A method for classifying seismic events based on multi-station cooperation

Abstract

The application discloses a kind of based on multiple station cooperation's earthquake event classification method, comprising the following steps: one, different types of seismic waveform records and label are obtained by station;Two, different types of seismic waveform records are preprocessed, and short-time Fourier transform is carried out, and the time-frequency diagram corresponding to the seismic waveform record of all stations is obtained;Three, construct graph dataset;Four, construct CNN-GNN hybrid network model;Five, training set in graph dataset is used to train CNN-GNN hybrid network model, and the trained CNN-GNN hybrid network model is obtained;Six, based on the trained CNN-GNN hybrid network model, subsequent to be measured earthquake event is classified.The method of the application is reasonable in design, uses CNN-GNN hybrid network model, improves classification performance by using the time-frequency diagram of multiple stations, while maintaining the flexibility of the number of stations used, solves the limitations of existing earthquake classification model.

Description

Technical Field

[0001] This invention belongs to the field of earthquake event classification technology, specifically relating to a method for earthquake event classification based on multi-station collaboration. Background Technology

[0002] In recent years, the rapid growth of earthquake monitoring data and computing power has made data-driven deep learning methods a powerful approach for event classification. CNN-based methods typically run on a single station and then require post-processing to obtain event-level classifications. Majority voting and probabilistic aggregation are two commonly used post-processing strategies. However, majority voting has limitations, requiring an odd number of stations and ignoring predicted probabilities. On the other hand, probabilistic aggregation risks being dominated by a few incorrectly identified stations, which could lead to event-level misclassifications. Furthermore, analysts often examine waveform consistency across multiple stations in practice to determine event types. For example, a classic criterion is that the first motion polarity in all azimuths around the P-wave source typically generated by a mine explosion (QB) is upward, while those with mixed polarities are considered EQ. Therefore, merging waveform data from multiple stations holds promise for improving the performance of classification models.

[0003] A common approach to multi-site classification is to overlay station records onto a fixed-dimensional input. However, this method requires a fixed number of stations. For example, currently, raw waveforms from four stations are stacked together, and a CNN is used to distinguish between natural earthquake EQ and mining explosion QB; subsequently, CNN features extracted from five stations are concatenated to classify volcanic seismic events. Another approach focuses on aggregating features from multiple stations into a unified event representation. This aggregation method explicitly models seismic events as a graph. Then, a graph neural network aggregates features by directly passing information between stations. However, because their models rely on graph convolutional network layers that require a fixed graph structure, they are still limited by a fixed number of stations. In summary, existing multi-site classification models are constrained by a fixed number of input stations. Therefore, they have poor versatility across different network configurations, are unapplicable when there are few stations, and fail to fully utilize additional records when there are many available stations.

[0004] Therefore, a seismic event classification method based on multi-station collaboration was designed, which is not limited by the number of fixed stations. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a multi-station collaborative earthquake event classification method to address the shortcomings of the prior art. The method has simple steps and reasonable design. It adopts a CNN-GNN hybrid network model and improves classification performance by utilizing earthquake time-frequency maps from multiple stations, while maintaining the flexibility of the number of stations used, thus overcoming the limitations of existing earthquake classification models.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a seismic event classification method based on multi-station collaboration, characterized in that the method includes the following steps: Step 1: Obtain different types of seismic waveform records and tags through the stations; wherein, a seismic event is received by multiple stations, corresponding to multiple seismic waveform records; each seismic event received by each station corresponds to one seismic waveform record, and the tags include natural earthquake EQ and mine explosion QB; Step 2: Preprocess the seismic waveform records of different types and perform short-time Fourier transform to obtain the time-frequency diagrams corresponding to the seismic waveform records of all stations; Step 3: Construct a graph dataset; Step 4: Construct a CNN-GNN hybrid network model; Step 5: Train the CNN-GNN hybrid network model using the training set in the graph dataset to obtain the trained CNN-GNN hybrid network model; Step 6: Classify subsequent earthquake events based on the trained CNN-GNN hybrid network model.

[0007] The above-mentioned multi-station collaborative seismic event classification method, further, in step two, the specific process is as follows: Step 201: Preprocess the seismic wave signal of any seismic waveform record to obtain the preprocessed seismic wave signal; wherein, the preprocessing includes removing the mean, removing the linear trend, removing the pinch-out, and bandpass filtering; Step 202: Resample the preprocessed seismic wave signal to the set sampling rate to obtain the resampled seismic wave signal; Step 203: Normalize the amplitude of the resampled seismic wave signal to obtain the normalized seismic wave signal; Step 204: Extract the normalized seismic wave signal to obtain the extracted seismic wave signal; Step 205: Convert the intercepted seismic wave signal to the ZRT coordinate system to obtain three component signals; Step 206: Set the window overlap rate to 12%, the window length to 2.50 seconds, and the window type to Hanning window. Perform short-time Fourier transform on the three component signals respectively to obtain three time-frequency diagrams. Step 207: Repeat steps 201 to 206 multiple times to obtain the time-frequency diagrams corresponding to the seismic waveform records of all stations.

[0008] The above-mentioned multi-station collaborative seismic event classification method, further, in step three, the specific process is as follows: Step 301: Construct a graph dataset based on the time-frequency graphs of multiple stations; wherein, one earthquake event corresponds to one graph dataset, and each graph dataset includes nodes and edges. The nodes are the stations that receive the earthquake event signal. If the distance between two stations is less than a set distance threshold, the nodes corresponding to the two stations are connected by an edge, and the weight of the edge is the reciprocal of the distance between the two stations. Step 302: Repeat step 301 to generate a graph dataset, and divide the graph dataset into a training set, a validation set, and a test set according to the chronological order of the earthquake events.

[0009] The above-mentioned multi-station collaborative seismic event classification method further includes step four, the specific process of which is as follows: The CNN-GNN hybrid network model includes a feature extraction module and a classification module. The feature extraction module includes a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a Flatten layer, and a linear layer. The first convolutional block, the second convolutional block, the third convolutional block, and the fourth convolutional block have the same structure and each includes a convolutional layer, a batch normalization layer, a ReLU activation function layer, a max pooling layer, and a dropout layer. The classification module includes a GATv2Conv layer, a layer normalization layer, a Dropout layer, an attention-focusing layer, and a second linear layer; The attention-focusing layer includes a first linear layer, a ReLU activation function layer, and a Softmax layer.

[0010] The above-mentioned multi-station collaborative seismic event classification method, further, step five, is as follows: Step 501: Input the time-frequency graph of each node in the same graph data in the training set into the CNN-GNN hybrid network model for feature extraction and classification, and output a 2D tensor corresponding to the graph data; Step 502, according to The cross-entropy loss is obtained. ;in, This represents the first element in the 2D tensor corresponding to the data in the i′ graph. Let represent the second element in the 2D tensor corresponding to the i′-th image data. If the true label corresponding to the i′-th image data is a natural earthquake, then... The value is 1; if the true label corresponding to the data in the i′th image is "mining explosion", then... The value is 0; Let i' and i' represent the total number of graph data in the training set. The integer is a positive integer, and 1 ≤ i′ ≤ ; Step 503: The computer uses the Adam optimizer and cross-entropy loss to train the CNN-GNN hybrid network model until the entire training set is trained, completing one iteration of training. Step 504: Repeat steps 501 to 503 for iterative training until the maximum number of training rounds is met, and obtain the trained CNN-GNN hybrid network model.

[0011] The above-mentioned multi-station collaborative seismic event classification method further includes step 501, which is as follows: Step 5011: Input the time-frequency graph of each node in the same graph data in the training set into the feature extraction module for feature extraction to obtain the feature vector of each node; Step 5012: Input the feature vectors corresponding to multiple nodes into the GATv2Conv layer for processing to obtain the updated feature vector for each node. Step 5013: Perform layer normalization on the updated feature vector corresponding to each node to obtain the normalized feature vector corresponding to each node. Step 5014: Process the normalized feature vector corresponding to each node through the Dropout layer to obtain the final feature vector corresponding to each node; Step 5015: Process the final feature vector corresponding to each node through the first linear layer, then according to... The non-standardized importance score corresponding to the i-th node is obtained. ;in, This represents the weight vector of the linear layer. This represents the bias term of the linear layer; This represents the final feature vector corresponding to the i-th node; Step 5016: Calculate the non-standardized importance score corresponding to the i-th node. After passing through the ReLU activation function layer, the importance score update value corresponding to the i-th node is obtained. ; Step 5017: Repeat steps 5015 and 5016 multiple times to obtain the updated importance scores for all nodes. Step 5018: Process the importance score update values ​​corresponding to all nodes through a Softmax layer to obtain the attention weight corresponding to each node; where the attention weight corresponding to the i-th node is denoted as... ; Step 5019, according to This yields the graph-level representation vector. Where N represents the total number of nodes in the same graph, i and N are both positive integers, and 1≤i≤N; Step 501A: Convert the graph-level representation vector The input is processed by the second linear layer to obtain a 2D tensor.

[0012] The above-mentioned multi-station collaborative seismic event classification method, further, step 5012, is as follows: Step A1: For each node in the same graph, according to The attention coefficients between the i-th node and the j-th node are obtained. ; ( () represents an exponential function with the natural constant as its base. ( () indicates the LeakReLU function operation. This indicates the concat concatenation operation. This represents the attention vector of the h-th head in the GATv2Conv layer. This represents the weight matrix of the h-th head in the GATv2Conv layer. This represents the feature vector of the i-th node. This represents the weight of the edge between the i-th node and the j-th node; Let T represent the set of neighbors of all nodes that have an edge with the i-th node; T represents the transpose; k represents The kth node in This represents the feature vector of the k-th node. The weight of the edge between the i-th node and the k-th node is represented by j; The j-th node in; Let h represent the feature vector of the j-th node; h is a positive integer and its value ranges from 1 to H. Step A2, according to We obtain the updated feature vector of the i-th node; where, This indicates the concat concatenation operation. This represents the exponential linear unit ELU activation function.

[0013] The above-mentioned multi-station collaborative seismic event classification method, further, step 5011, is as follows: Step B1: Input the time-frequency map of each node in the training set into the first convolutional block for processing to obtain the first feature map; Step B2: Input the first feature map into the second convolutional block for processing to obtain the second feature map; Step B3: Input the second feature map into the third convolutional block for processing to obtain the third feature map; Step B4: Input the third feature map into the fourth convolutional block for processing to obtain the fourth feature map; Step B5: Input the fourth feature map into the Flatten layer for processing to obtain a one-dimensional tensor; Step B6: Process the one-dimensional tensor through a linear layer to obtain the feature vector of each node.

[0014] Compared with the prior art, the present invention has the following advantages: 1. The method of this invention has simple steps and a reasonable design. It solves the problem that most current CNNs operate on a single station, and post-processing is usually required to obtain event categories. This method directly obtains event categories end-to-end without the need for post-processing strategies. Furthermore, since most CNNs operate on a single station and cannot utilize the collaborative relationships between multiple stations, this invention utilizes the collaborative relationships between multiple stations, further improving classification performance.

[0015] 2. Currently, even though some deep learning methods utilize multiple stations for classification, these methods require a fixed number of stations in advance. If the number of stations is less than the required number, the model cannot be applied; if the number of stations is greater than the required number, the additional information cannot be used, resulting in waste. The method of this invention overcomes this limitation, as it has no constraint on the number of stations, making it flexible.

[0016] 3. When constructing graph data, this invention considers the distance relationship between stations and explicitly incorporates the layout of stations into the graph data, thereby enabling collaborative analysis using multiple stations.

[0017] 4. In the earthquake event classification module, this invention uses the GATv2Conv layer of the graph attention network in graph neural network for classification, which has no limit on the number of stations, thus making the CNN-GNN hybrid network model flexibly adaptable to any station network configuration.

[0018] In summary, the method of this invention is simple in steps and reasonable in design. It adopts a CNN-GNN hybrid network model and improves classification performance by utilizing seismic time-frequency maps from multiple stations, while maintaining the flexibility of the number of stations used, thus overcoming the limitations of existing seismic classification models.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 This is a schematic diagram of the feature extraction module of the present invention.

[0022] Figure 3 This is a schematic diagram of the classification module of the present invention. Detailed Implementation

[0023] like Figure 1 As shown, the present invention provides a seismic event classification method based on multi-station collaboration, comprising the following steps: Step 1: Obtain different types of seismic waveform records and tags through the stations; wherein, a seismic event is received by multiple stations, corresponding to multiple seismic waveform records; each seismic event received by each station corresponds to one seismic waveform record, and the tags include natural earthquake EQ and mine explosion QB; Step 2: Preprocess the seismic waveform records of different types and perform short-time Fourier transform to obtain the time-frequency diagrams corresponding to the seismic waveform records of all stations; Step 3: Construct a graph dataset; Step 4: Construct a CNN-GNN hybrid network model; Step 5: Train the CNN-GNN hybrid network model using the training set in the graph dataset to obtain the trained CNN-GNN hybrid network model; Step 6: Classify subsequent earthquake events based on the trained CNN-GNN hybrid network model.

[0024] In this embodiment, step two is as follows: Step 201: Preprocess the seismic wave signal of any seismic waveform record to obtain the preprocessed seismic wave signal; wherein, the preprocessing includes removing the mean, removing the linear trend, removing the pinch-out, and bandpass filtering; Step 202: Resample the preprocessed seismic wave signal to the set sampling rate to obtain the resampled seismic wave signal; Step 203: Normalize the amplitude of the resampled seismic wave signal to obtain the normalized seismic wave signal; Step 204: Extract the normalized seismic wave signal to obtain the extracted seismic wave signal; Step 205: Convert the intercepted seismic wave signal to the ZRT coordinate system to obtain three component signals; Step 206: Set the window overlap rate to 12%, the window length to 2.50 seconds, and the window type to Hanning window. Perform short-time Fourier transform on the three component signals respectively to obtain three time-frequency diagrams. Step 207: Repeat steps 201 to 206 multiple times to obtain the time-frequency diagrams corresponding to the seismic waveform records of all stations.

[0025] In this embodiment, step three is as follows: Step 301: Construct a graph dataset based on the time-frequency graphs of multiple stations; wherein, one earthquake event corresponds to one graph dataset, and each graph dataset includes nodes and edges. The nodes are the stations that receive the earthquake event signal. If the distance between two stations is less than a set distance threshold, the nodes corresponding to the two stations are connected by an edge, and the weight of the edge is the reciprocal of the distance between the two stations. Step 302: Repeat step 301 to generate a graph dataset, and divide the graph dataset into a training set, a validation set, and a test set according to the chronological order of the earthquake events.

[0026] like Figure 2 and Figure 3 As shown in the figure, in this embodiment, step four is specifically performed as follows: The CNN-GNN hybrid network model includes a feature extraction module and a classification module. The feature extraction module includes a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a Flatten layer, and a linear layer. The first convolutional block, the second convolutional block, the third convolutional block, and the fourth convolutional block have the same structure and each includes a convolutional layer, a batch normalization layer, a ReLU activation function layer, a max pooling layer, and a dropout layer. The classification module includes a GATv2Conv layer, a layer normalization layer, a Dropout layer, an attention-focusing layer, and a second linear layer; The attention-focusing layer includes a first linear layer, a ReLU activation function layer, and a Softmax layer.

[0027] In this embodiment, step five is as follows: Step 501: Input the time-frequency graph of each node in the same graph data in the training set into the CNN-GNN hybrid network model for feature extraction and classification, and output a 2D tensor corresponding to the graph data; Step 502, according to The cross-entropy loss is obtained. ;in, This represents the first element in the 2D tensor corresponding to the data in the i′ graph. Let represent the second element in the 2D tensor corresponding to the i′-th image data. If the true label corresponding to the i′-th image data is a natural earthquake, then... The value is 1; if the true label corresponding to the data in the i′th image is "mining explosion", then... The value is 0; Let i' and i' represent the total number of graph data in the training set. The integer is a positive integer, and 1 ≤ i′ ≤ ; Step 503: The computer uses the Adam optimizer and cross-entropy loss to train the CNN-GNN hybrid network model until the entire training set is trained, completing one iteration of training. Step 504: Repeat steps 501 to 503 for iterative training until the maximum number of training rounds is met, and obtain the trained CNN-GNN hybrid network model.

[0028] In this embodiment, step 501 is specifically performed as follows: Step 5011: Input the time-frequency graph of each node in the same graph data in the training set into the feature extraction module for feature extraction to obtain the feature vector of each node; Step 5012: Input the feature vectors corresponding to multiple nodes into the GATv2Conv layer for processing to obtain the updated feature vector for each node. Step 5013: Perform layer normalization on the updated feature vector corresponding to each node to obtain the normalized feature vector corresponding to each node. Step 5014: Process the normalized feature vector corresponding to each node through the Dropout layer to obtain the final feature vector corresponding to each node; Step 5015: Process the final feature vector corresponding to each node through the first linear layer, then according to... The non-standardized importance score corresponding to the i-th node is obtained. ;in, This represents the weight vector of the linear layer. This represents the bias term of the linear layer; This represents the final feature vector corresponding to the i-th node; Step 5016: Calculate the non-standardized importance score corresponding to the i-th node. After passing through the ReLU activation function layer, the importance score update value corresponding to the i-th node is obtained. ; Step 5017: Repeat steps 5015 and 5016 multiple times to obtain the updated importance scores for all nodes. Step 5018: Process the importance score update values ​​corresponding to all nodes through a Softmax layer to obtain the attention weight corresponding to each node; where the attention weight corresponding to the i-th node is denoted as... ; Step 5019, according to This yields the graph-level representation vector. Where N represents the total number of nodes in the same graph, i and N are both positive integers, and 1≤i≤N; Step 501A: Convert the graph-level representation vector The input is processed by the second linear layer to obtain a 2D tensor.

[0029] In this embodiment, step 5012 is specifically performed as follows: Step A1: For each node in the same graph, according to The attention coefficients between the i-th node and the j-th node are obtained. ; ( () represents an exponential function with the natural constant as its base. ( () indicates the LeakReLU function operation. This indicates the concat concatenation operation. This represents the attention vector of the h-th head in the GATv2Conv layer. This represents the weight matrix of the h-th head in the GATv2Conv layer. This represents the feature vector of the i-th node. This represents the weight of the edge between the i-th node and the j-th node; Let T represent the set of neighbors of all nodes that have an edge with the i-th node; T represents the transpose; k represents The kth node in This represents the feature vector of the k-th node. The weight of the edge between the i-th node and the k-th node is represented by j; The j-th node in; Let h represent the feature vector of the j-th node; h is a positive integer and its value ranges from 1 to H. Step A2, according to We obtain the updated feature vector of the i-th node; where, This indicates the concat concatenation operation. This represents the exponential linear unit ELU activation function.

[0030] In this embodiment, step 5011 is specifically performed as follows: Step B1: Input the time-frequency map of each node in the training set into the first convolutional block for processing to obtain the first feature map; Step B2: Input the first feature map into the second convolutional block for processing to obtain the second feature map; Step B3: Input the second feature map into the third convolutional block for processing to obtain the third feature map; Step B4: Input the third feature map into the fourth convolutional block for processing to obtain the fourth feature map; Step B5: Input the fourth feature map into the Flatten layer for processing to obtain a one-dimensional tensor; Step B6: Process the one-dimensional tensor through a linear layer to obtain the feature vector of each node.

[0031] In this embodiment, the sampling rate set in step 202 is 100Hz; the distance threshold set in step 301 is 300km; these can be adjusted according to actual requirements.

[0032] In this embodiment, specifically, in step 503, the optimizer learning rate is set to 0.001, the learning rate decay parameter is set to 0.004, the batch size for model training is set to 124, and the maximum number of training epochs in step 503 is 100.

[0033] In this embodiment, the size of the feature vector of each node in step 5011 is 64×1. In step 5012, the size of the updated feature vector corresponding to the node is 128×1; the number of heads H of the GATv2Conv layer is set to 2; In step 5014, the final feature vector corresponding to the node has a size of 128×1; the probability P in the Dropout layer is set to 0.2.

[0034] In this embodiment, Linville (2021) published a catalog of seismic events in Utah and its surrounding areas from October 2012 to October 2020. This catalog contains 6,881 earthquake EQs, 9,145 mine explosion QBs, and 220 explosions, all of which were manually reviewed and tagged by analysts. To construct our dataset, we selected earthquake EQs and mine explosion QBs from the catalog and downloaded the corresponding waveforms from the EarthScope Consortium-operated Seismic Data Service using the provided P-wave arrival and station metadata. Continuous waveforms were downloaded using a 130-second duration defined by P-wave arrival (from the preceding 10 seconds to the following 120 seconds) to facilitate subsequent truncation. The dataset consists of 181,470 seismic waveform records (only the vertical component was counted for each station; this rule applies to all reported waveform counts), including 89,348 of the 6,877 EQs and 92,122 of the 9,141 QBs.

[0035] In this embodiment, the length of the seismic wave signal intercepted in step 204 is 10 seconds before the arrival of the P-wave and 80 seconds after the arrival of the P-wave, for a total of 90 seconds.

[0036] In this embodiment, the graph dataset is segmented according to the earthquake event level and in chronological order. Specifically, the graph dataset constructed from earthquake events on or before December 31, 2018, constitutes the training set; the graph dataset constructed from earthquake events that occurred in 2019 constitutes the validation set; and the graph dataset constructed from earthquake events on or after January 1, 2020, constitutes the test set. The process for constructing the validation and test sets can refer to the process for constructing the training set.

[0037] The training set contains 5,685 EQ graphs and 7,399 QB graphs for training; the validation set contains 932 EQ graphs and 878 QB graphs for validation; and the test set contains 260 EQ graphs and 864 QB graphs for testing.

[0038] In this embodiment, it should be noted that mean removal, linear trend removal, sharpening removal, and bandpass filtering are all conventional preprocessing methods.

[0039] In this embodiment, it should be noted that the size of the feature map in step 5011 is represented by the number of channels × width × height, and the time-frequency map [C, H0, W0] describes the shape of the input tensor with C components, frequency dimension H0 and time dimension W0.

[0040] The time-frequency plot is 3×48×40, the first feature map is 18×24×20, the second feature map is 36×12×10, the third feature map is 54×6×5, and the fourth feature map is 54×3×2; the size of the one-dimensional tensor is 324×1. In this embodiment, it should be noted that the size of the convolution kernel in the convolutional layer of the first convolutional block is 2×2, the number of convolution kernels is 18, the stride is 1, and the padding is "same". The second convolutional block has a 2×2 kernel size, 36 kernels, a stride of 1, and padding of "same". The third convolutional block has a 2×2 kernel size, 54 kernels, a stride of 1, and padding of "same". The third convolutional block has a 2×2 kernel size, 54 kernels, a stride of 1, and padding of "same". In this embodiment, it should be noted that the size of the max pooling layer is 2×2, the step size is 2, and the fill size is 0.

[0041] In this embodiment, Table 1 shows the performance metrics of the model of the present invention on the test set of the graph dataset. (The test set contains 260 EQ graphs and 864 QB graphs for testing). Therefore, the method of the present invention is superior.

[0042] Table 1. Performance metrics of the model of this invention on the graph dataset test set.

[0043] In summary, the method of this invention is simple in steps and reasonable in design. It adopts a CNN-GNN hybrid network model and improves classification performance by utilizing seismic time-frequency maps from multiple stations, while maintaining the flexibility of the number of stations used, thus overcoming the limitations of existing seismic classification models.

[0044] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A seismic event classification method based on multi-station collaboration, characterized in that, The method includes the following steps: Step 1: Obtain different types of seismic waveform records and tags through the stations; wherein, a seismic event is received by multiple stations, corresponding to multiple seismic waveform records; each seismic event received by each station corresponds to one seismic waveform record, and the tags include natural earthquake EQ and mine explosion QB; Step 2: Preprocess the seismic waveform records of different types and perform short-time Fourier transform to obtain the time-frequency diagrams corresponding to the seismic waveform records of all stations; Step 3: Construct a graph dataset; Step 4: Construct a CNN-GNN hybrid network model; Step 5: Train the CNN-GNN hybrid network model using the training set in the graph dataset to obtain the trained CNN-GNN hybrid network model; Step 6: Classify subsequent earthquake events based on the trained CNN-GNN hybrid network model.

2. The seismic event classification method based on multi-station collaboration according to claim 1, characterized in that: Step two, the specific process is as follows: Step 201: Preprocess the seismic wave signal of any seismic waveform record to obtain the preprocessed seismic wave signal; wherein, the preprocessing includes removing the mean, removing the linear trend, removing the pinch-out, and bandpass filtering; Step 202: Resample the preprocessed seismic wave signal to the set sampling rate to obtain the resampled seismic wave signal; Step 203: Normalize the amplitude of the resampled seismic wave signal to obtain the normalized seismic wave signal; Step 204: Extract the normalized seismic wave signal to obtain the extracted seismic wave signal; Step 205: Convert the intercepted seismic wave signal to the ZRT coordinate system to obtain three component signals; Step 206: Set the window overlap rate to 12%, the window length to 2.50 seconds, and the window type to Hanning window. Perform short-time Fourier transform on the three component signals respectively to obtain three time-frequency diagrams. Step 207: Repeat steps 201 to 206 multiple times to obtain the time-frequency diagrams corresponding to the seismic waveform records of all stations.

3. The seismic event classification method based on multi-station collaboration according to claim 2, characterized in that: Step three, the specific process is as follows: Step 301: Construct a graph dataset based on the time-frequency graphs of multiple stations; wherein, one earthquake event corresponds to one graph dataset, and each graph dataset includes nodes and edges. The nodes are the stations that receive the earthquake event signal. If the distance between two stations is less than a set distance threshold, the nodes corresponding to the two stations are connected by an edge, and the weight of the edge is the reciprocal of the distance between the two stations. Step 302: Repeat step 301 to generate a graph dataset, and divide the graph dataset into a training set, a validation set, and a test set according to the chronological order of the earthquake events.

4. A seismic event classification method based on multi-station collaboration as described in claim 3, characterized in that: Step four, the specific process is as follows: The CNN-GNN hybrid network model includes a feature extraction module and a classification module. The feature extraction module includes a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a Flatten layer, and a linear layer. The first convolutional block, the second convolutional block, the third convolutional block, and the fourth convolutional block have the same structure and each includes a convolutional layer, a batch normalization layer, a ReLU activation function layer, a max pooling layer, and a dropout layer. The classification module includes a GATv2Conv layer, a layer normalization layer, a Dropout layer, an attention-focusing layer, and a second linear layer; The attention-focusing layer includes a first linear layer, a ReLU activation function layer, and a Softmax layer.

5. A seismic event classification method based on multi-station collaboration according to claim 4, characterized in that: Step five, the specific process is as follows: Step 501: Input the time-frequency graph of each node in the same graph data in the training set into the CNN-GNN hybrid network model for feature extraction and classification, and output a 2D tensor corresponding to the graph data; Step 502, according to The cross-entropy loss is obtained. ;in, This represents the first element in the 2D tensor corresponding to the data in the i′ graph. Let represent the second element in the 2D tensor corresponding to the i′-th image data. If the true label corresponding to the i′-th image data is a natural earthquake, then... The value is 1; if the true label corresponding to the data in the i′th image is "mining explosion", then... The value is 0; Let i' and i' represent the total number of graph data in the training set. The integer is a positive integer, and 1 ≤ i′ ≤ ; Step 503: The computer uses the Adam optimizer and cross-entropy loss to train the CNN-GNN hybrid network model until the entire training set is trained, completing one iteration of training. Step 504: Repeat steps 501 to 503 for iterative training until the maximum number of training rounds is met, and obtain the trained CNN-GNN hybrid network model.

6. A seismic event classification method based on multi-station collaboration according to claim 5, characterized in that: Step 501, the specific process is as follows: Step 5011: Input the time-frequency graph of each node in the same graph data in the training set into the feature extraction module for feature extraction to obtain the feature vector of each node; Step 5012: Input the feature vectors corresponding to multiple nodes into the GATv2Conv layer for processing to obtain the updated feature vector for each node. Step 5013: Perform layer normalization on the updated feature vector corresponding to each node to obtain the normalized feature vector corresponding to each node. Step 5014: Process the normalized feature vector corresponding to each node through the Dropout layer to obtain the final feature vector corresponding to each node; Step 5015: Process the final feature vector corresponding to each node through the first linear layer, then according to... The non-standardized importance score corresponding to the i-th node is obtained. ;in, This represents the weight vector of the linear layer. This represents the bias term of the linear layer; This represents the final feature vector corresponding to the i-th node; Step 5016: Calculate the non-standardized importance score corresponding to the i-th node. After passing through the ReLU activation function layer, the importance score update value corresponding to the i-th node is obtained. ; Step 5017: Repeat steps 5015 and 5016 multiple times to obtain the updated importance scores for all nodes. Step 5018: Process the importance score update values ​​corresponding to all nodes through a Softmax layer to obtain the attention weight corresponding to each node; where the attention weight corresponding to the i-th node is denoted as... ; Step 5019, according to This yields the graph-level representation vector. Where N represents the total number of nodes in the same graph, i and N are both positive integers, and 1≤i≤N; Step 501A: Convert the graph-level representation vector The input is processed by the second linear layer to obtain a 2D tensor.

7. A seismic event classification method based on multi-station collaboration according to claim 6, characterized in that: Step 5012, the specific process is as follows: Step A1: For each node in the same graph, according to The attention coefficients between the i-th node and the j-th node are obtained. ; ( () represents an exponential function with the natural constant as its base. ( () indicates the LeakReLU function operation. This indicates the concat concatenation operation. This represents the attention vector of the h-th head in the GATv2Conv layer. This represents the weight matrix of the h-th head in the GATv2Conv layer. This represents the feature vector of the i-th node. This represents the weight of the edge between the i-th node and the j-th node; Let T represent the set of neighbors of all nodes that have an edge with the i-th node; T represents the transpose; k represents The kth node in This represents the feature vector of the k-th node. The weight of the edge between the i-th node and the k-th node is represented by j; The j-th node in; Let h represent the feature vector of the j-th node; h is a positive integer and its value ranges from 1 to H. Step A2, according to We obtain the updated feature vector of the i-th node; where, This indicates the concat concatenation operation. This represents the exponential linear unit ELU activation function.

8. A seismic event classification method based on multi-station collaboration according to claim 6, characterized in that: Step 5011, the specific process is as follows: Step B1: Input the time-frequency map of each node in the training set into the first convolutional block for processing to obtain the first feature map; Step B2: Input the first feature map into the second convolutional block for processing to obtain the second feature map; Step B3: Input the second feature map into the third convolutional block for processing to obtain the third feature map; Step B4: Input the third feature map into the fourth convolutional block for processing to obtain the fourth feature map; Step B5: Input the fourth feature map into the Flatten layer for processing to obtain a one-dimensional tensor; Step B6: Process the one-dimensional tensor through a linear layer to obtain the feature vector of each node.

Images