Earthquake Magnitude Detection Method and System Based on Deep Learning and Multi-Feature Fusion
By employing a multi-feature fusion earthquake magnitude detection method, deep learning technology is used to preprocess and extract features from earthquake waveform data. Combined with multiple network models, high-precision, one-time magnitude detection is achieved, solving the problems of large dispersion and low accuracy in magnitude calculation results in existing technologies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTHEASTERN UNIV CHINA
- Filing Date
- 2023-07-06
- Publication Date
- 2026-05-26
AI Technical Summary
Existing deep learning methods suffer from problems such as large dispersion of results, low accuracy, and the need for multiple revisions in earthquake magnitude calculation, making it difficult to achieve accurate and timely magnitude calculation.
A deep learning-based multi-feature fusion method is adopted. The original three-component seismic waveform data is preprocessed and combined with short-time Fourier transform, recursive graph and S-transform feature extraction networks. The time dependence is captured by Bi-LSTM network and one-time magnitude detection is achieved through spatial attention module and fully connected layer.
It improves the accuracy of earthquake magnitude calculation, reduces human intervention, can accurately calculate the magnitude in one go, reduces errors, and improves the reliability of the calculation.
Smart Images
Figure CN116859459B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning technology, and in particular to an earthquake magnitude detection method and system based on deep learning and multi-feature fusion. Background Technology
[0002] Earthquakes are among the most destructive natural disasters, occurring instantaneously and causing immense loss of life. The destructive power of an earthquake is related to its magnitude, and determining the magnitude significantly impacts subsequent scientific research and forecasting. After an earthquake occurs, the released energy propagates in all directions as waves. Seismic waves are mainly divided into two categories: body waves, which propagate within the Earth and can be further subdivided into P-waves (longitudinal waves) and S-waves (transverse waves), and surface waves, which propagate on the Earth's surface. The propagation speed of each type of seismic wave is limited by the medium, with P-waves being the fastest, followed by S-waves, and surface waves being the slowest. However, their destructive power is opposite: surface waves cause the most damage, followed by S-waves, and P-waves cause the least.
[0003] Earthquake magnitude determination has long been affected by factors such as earthquake focal depth, geological propagation medium, and seismic wave receiving instruments, making it difficult to achieve a unified standard. Different seismic stations, due to variations in their location, attenuation and response to seismic waves, and the different frequency bands of their seismic wave receiving instruments, will measure magnitudes differently, with significant variations in some regions. The earthquake magnitude calculation process requires personnel to possess extensive theoretical knowledge of seismology and accumulated practical experience; this method is highly susceptible to subjective factors, resulting in less than satisfactory accuracy of the analysis results.
[0004] With the rapid advancement of deep learning technology, deep learning methods have been widely applied in the field of seismology. Although current methods can roughly determine the magnitude of an earthquake, they generally suffer from large dispersion, resulting in low accuracy. Even some methods with higher accuracy often require multiple revisions of the magnitude calculation results and cannot provide accurate earthquake magnitude calculation results in a timely manner. Summary of the Invention
[0005] The problem this invention aims to solve is to address the shortcomings of the existing technology by providing a deep learning-based multi-feature fusion method and system for earthquake magnitude detection, thereby resolving the issues of the need for multiple revisions of magnitude calculation results and low accuracy in magnitude magnitude calculation in current deep learning methods.
[0006] To achieve the above-mentioned objectives of the present invention, a first aspect of the present invention provides an earthquake magnitude detection method based on deep learning and multi-feature fusion, the method comprising the following steps:
[0007] Step 1: Acquire raw three-component seismic waveform data;
[0008] Step 2 involves preprocessing the original three-component seismic waveform data to obtain preprocessed three-component seismic waveform data.
[0009] Step 3: Obtain the training set from the preprocessed three-component seismic waveform data;
[0010] Step 4: Build a magnitude detection model;
[0011] Step 5: Train the model until the set number of training rounds is reached, and save the best-performing model from the trained rounds as the final magnitude detection model.
[0012] Step 6 involves preprocessing the three-component waveform data of the earthquake to be detected and then inputting it into the final magnitude detection model to obtain the magnitude detection result.
[0013] Furthermore, according to the earthquake magnitude detection method based on deep learning and multi-feature fusion, the method for preprocessing the original three-component earthquake waveform data is as follows: removing the three-component earthquake waveform data containing fragments or with a signal-to-noise ratio lower than a set threshold; dividing the three-component earthquake waveform data by the sensitivity of the seismic wave receiver to correct the error caused by the instrument's own influencing factors; and then performing Butterworth filtering, mean denoting, and normalization operations on the three-component earthquake waveform data.
[0014] Furthermore, according to the earthquake magnitude detection method based on deep learning and multi-feature fusion, step 4 specifically includes the following steps:
[0015] Step 4.1 Input the training set into the feature extraction network based on short-time Fourier transform, the feature extraction network based on recursive graph, and the feature extraction network based on S-transform respectively. After the three feature extraction networks extract the corresponding two-dimensional tensors based on prior knowledge, the widening convolution module WidenConv is used to further extract features at different scales and output the extracted Fourier transform feature tensor, recursive graph feature tensor, and S-transform feature tensor.
[0016] Step 4.2 Input the training set into the Bi-LSTM network and output the extracted Bi-LSTM feature tensor;
[0017] Step 4.3 merge the three feature tensors generated in Step 4.1 into a single merged feature tensor and input it into the spatial attention module to obtain a spatial attention feature tensor;
[0018] Step 4.4: Concatenate the Bi-LSTM feature tensor obtained in Step 4.2 with the spatial attention feature tensor obtained in Step 4.3 to obtain the final feature tensor and input it into the fully connected layer.
[0019] A second aspect of the present invention provides an earthquake magnitude detection system based on deep learning and multi-feature fusion, the system comprising: a data preprocessing module and an earthquake magnitude detection module.
[0020] Furthermore, according to the aforementioned earthquake magnitude detection system based on deep learning and multi-feature fusion, the data preprocessing module is used to preprocess the original three-component earthquake waveform data and send the preprocessed three-component earthquake waveform data to the earthquake magnitude detection module.
[0021] Furthermore, according to the aforementioned earthquake magnitude detection system based on deep learning and multi-feature fusion, the earthquake magnitude detection module is used to detect the corresponding magnitude based on the pre-processed earthquake three-component waveform data and output the magnitude detection result.
[0022] Furthermore, according to the aforementioned earthquake magnitude detection system based on deep learning and multi-feature fusion, the earthquake magnitude detection module further includes:
[0023] A feature extraction network based on short-time Fourier transform is used to connect the time and frequency domains of the pre-processed three-component seismic waveform data through short-time Fourier transform, reflecting the two-dimensional features of the three-component seismic waveform data in the time and frequency domains, and obtaining a two-dimensional Fourier transform tensor. The two-dimensional Fourier transform tensor is then used to extract a Fourier transform feature tensor through the widening convolution WildConv module and sent to the feature tensor merging module.
[0024] A recursive graph-based feature extraction network is used to capture the time-varying characteristics of the preprocessed seismic three-component waveform data through recursive graph transformation, resulting in a recursive graph. Figure 2 Dimensional tensor; recursion Figure 2 The dimensional tensor is used to extract a recursive graph feature tensor through the WideConv module and then sent to the feature tensor merging module.
[0025] The feature extraction network based on S-transform is used to simultaneously observe the energy distribution of the pre-processed three-component seismic waveform data in both the time and frequency domains through S-transform, resulting in a two-dimensional S-transform tensor. The two-dimensional S-transform tensor is then used to extract an S-transform feature tensor through the widening convolution WildConv module and sent to the feature tensor merging module.
[0026] The feature tensor merging module receives Fourier transform feature tensors from a feature extraction network based on short-time Fourier transform, recursive graph feature tensors from a feature extraction network based on recursive graph, and S-transform feature tensors from a feature extraction network based on S-transform; merges the three received feature tensors to obtain a merged feature tensor and sends the merged feature tensor to the spatial attention module.
[0027] The spatial attention module receives the merged feature tensor sent by the feature tensor merging module. By introducing spatial attention, the spatial attention module focuses more on the key spatial features of the residual features, obtains a spatial attention feature tensor, and sends the spatial attention feature tensor to the fully connected module.
[0028] The Bi-LSTM network is used to perform bidirectional modeling based on the preprocessed three-component seismic waveform data, capture time-related dependencies and contextual information in the input, extract a Bi-LSTM feature tensor, and send the Bi-LSTM feature tensor to the fully connected module.
[0029] Fully connected module: used to concatenate the spatial attention feature tensor sent by the spatial attention module and the Bi-LSTM feature tensor sent by the Bi-LSTM network into a final feature tensor; the final feature tensor is mapped through the fully connected layer to realize the mapping from any input to any output, thereby outputting the earthquake magnitude.
[0030] Compared with the prior art, the technical solution adopted in this invention has the following technical effects:
[0031] This invention constructs a multi-feature fusion-based earthquake magnitude detection model based on deep learning, which can calculate the magnitude in one go without multiple revisions or manual intervention. Considering the impact of instruments and noise on the magnitude calculation results, the three-component earthquake waveform data is preprocessed, and prior earthquake knowledge is added to the network, thereby improving the calculation accuracy. Attached Figure Description
[0032] Figure 1 This is a block diagram of an earthquake magnitude detection system based on deep learning and multi-feature fusion in an embodiment of the present invention.
[0033] Figure 2 This is a schematic diagram of the earthquake magnitude detection module in an embodiment of the present invention;
[0034] Figure 3 This is a flowchart of an earthquake magnitude detection method based on deep learning and multi-feature fusion in an embodiment of the present invention. Detailed Implementation
[0035] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.
[0036] The software environment for this implementation is the Windows 10 system.
[0037] Figure 1 The earthquake magnitude detection system based on deep learning multi-feature fusion shown includes:
[0038] The data preprocessing module is used to preprocess the raw three-component seismic waveform data and send the preprocessed three-component seismic waveform data to the earthquake magnitude detection module. The preprocessing method specifically includes: removing three-component seismic waveform data containing fragments or with a signal-to-noise ratio below a set threshold; correcting errors caused by instrument-related factors; and noise reduction and filtering.
[0039] The earthquake magnitude detection module is used to detect the corresponding magnitude based on preprocessed earthquake three-component waveform data and output the magnitude detection result.
[0040] like Figure 2 As shown, the earthquake magnitude detection module in this embodiment further includes:
[0041] A feature extraction network based on short-time Fourier transform is used to connect the time and frequency domains of the pre-processed three-component seismic waveform data through short-time Fourier transform, reflecting the two-dimensional features of the three-component seismic waveform data in the time and frequency domains, and obtaining a two-dimensional Fourier transform tensor. The two-dimensional Fourier transform tensor is then used to extract a Fourier transform feature tensor through the widening convolution WildConv module and sent to the feature tensor merging module.
[0042] A recursive graph-based feature extraction network is used to transform preprocessed seismic three-component waveform data through a recursive graph to capture the time-varying characteristics of the seismic three-component waveform data. This provides prior knowledge about the similarity, information content, and predictability of the seismic three-component waveform data, resulting in a recursive graph. Figure 2 Dimensional tensor; recursion Figure 2 The dimensional tensor is used to extract a recursive graph feature tensor through the WideConv module and then sent to the feature tensor merging module.
[0043] The feature extraction network based on S-transform is used to simultaneously observe the energy distribution of the pre-processed three-component seismic waveform data in both the time and frequency domains through S-transform, resulting in a two-dimensional S-transform tensor. The two-dimensional S-transform tensor is then used by the WidenConv module to extract an S-transform feature tensor and send it to the feature tensor merging module.
[0044] The feature tensor merging module receives Fourier transform feature tensors from a feature extraction network based on short-time Fourier transform, recursive graph feature tensors from a feature extraction network based on recursive graph, and S-transform feature tensors from a feature extraction network based on S-transform; merges the three received feature tensors to obtain a merged feature tensor and sends the merged feature tensor to the spatial attention module.
[0045] The spatial attention module receives the merged feature tensor sent by the feature tensor merging module. By introducing spatial attention, the spatial attention module focuses more on the key spatial features of the residual features, obtains a spatial attention feature tensor, and sends the spatial attention feature tensor to the fully connected module.
[0046] The Bi-LSTM network is used to perform bidirectional modeling based on the preprocessed three-component seismic waveform data, capture time-related dependencies and contextual information in the input, extract a Bi-LSTM feature tensor, and send the Bi-LSTM feature tensor to the fully connected module.
[0047] Fully connected module: used to concatenate the spatial attention feature tensor sent by the spatial attention module and the Bi-LSTM feature tensor sent by the Bi-LSTM network into a final feature tensor; the final feature tensor is mapped through the fully connected layer to realize the mapping from arbitrary input to arbitrary output, thereby outputting the earthquake magnitude.
[0048] This embodiment presents a deep learning-based multi-feature fusion method for earthquake magnitude detection, such as... Figure 3 As shown, the method includes the following steps:
[0049] Step 1: Acquire raw three-component seismic waveform data;
[0050] In this embodiment, the method for acquiring raw three-component earthquake waveform data is as follows: query the earthquake catalog information to determine the time when the P-wave of the three-component earthquake arrives at the earthquake monitoring station, i.e., the P-wave arrival time; after determining the P-wave arrival time, extract the data from 3 seconds before to 7 seconds after the P-wave arrival time of each component. The data sampling frequency is 100Hz, with a total of 1000 sampling points, to obtain the final raw three-component earthquake waveform data.
[0051] Step 2 involves preprocessing the original three-component seismic waveform data to obtain preprocessed three-component seismic waveform data.
[0052] Because the recording of seismic three-component waveform data by the station is affected by various external factors, causing interruptions in the seismic three-component waveform data, and because the waveform of the seismic wave may be affected by significant noise during propagation, resulting in unclear characteristics of the seismic waveform, data preprocessing is performed on the original seismic three-component waveform data obtained in step 1. In this embodiment, the method for data preprocessing of the original seismic three-component waveform data is as follows: removing seismic three-component waveform data containing interruptions and with a signal-to-noise ratio lower than 5dB; dividing the seismic three-component waveform data by the sensitivity of the seismic wave receiver to correct for errors caused by the instrument's own influencing factors; and then performing Butterworth filtering, mean denoting, and normalization operations on the seismic three-component waveform data. The preprocessed seismic three-component waveform data obtained is 1*1000*3 seismic three-component waveform data, where 1*1000 represents a one-dimensional vector with a length of 1000, and 3 represents 3 channels.
[0053] Step 3: Obtain the training set from the preprocessed three-component seismic waveform data;
[0054] In this embodiment, the method for obtaining the training set is as follows: all waveform data with a magnitude of 4 or above in the three-component earthquake waveform data are retained, and waveform data with a magnitude of 1-2 and 2-3 are randomly sampled to obtain three-component earthquake waveform data with a magnitude of 1-2, 2-3, and 4 or above in a ratio of 1:1:1. The dataset is then divided to obtain the training set.
[0055] Step 4: Build a magnitude detection model;
[0056] The specific steps include:
[0057] Step 4.1 Input the training set into the feature extraction network based on short-time Fourier transform, the feature extraction network based on recursive graph, and the feature extraction network based on S-transform respectively. After the three feature extraction networks extract the corresponding two-dimensional tensors based on prior knowledge, the widening convolution module WidenConv is used to further extract features at different scales and output the extracted Fourier transform feature tensor, recursive graph feature tensor, and S-transform feature tensor.
[0058] The specific method in this embodiment is as follows:
[0059] For a feature extraction network based on Short-Time Fourier Transform (SFT), the training set is input, and the SFT is used to obtain the two-dimensional Fourier transform tensor sp{x(t)} for each channel, as shown in the following formula:
[0060]
[0061] sp{x(t)}(τ,ω)=|X(τ,ω)| 2 (2)
[0062] Where w(t-τ) uses a Gaussian window function; ω represents the angular frequency; τ represents the size of the time window; e -jωt denoted by t; x(t) represents the signal value of each waveform component at time t; X(τ, ω) is defined as the short-time Fourier transform, which is a two-dimensional function of time t and frequency ω.
[0063] The Fourier transform 2D tensor sp{x(t)} is then passed to the WildConv module. This input Fourier transform 2D tensor is downsampled through a 3x3 convolutional layer and a ReLU activation layer, resulting in a 64*64*3 tensor. It then passes through six parallel convolutional layers with different kernel sizes, followed by a 3x3 convolutional layer, yielding a 64*64*64 Fourier transform feature tensor output1. Each parallel convolutional layer has the following characteristics:
[0064] c = ∑ k＝1，3，5 ω k x+∑ k＝1，3，5 b k (3)
[0065] Where c and x represent the output and input of the parallel convolutional layer, respectively; ω1, ω3, and ω5 represent the weights of the three parallel convolutional kernels, respectively; k is the size of the corresponding convolutional kernel; and b1, b3, and b5 represent the biases of the three parallel convolutional kernels, respectively.
[0066] For a feature extraction network based on a recursive graph, the training set is input to obtain the recursive graph. Figure 2 2D tensor RP i，j The formula is shown below:
[0067] RP i , j =||x i -x j ||, i, j=1, 2,..., td (4)
[0068] y = x(t) (5)
[0069] x i =(x(i),x(i+1),...,x(i+d)) (6)
[0070] Where y = x(t) is the data for each waveform component; d is the value of x. i The length of ||·|| is the first-order norm of the vector.
[0071] Then recursion Figure 2 2D tensor RP i，j Pass the input to the WidenConv module, and the input is recursive. Figure 2The dimensional tensor is downsampled to 64*64*3 after passing through a convolutional layer with a kernel size of 3x3 and an activation layer with the ReLU activation function. Then, it passes through six parallel convolutional layers with different kernel sizes, and is then input into a 3x3 convolutional layer, finally yielding a 64*64*64 recursive graph feature tensor output2.
[0072] For a feature extraction network based on the S-transform, the training set is input, and the S-transform is used to obtain the two-dimensional tensor S{x(t)} of each channel, as shown in the following formula:
[0073]
[0074] Where f is the reciprocal of the variance of the Gaussian window.
[0075] The S-transform 2D tensor S{x(t)} is then passed to the WildConv module. The input S-transform 2D tensor is downsampled through a convolutional layer with a kernel size of 3x3 and an activation layer with the ReLU activation function, resulting in a 64*64*3 tensor. It then passes through six parallel convolutional layers with different kernel sizes, and is then input into a 3x3 convolutional layer, finally yielding a 64*64*64 S-transform feature tensor output3.
[0076] Step 4.2 Input the training set into the Bi-LSTM network and output an extracted Bi-LSTM feature tensor;
[0077] In this embodiment, the training set is input into the Bi-LSTM network to obtain a 1*128 Bi-LSTM feature tensor output4.
[0078] Step 4.3 merge the three feature tensors generated in Step 4.1 into a single merged feature tensor and input it into the spatial attention module to obtain a spatial attention feature tensor;
[0079] In this specific implementation: the Fourier transform feature tensor output1, the recursive graph feature tensor output2, and the S-transform feature tensor output3 are merged into a 64*64*64 merged feature tensor, which is then input into the spatial attention module. In the spatial attention module, a 1x1 convolution is used to reduce the number of channels, and then a 3x3 convolution is used to expand the receptive field. High-resolution feature information is obtained directly using skip connections. A sigmoid function is used to obtain 64 weights between 0 and 1, in the format of 1*1*64, which serve as the weights for each channel. Each element of the corresponding channel is multiplied by the weight to obtain the output feature tensor. Then, it is flattened into a one-dimensional feature tensor through a Flatten layer, and the final output is a 1*128 spatial attention feature tensor output5.
[0080] Step 4.4: Concatenate the Bi-LSTM feature tensor obtained in Step 4.2 with the spatial attention feature tensor obtained in Step 4.3 to obtain the final feature tensor and input it into the fully connected layer;
[0081] In this implementation, the Bi-LSTM feature tensor output4 and the spatial attention feature tensor output5 are concatenated to obtain a final feature tensor of 1*256, which is then input into the fully connected layer FC, and the output is the magnitude.
[0082] Step 5: Train the model until the set number of training rounds is reached, and save the best-performing model from the trained rounds as the final magnitude detection model.
[0083] In this specific implementation, the model is trained until it reaches 60 iterations, and the model with the best performance in the trained iterations is saved.
[0084] The model with the best performance is determined by minimizing the loss function MSELoss. The formula is shown below:
[0085]
[0086] Among them, y i This refers to the data label value, which is the actual magnitude value; This represents the output value of the magnitude detection model; m is the batch size set during training.
[0087] Step 6: After preprocessing the three-component waveform data of the earthquake to be detected according to the data preprocessing method described in Step 2, input it into the final magnitude detection model to obtain the magnitude detection result;
[0088] In this embodiment, the three-component waveform data of earthquakes recorded by seismic stations in China from August 2010 to April 2015, provided by the China Earthquake Networks Center, are used. Some of the magnitude detection results are shown in Table 1.
[0089] Table 1
[0090]
[0091] The mean square error was calculated based on all detected results, and the final mean square error was 0.18, verifying that the method of the present invention can effectively detect the magnitude of an earthquake within an acceptable error range.
Claims
1. A multi-feature fusion earthquake magnitude detection method based on deep learning, characterized in that, The method includes the following steps: Step 1: Acquire raw three-component seismic waveform data; Step 2: Perform data preprocessing on the original three-component seismic waveform data to obtain preprocessed three-component seismic waveform data; Step 3: Obtain the training set from the preprocessed three-component seismic waveform data; Step 4: Build a magnitude detection model; Step 4 specifically includes the following steps: Step 4.1 Input the training set into the feature extraction network based on short-time Fourier transform, the feature extraction network based on recursive graph, and the feature extraction network based on S-transform respectively. After the three feature extraction networks extract the corresponding two-dimensional tensors based on prior knowledge, the widening convolution module WidenConv is used to further extract features at different scales and output the extracted Fourier transform feature tensor, recursive graph feature tensor, and S-transform feature tensor. Step 4.2 Input the training set into the Bi-LSTM network and output the extracted Bi-LSTM feature tensor; Step 4.3 Merge the three feature tensors generated in Step 4.1 into a single merged feature tensor and input it into the spatial attention module to obtain a spatial attention feature tensor; Step 4.4 Concatenate the Bi-LSTM feature tensor obtained in Step 4.2 with the spatial attention feature tensor obtained in Step 4.3 to obtain the final feature tensor and input it into the fully connected layer; Step 5: Train the model until the set number of training rounds is reached, and save the best-performing model from the trained rounds as the final magnitude detection model. Step 6: After preprocessing the three-component waveform data of the earthquake to be detected, input it into the final magnitude detection model to obtain the magnitude detection result.
2. The earthquake magnitude detection method based on deep learning and multi-feature fusion according to claim 1, characterized in that, The method for preprocessing the original three-component seismic waveform data is as follows: removing three-component seismic waveform data containing fragments or with a signal-to-noise ratio lower than a set threshold; dividing the three-component seismic waveform data by the sensitivity of the seismic wave receiver to correct errors caused by the instrument's own influencing factors; and then performing Butterworth filtering, mean denoting, and normalization operations on the three-component seismic waveform data.
3. A deep learning-based multi-feature fusion earthquake magnitude detection system, used to implement the deep learning-based multi-feature fusion earthquake magnitude detection method according to any one of claims 1-2, characterized in that, The system includes a data preprocessing module and an earthquake magnitude detection module.
4. The earthquake magnitude detection system based on deep learning and multi-feature fusion according to claim 3, characterized in that, The data preprocessing module is used to preprocess the raw three-component earthquake waveform data and send the preprocessed three-component earthquake waveform data to the earthquake magnitude detection module.
5. The earthquake magnitude detection system based on deep learning and multi-feature fusion according to claim 3, characterized in that, The earthquake magnitude detection module is used to detect the corresponding magnitude based on the pre-processed three-component earthquake waveform data and output the magnitude detection result.
6. The earthquake magnitude detection system based on deep learning and multi-feature fusion according to claim 5, characterized in that, The earthquake magnitude detection module further includes: A feature extraction network based on short-time Fourier transform is used to connect the time and frequency domains of the pre-processed three-component seismic waveform data through short-time Fourier transform, reflecting the two-dimensional features of the three-component seismic waveform data in the time and frequency domains, and obtaining a two-dimensional Fourier transform tensor. The two-dimensional Fourier transform tensor is then used to extract a Fourier transform feature tensor through the widening convolution WildConv module and sent to the feature tensor merging module. A recursive graph-based feature extraction network is used to capture the time-varying characteristics of the pre-processed three-component seismic waveform data through recursive graph transformation, resulting in a two-dimensional recursive graph tensor. The two-dimensional recursive graph tensor is then used to extract a recursive graph feature tensor through the widening convolution WildConv module and sent to the feature tensor merging module. The feature extraction network based on S-transform is used to simultaneously observe the energy distribution of the pre-processed three-component seismic waveform data in both the time and frequency domains through S-transform, resulting in a two-dimensional S-transform tensor. The two-dimensional S-transform tensor is then used to extract an S-transform feature tensor through the widening convolution WildConv module and sent to the feature tensor merging module. The feature tensor merging module receives Fourier transform feature tensors sent by a feature extraction network based on short-time Fourier transform, recursive graph feature tensors sent by a feature extraction network based on recursive graph, and S-transform feature tensors sent by a feature extraction network based on S-transform; merges the three received feature tensors to obtain a merged feature tensor and sends the merged feature tensor to the spatial attention module. The spatial attention module receives the merged feature tensor sent by the feature tensor merging module. By introducing spatial attention, the spatial attention module focuses more on the key spatial features of the residual features, obtains a spatial attention feature tensor, and sends the spatial attention feature tensor to the fully connected module. The Bi-LSTM network is used to perform bidirectional modeling based on the preprocessed three-component seismic waveform data, capture time-related dependencies and contextual information in the input, extract a Bi-LSTM feature tensor, and send the Bi-LSTM feature tensor to the fully connected module. Fully connected module: used to concatenate the spatial attention feature tensor sent by the spatial attention module and the Bi-LSTM feature tensor sent by the Bi-LSTM network into a final feature tensor; the final feature tensor is mapped through the fully connected layer to realize the mapping from any input to any output, thereby outputting the earthquake magnitude.