections, and edge features; each sub-region is treated as a graph node, whether the sub-regions are connected is treated as a graph connection, and the frequency features and the GR relationship of the sub-regions are treated as features on the graph nodes. The graph aggregation formula used is:

[0032]

[0033]

[0034] in Represents all graph nodes, i.e., each sub-region; x i The main node is the sub-region; x j Nodes that are connected to the master node are called neighboring nodes; x i+1 The update result for the master node; in the specific implementation, when using the above formula for calculation, each sub-region's representative node acts as the master node once, and also acts as a neighbor node multiple times; W1, W2, W3, and W4 are learnable parameters, specifically a matrix initialized with random numbers, which is then fitted and optimized through training, and the final result is obtained after training is completed; α i，j This represents the correlation between node features of two nodes calculated using learnable parameters, the master node, and neighboring nodes; d represents the length of the node feature.

[0035] Step 3: For each sub-region divided in Step 1, train a corresponding number of graph neural networks to predict whether an earthquake will occur in each sub-region.

[0036] In this step, since it is a multi-task, multi-classification problem, it is necessary to consider the balance of distribution in each region when making samples, so as to build a certain number of negative samples so that N networks can be trained to predict whether an earthquake will occur in each sub-region.

[0037] Once the graph neural network is trained, its output is a percentage. Therefore, setting different thresholds (a probability greater than 99% is required for an earthquake to occur) will yield different prediction results. When a specific threshold is selected, a concrete prediction result can be obtained, such as... Figure 3 The diagram shown is a schematic representation of the evaluation indicators provided in an embodiment of the present invention. Further, using... Figure 3 The evaluation metrics used in this study yield a specific prediction accuracy, including the accuracy rate and the false alarm rate. The specific formula is shown below:

[0038] Accuracy rate = Number of samples where an earthquake was predicted and actually occurred / Total number of samples where earthquakes occurred

[0039] False alarm rate = Number of samples where an earthquake was predicted but did not actually occur / Number of samples where no earthquake actually occurred

[0040] By using different accuracy and false alarm rates, receiver operating characteristic (ROC) curves can be plotted on the validation set, such as... Figure 4 The figure shown is a schematic diagram of the ROC curve of the trained network on the validation set according to an embodiment of the present invention. Figure 4 The ROC curve in the algorithm selects a threshold that balances the accuracy and false alarm rate for different sub-regions. Specifically, it calculates the accuracy and false alarm rate for all threshold settings by setting a step size, selects the accuracy and false alarm rate using the ROC curve, and then finds the corresponding threshold for this result. In other words, for a specific sub-region, setting a certain threshold ensures that the false alarm rate is below 5% while the accuracy reaches above 30%.

[0041] After the threshold selection is completed, the threshold will be used in subsequent testing. When the probability of an earthquake occurring output by the graph neural network is greater than the threshold, it is considered that an earthquake will occur; conversely, when the probability is less than the threshold, it is determined that an earthquake will not occur. This is how to determine whether an earthquake will occur in a certain sub-region.

[0042] For example, suppose an earthquake occurred in two out of eight months in region 1, and the method predicts one of those earthquakes, the accuracy rate is 50%. This also means that if no earthquake occurred in the region for six months, and the method predicts an earthquake in one of those months, the false alarm rate is 16.6%. Based on this, different thresholds will yield different prediction results: if the graph neural network predicts an earthquake with a 98% probability, and the threshold is set to predict an earthquake only if the probability is greater than or equal to 99%, then no earthquake will occur in that region; however, if the threshold is set to predict an earthquake only if the probability is greater than or equal to 98%, then an earthquake will occur in that region.

[0043] like Figure 5The diagram illustrates the test results of various sub-regions on the validation set in this embodiment of the invention. Different thresholds are set to determine whether an earthquake will occur, resulting in multiple sets of different predictions. Current test results show that, since different thresholds yield different predictions, and considering practical needs and the standards of the seismic network center, two different models can be obtained: a low false alarm scheme (25% accuracy and 8% false alarm rate) and a high accuracy scheme (35.7% accuracy and 14% false alarm rate). Changing the threshold increases both the accuracy and false alarm rates. Therefore, setting a suitable threshold can achieve an accuracy rate of 30% and limit the false alarm rate to around 10%.

[0044] Step 4: After completing the graph neural network training, input the earthquake data of the target area within one year into the graph neural network to predict whether an earthquake of magnitude 5 or above will occur in each sub-region of the target area within the next month.

[0045] In this step, the different prediction results of each sub-region of the target area are further integrated. Specifically, the graph neural network of each sub-region is used to predict whether an earthquake will occur in the target area. Then, the prediction results of each sub-region are integrated together as the final output result, which is the prediction result that an earthquake of magnitude 5 or above will occur in the target area within the next month.

[0046] In the specific implementation process, since earthquakes of magnitude 6 or above are relatively rare, the prediction of earthquakes of magnitude 5 or above is taken as the target of the method verification in this embodiment. Considering the number of earthquakes and the completeness of earthquake prediction, earthquakes of magnitude 3 or above are used to predict earthquakes of magnitude 5 or above.

[0047] like Figure 6 The diagram illustrates a comparison between the method described in this embodiment of the invention and a convolutional neural network prediction scheme in the prior art. The comparison shows that although the convolutional neural network scheme can achieve similar prediction results to the graph neural network of this application on the validation set data, it fails completely on the test set data. This is because the event interval between the training data and the test data is large, so the prediction accuracy of any scheme will decrease. Under these circumstances, the graph neural network scheme of this application can still maintain a certain accuracy, indicating that it has a certain degree of robustness.

[0048] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0049] In summary, the method described in this embodiment of the invention directly processes the earthquake catalog, thus eliminating the need for complex data feature extraction. Furthermore, based on the relationships between sub-regions, this method effectively utilizes previous research findings in the target region by employing graph neural networks, and leverages existing fault structures nationwide to enhance the network's interpretability. Compared to existing methods, the method described in this embodiment of the invention provides more accurate prediction results.

[0050] Furthermore, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0051] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. An earthquake prediction method based on graph neural networks, characterized in that, The method includes: Step 1: Perform sub-region division and earthquake catalog preprocessing on the target area to obtain frequency characteristics and GR relationship curves of sub-regions, and construct training set data for graph neural network training; Step 2: Provide training set data to the graph neural network for training; In step 2, during the training of the graph neural network, the graph neural network is based on the graph concept in graph theory. Each sub-region is treated as a graph node, the connectivity of the sub-regions is used as the graph's connection relationship, and the frequency features and the GR relationship of the sub-regions are used as features on the graph nodes. The graph aggregation formula used is: ； ； in, Represents all graph nodes; x i Master node; x j Nodes that are connected to the master node are called neighboring nodes; x i+1 The update result of the master node; in which each sub-region represents a node that acts as the master node once, and also acts as a neighbor node multiple times. W1, W2, W3, and W4 are learnable parameters, specifically a matrix initialized with random numbers. The matrix is ​​then fitted and optimized through training, and the final result is obtained after training is completed. This represents the correlation of node features between two nodes calculated using learnable parameters, the master node, and neighboring nodes. d Indicates the length of the node feature; Step 3: For each sub-region divided in Step 1, train a corresponding number of graph neural networks to predict whether an earthquake will occur in each sub-region. Step 4: After completing the graph neural network training, input the earthquake data of the target area within one year into the graph neural network to predict whether an earthquake of magnitude 5 or above will occur in each sub-region of the target area within the next month.

2. The earthquake prediction method based on graph neural networks according to claim 1, characterized in that, In step 1, the preprocessing process of the earthquake catalog is as follows: earthquakes of magnitude 3 or above are selected as input to the dataset, and earthquakes of magnitude 5 or above are selected as labels for the dataset. For each sub-region, the frequency of earthquakes occurring within 1 month, 3 months, 6 months, and 1 year is counted. The horizontal axis represents the magnitude of the earthquake, and the vertical axis represents the number of earthquakes of different magnitudes. The frequency is normalized by dividing the earthquake frequency in different time intervals by the corresponding time range. The Gutenberg-Richter (GR) relationship curves for each sub-region were calculated over 1 month, 3 months, 6 months, and 1 year, and used together with the frequency as training set data.

3. The earthquake prediction method based on graph neural networks according to claim 1, characterized in that, In step 1, the graph neural network is trained on a monthly basis, with the total data spanning 620 months from January 1971 to August 2022. Among them, samples in which no magnitude 5 earthquake occurred in the two months before and after were considered negative samples. After deleting data that was neither a positive nor a negative sample, all samples were divided into the following time periods: A total of 300 samples from January 1971 to August 2013 were used as the training set data; A total of 50 samples from September 2013 to April 2020 were used as the validation set data; A total of 28 samples from May 2020 to August 2022 were used as the test set data.

4. The earthquake prediction method based on graph neural networks according to claim 1, characterized in that, In step 3, after the graph neural network is trained, its output is a percentage. Therefore, setting different thresholds will yield different prediction results. When a specific threshold is selected, a specific prediction result can be obtained, thus generating the accuracy and false alarm rate. The specific formulas are as follows: Accuracy rate = Number of samples where an earthquake was predicted and actually occurred / Total number of samples where an earthquake occurred; False alarm rate = Number of samples where an earthquake was predicted but did not actually occur / Number of samples where no earthquake actually occurred; Then, using different accuracy and false alarm rates, receiver operating characteristic (ROC) curves are plotted on the validation set. Based on the ROC curves, a threshold that balances accuracy and false alarm rate is selected for different sub-regions. Specifically, by setting a step size, the accuracy and false alarm rate are calculated for all threshold settings. After selecting the accuracy and false alarm rate through the ROC curve, the corresponding threshold for this result is found. After the threshold selection is completed, the threshold will be used in subsequent testing. When the probability of an earthquake occurring output by the graph neural network is greater than the threshold, it is considered that an earthquake will occur; conversely, when the probability is less than the threshold, it is determined that an earthquake will not occur. This is how to determine whether an earthquake will occur in a certain sub-region.

5. The earthquake prediction method based on graph neural networks according to claim 1, characterized in that, In step 4, the different prediction results of each sub-region of the target area are integrated. Specifically, the graph neural network of each sub-region is used to predict whether an earthquake will occur in the target area. The prediction results of each sub-region are then integrated together as the final output result, which is the prediction result that an earthquake of magnitude 5 or above will occur in the target area within the next month.

Images