Seismic phase sequence detection method based on interpretable attention neural network

Abstract

The present application relates to a kind of seismic phase sequence detection methods based on interpretable attention neural network, belong to seismic waveform data processing and analysis field.The present application is by setting time window and step, the seismic waveform original data of acquisition is slid and intercepts as several detection windows;Seismic waveform data is input into the encoder Encoder of model, and the feature sequence extraction of spatial geometry, time series and global and local attention information is sequentially carried out;The extracted feature sequence is input into the decoder Decoder of model, is up-sampled to it, and finally the detection identification and time estimation result of phase sequence are output by model prediction layer;Gradient class activation mapping diagram visualization technology is added to the key convolution layer in model decoder, and the decision-making behavior of the phase sequence recognition and time estimation result of model output provides interpretability.The present application guarantees the performance of model and improves the internal transparency of model, and provides the interpretability for model decision-making behavior.

Description

Technical Field

[0001] This invention belongs to the field of earthquake waveform data processing and analysis, specifically relating to an earthquake phase sequence detection method based on an interpretable attention neural network. Background Technology

[0002] With the continuous establishment and improvement of global earthquake monitoring networks and infrastructure, more and more low-cost and high-efficiency sensor technologies have been developed, and massive amounts of data are constantly emerging. The field of seismology is facing the "3Vs" challenge brought about by the big data era: volume, variety, and velocity.

[0003] Against this backdrop, Artificial Intelligence (AI) technology has been increasingly applied to the field of seismology. Scholars both domestically and internationally have explored this area for over a decade, bringing AI technologies, represented by machine learning and deep learning, into almost every subfield of seismology. AI technology has demonstrated superior capabilities compared to classical methods, particularly in seismic tasks such as earthquake event detection, phase picking, and earthquake event recognition. Phase picking is a crucial step in the automated processing of seismic data; accurately detecting phases and determining arrival times is essential for earthquake event detection and is currently a popular application area for AI technology in seismic waveform data processing. However, most current AI-based phase recognition work primarily focuses on independently picking P and S phases, which suffers from problems such as single-phase picking and high false positive and false negative rates.

[0004] Furthermore, due to the highly complex structure of AI models and their inherent lack of transparency, it is impossible to explain or prove the rationale behind their decisions to humans. Therefore, they are often treated as "black boxes," making it difficult to demonstrate their rationality in important decisions. Without proper explanation, identifying and preventing model errors becomes impossible. In 2022, the National Natural Science Foundation of China (NSFC) released the guidelines for the Major Research Program "Explainable and Generalizable Next-Generation Artificial Intelligence Methods." Addressing the scientific problems of poor interpretability, poor robustness to adversarial examples, high data and computing power requirements, and weak theoretical foundations of AI methods, particularly deep learning, the program aims to establish high-precision, explainable, and generalizable new AI methods that do not rely on large amounts of labeled data through a combination of rule-based and learning approaches. It also seeks to develop the databases and model training platforms needed for next-generation AI methods and improve the infrastructure driving next-generation AI methods. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] The technical problem this invention aims to solve is how to provide an earthquake phase sequence detection method based on an interpretable attention neural network, in order to address the issue that AI-based phase identification work suffers from highly complex AI model structures, inherently lacking transparency, and unable to explain and prove their decision-making basis to humans.

[0007] (II) Technical Solution

[0008] To address the aforementioned technical problems, this invention proposes a method for detecting seismic phase sequences based on an interpretable attention neural network, which includes the following steps:

[0009] Step 1: By setting the time window and step size, the acquired raw seismic waveform data is slidably truncated into several windows to be detected, and the truncated data is preprocessed.

[0010] Step 2: Input the preprocessed seismic waveform data into the model's encoder, and extract the feature sequences of spatial geometry, time series, and global and local attention information in sequence;

[0011] Step 3: Input the feature sequence extracted in Step 2 into the decoder of the model, upsample it, and finally output the detection and identification of the seismic phase sequence and the time estimation results through the model prediction layer;

[0012] Step 4: Add gradient class activation maps to the key convolutional layers in the model decoder and visualize them to provide interpretability for the decision-making behavior of the model output seismic phase sequence identification and time estimation results.

[0013] (III) Beneficial Effects

[0014] This invention proposes an earthquake phase sequence detection method based on an interpretable attention neural network. This invention fully utilizes the advantage of artificial intelligence technology in the context of the emergence of massive earthquake waveform data, which can train a high-performance deep neural network model based on data-driven training, thereby efficiently and accurately detecting earthquake phases and determining the arrival time. At the same time, considering the inherent black-box nature of deep neural network models, interpretable optimization is carried out without affecting the original network structure and prediction accuracy of the model.

[0015] The beneficial effects of this invention include:

[0016] (1) In the encoder part of the model, the feature extraction process takes into account spatial geometric features, time series features, global attention features and local attention features at the same time, and has a strong feature discovery capability for one-dimensional time series data of earthquake waveform type.

[0017] (2) In the decoder part of the model, a one-dimensional deconvolutional neural network with the same number of layers as the encoder is used, which can restore the feature sequence extracted by the encoder part to the same dimension as the original input waveform data through upsampling, thereby realizing the picking of the phase arrival time.

[0018] (3) In the prediction layer of the model, a combination of fully connected layer and softmax function activation is adopted, which enables the model to distinguish multiple seismic phases at the same time, thereby realizing the identification of multiple seismic phases.

[0019] (4) Combining (1) to (3), compared with the current AI phase picking method which can only pick up two independent phases, P and S, the present invention can realize the overall detection and timely picking of phase sequences composed of multiple phases in seismic waveform data;

[0020] (5) Without affecting the original network structure and prediction accuracy of the model, an interpretable attention neural network for seismic waveform data phase sequence detection was designed. The waveform data regions that the key convolutional layers in the model focus on when contributing to the prediction results are visualized by a class activation map. This not only ensures the performance of the model but also improves the internal transparency of the model, providing interpretability for the model's decision-making behavior. Attached Figure Description

[0021] Figure 1 This is a flowchart of an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram of the phase sequence detection and identification and time estimation results obtained by processing the input raw seismic waveform data using a deep attention neural network model in an embodiment of the present invention.

[0023] Figure 3 The diagram shows the results of seismic phase sequence detection and identification and time estimation obtained by processing the original input seismic waveform data using a deep attention neural network model with an interpretable optimization mechanism added in this embodiment of the invention, as well as the visualization interpretation effect when the model detects two types of seismic phases, Pg and Sg. Detailed Implementation

[0024] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0025] To address the above problems, this invention proposes a method for seismic waveform data phase sequence detection based on an interpretable attention neural network. The specific steps are as follows:

[0026] Step 1: By setting the time window and step size, the acquired raw seismic waveform data is slidably truncated into several windows to be detected, and the truncated data is preprocessed.

[0027] Step 2: Input the preprocessed seismic waveform data into the encoder of the model, and extract the feature sequences of spatial geometry, time series, and global and local attention information in sequence.

[0028] Step 3: Input the feature sequence extracted in Step 2 into the decoder of the model, upsample it, and finally output the detection and identification of the seismic phase sequence and the time estimation results through the model prediction layer;

[0029] Step four involves adding gradient class activation maps to the key convolutional layers in the model decoder and visualizing them to provide interpretability for the decision-making behavior of the model's output seismic phase sequence identification and time estimation results.

[0030] Furthermore, the time window set in step one is 8.192 seconds, the step size is 0.1 seconds, and the preprocessing includes delinearization, filtering, and normalization operations on the seismic waveform data after sliding capture.

[0031] Furthermore, the model encoder described in step two consists of a multi-layer convolutional neural network, a recurrent neural network, and an attention neural network. The encoder sequentially extracts feature sequences from the seismic waveform data, including spatial geometry, temporal series data, and global and local attention information. The specific implementation is as follows:

[0032] S2.1 First, the preprocessed seismic waveform data is input into a one-dimensional convolutional neural network (CNN) consisting of 7 layers of Conv1D and 5 layers of ResCNN to extract spatial geometric features from the original waveform data by downsampling.

[0033] S2.2 Then, the downsampled geometric feature sequence extracted in step S2.1 is input into a recurrent neural network (RNN) consisting of two layers of Bi-LSTM and one layer of LSTM to further extract time series features.

[0034] S2.3 Next, the time series features extracted in step S2.2 are input into the attention neural network consisting of two layers of global Transformer, one layer of LSTM and one layer of local Transformer, respectively, to extract global (event level) attention and local (seismic phase level) attention feature information in sequence;

[0035] Furthermore, the model decoder described in step three consists of a 7-layer Conv1D one-dimensional deconvolutional neural network, and the model prediction layer consists of a 1-layer fully connected layer and a softmax activation function layer.

[0036] Furthermore, in step four, gradient class activation maps are added to and visualized for key convolutional layers in the model decoder. Without altering the original network structure and model parameters, the backpropagation gradients of the key convolutional layers are extracted to calculate feature map weights. The gradient class activation maps are obtained through weighted calculation and visualized, providing interpretability for the model's decision-making behavior in seismic phase sequence identification and time-of-arrival estimation. The specific implementation is as follows:

[0037] S4.1 First, select the convolutional layer of interest in the model decoder as the key convolutional layer;

[0038] S4.2, Based on a certain fully connected layer category y c The output result of the channel corresponding to (c represents the category number) is the output feature map A of the above key convolutional layer. k (k represents the channel number) Calculate the inverse gradient matrix The dimensions (width and height) of this gradient matrix are related to the feature map A. k Consistent;

[0039] S4.3, By analyzing the gradient matrix Perform global average pooling to obtain feature map A. k weight value for:

[0040]

[0041] Where Z = u × v, u and v represent feature map A respectively. k Width and height.

[0042] S4.4. Calculate the weight values ​​of all channel feature maps using S4.3, and then use these calculated weight values ​​to weight each channel feature map to obtain the class activation map L for category c. c for:

[0043]

[0044] S4.5 The class activation map calculated in S4.4 is overlaid on the original seismic waveform data in the form of a heat map for visualization. It can be found that the part with higher heat values ​​in the figure is the waveform region that the model focuses on when the output prediction result is class c. Visualization improves the local transparency of the black box model and provides decision-making basis and necessary explanation for the model's prediction behavior.

[0045] S4.6 By repeating steps S4.1 to S4.5, the corresponding class activation map can be calculated for any fully connected layer of interest according to any fully connected layer category and visualized, thereby improving the interpretability of model decision behavior while enhancing model transparency.

[0046] Example 1:

[0047] like Figure 1 The embodiment includes the following steps:

[0048] Step 1: By setting the time window and step size, the acquired raw seismic waveform data is slidably truncated into several windows to be detected, and the truncated data is preprocessed.

[0049] (1.1) Set the time window to 8.192 seconds, and use a step size of 0.1 seconds to extract several windows of the original seismic waveform data through a sliding window;

[0050] (1.2) Perform preprocessing operations such as delinearization and filtering on the original seismic waveform data, and then normalize the data by using the maximum value of each segment.

[0051] Step 2: Input the preprocessed seismic waveform data into the model's encoder, and extract the feature sequences of spatial geometry, time series, and global and local attention information in sequence.

[0052] (2.1) The preprocessed seismic waveform data is input into a one-dimensional convolutional neural network (CNN) consisting of a pre-trained 7-layer Conv1D and 5-layer ResCNN to extract spatial geometric features from the original waveform data by downsampling.

[0053] (2.2) The spatial geometric feature sequence extracted in (2.1) is further input into a recurrent neural network (RNN) composed of a pre-trained 2-layer Bi-LSTM and a 1-layer LSTM to further extract time series features;

[0054] (2.3a) The temporal feature sequence extracted in (2.2) is further input into a global attention network composed of two pre-trained Transformer layers to extract event-level global attention feature sequences.

[0055] (2.3b) The global attention feature sequence extracted in (2.3a) is further input into the local attention neural network composed of a pre-trained 1-layer LSTM and a 1-layer local Transformer to extract the local attention feature sequence oriented towards the seismic phase level.

[0056] Step 3: Input the feature sequence extracted in Step 2 into the model's decoder, upsample it, and finally output the detection and identification of the seismic phase sequence and the estimated time through the model's prediction layer.

[0057] S31. Input the feature sequence extracted by the encoder in step two into the decoder composed of a pre-trained 7-layer Conv1D one-dimensional deconvolutional neural network. Upsample it to restore it to the same dimension as the original input waveform data, thereby realizing the picking of the phase arrival time.

[0058] S32. Input the deconvolution feature sequence output by S31 into the model prediction layer, which consists of a pre-trained fully connected layer and a Softmax activation function layer, thereby realizing the identification of different seismic phase categories. In this embodiment of the invention, four seismic phases, Pn, Pg, Sn, and Sg, are used as examples.

[0059] S33. Through S31 and S32, the model finally outputs the detection and identification of the phase sequence (Pn, Pg, Sn, Sg) of the original seismic waveform data and the estimated arrival time. The effect is as follows: Figure 2 As shown.

[0060] Step four involves adding gradient class activation map visualization technology to the key convolutional layers in the model decoder to provide interpretability for the decision-making behavior of the model output seismic phase sequence identification and time estimation results.

[0061] S41. In this embodiment of the invention, the last convolutional layer of the model decoder is selected as the key convolutional layer;

[0062] S42. Based on the channel output results corresponding to the Pg phase category in the fully connected layer, respectively, analyze the output feature map A of the key convolutional layer. k (k represents the channel number) Calculate the inverse gradient matrix The dimensions (width and height) of this gradient matrix are related to the feature map A. k Consistent;

[0063] S43, By adjusting the gradient matrix Perform global average pooling to obtain feature map A. k weight value for:

[0064]

[0065] Where Z = u × v, u and v represent feature map A respectively. k Width and height.

[0066] S44. Repeat step S43 to calculate the weight values ​​of all channel feature maps of the key convolutional layer, and use the calculated weight values ​​to weight each channel feature map to obtain the class activation map L of the Pg seismic phase category. Pg for:

[0067]

[0068] S45. Repeat steps S41 to S44 to obtain the class activation mapping map L of the Sg seismic phase category. Sg ;

[0069] S46. The calculated Pg and Sg seismic phase class activation maps are overlaid on the original seismic waveform data as heatmaps for visualization. This reveals that the regions with higher heatmap values ​​are the waveform areas the model focuses on when predicting Pg or Sg seismic phase classes. This visualization enhances the local transparency of the black-box model, providing decision-making basis and necessary explanations for the model's predictive behavior. The effect is as follows: Figure 3 As shown.

[0070] Figure 2 This diagram illustrates the phase sequence detection and arrival estimation results obtained by processing the input raw seismic waveform data using a deep attention neural network model in this embodiment of the invention. The first row represents the raw waveform data, the second row represents the phase category and the true arrival value, and the third row represents the predicted phase category and arrival information output by the model. The confidence level and error range of the prediction results are represented in Gaussian distribution form.

[0071] Figure 3 This invention presents a deep attention neural network model with an added interpretable optimization mechanism. The model processes the input raw seismic waveform data to obtain seismic phase sequence detection and arrival time estimation results. A visual interpretation diagram is also provided for the model's detection of Pg and Sg seismic phases. The first row represents the raw waveform data; the second row represents the true values ​​of the seismic phase category and arrival time; the third row represents the model's predicted seismic phase category and arrival time information, expressed as a Gaussian distribution with confidence and error range; the fourth and fifth rows represent the significance responses of the model's key feature layers to the arrival times of Pg and Sg seismic phases, respectively, with red indicating strong significance and blue indicating weak significance.

[0072] Example 2:

[0073] A method for seismic waveform data phase sequence detection based on an interpretable attention neural network includes the following specific steps:

[0074] Step 1: By setting the time window and step size, the acquired raw seismic waveform data is slidably truncated into several windows to be detected, and the truncated data is preprocessed.

[0075] Step 2: Input the preprocessed seismic waveform data into the model's encoder, and extract the feature sequences of spatial geometry, time series, and global and local attention information in sequence.

[0076] Step 3: Input the feature sequence extracted in Step 2 into the model's decoder, upsample it, and finally output the detection and identification of the seismic phase sequence and the time estimation results through the model prediction layer;

[0077] Step four involves adding gradient class activation map visualization technology to the key convolutional layers in the model decoder to provide interpretability for the decision-making behavior of the model output seismic phase sequence identification and time estimation results.

[0078] Furthermore, the time window set in step one is 8.192 seconds, the step size is 0.1 seconds, and the preprocessing includes delinearization, filtering, and normalization operations on the seismic waveform data after sliding capture.

[0079] Furthermore, the model encoder described in step two consists of a multi-layer convolutional neural network, a recurrent neural network, and an attention neural network. The encoder sequentially extracts feature sequences from the seismic waveform data, including spatial geometry, temporal series data, and global and local attention information. The specific implementation is as follows:

[0080] (1.1) First, the seismic waveform data obtained by sliding is input into a one-dimensional convolutional neural network (CNN) consisting of 7 layers of Conv1D and 5 layers of ResCNN to extract spatial geometric features by downsampling the original waveform data;

[0081] (1.2) Then the downsampled geometric feature sequence extracted in step (1.1) is input into a recurrent neural network (RNN) consisting of two layers of Bi-LSTM and one layer of LSTM to further extract time series features;

[0082] (1.3) Next, the time series features extracted in step (1.2) are input into the attention neural network consisting of two layers of global Transformer, one layer of LSTM and one layer of local Transformer, respectively, to extract global (event level) attention and local (seismic phase level) attention feature information in turn;

[0083] Furthermore, the model decoder described in step three consists of a 7-layer Conv1D one-dimensional deconvolutional neural network, and the model prediction layer consists of a 1-layer fully connected layer and a softmax activation function layer.

[0084] Furthermore, the visualization technique of adding gradient class activation maps to key convolutional layers in the model decoder, as described in step four, can extract the backpropagation gradient calculation feature map weights of key convolutional layers without changing the original network structure and model parameters. The gradient class activation map is obtained through weighted calculation and then visualized, providing interpretability for the model's decision-making behavior in seismic phase sequence identification and time estimation. The specific implementation method is as follows:

[0085] (1.1) First, select the convolutional layer of interest in the model decoder as the key convolutional layer;

[0086] (1.2) Based on a certain fully connected layer category y c The output result of the channel corresponding to (c represents the category number) is the output feature map A of the above key convolutional layer. k (k represents the channel number) Calculate the inverse gradient matrix The dimensions (width and height) of this gradient matrix are related to the feature map A. k Consistent;

[0087] (1.3) By analyzing the gradient matrix Perform global average pooling to obtain feature map A. k weight value for:

[0088]

[0089] Where Z = u × v, u and v represent feature map A respectively. k Width and height.

[0090] (1.4) The weight values ​​of all channel feature maps of the key convolutional layer are calculated using (1.3), and the calculated weight values ​​are used to weight each channel feature map to obtain the class activation map L of category c. c for:

[0091]

[0092] (1.5) The class activation map calculated in (1.4) is superimposed on the original seismic waveform data in the form of a heat map for visualization. It can be found that the part with higher heat values ​​in the figure is the waveform area that the model focuses on when the output prediction result is class c. Visualization improves the local transparency of the black box model and provides decision-making basis and necessary explanation for the model's prediction behavior.

[0093] (1.6) By repeating steps (1.1) to (1.5), the corresponding class activation map of any fully connected layer can be calculated and visualized for any convolutional layer of interest, thereby improving the interpretability of model decision behavior while enhancing model transparency.

[0094] This invention fully leverages the advantages of artificial intelligence technology in the era of massive seismic waveform data emergence. It enables data-driven training of high-performance deep neural network models, allowing for efficient and accurate detection of seismic phases and arrival times. Simultaneously, considering the inherent black-box nature of deep neural network models, interpretable optimization is performed without affecting the original network structure and prediction accuracy. Its key features are:

[0095] (1) In the encoder part of the model, the feature extraction process takes into account spatial geometric features, time series features, global attention features and local attention features at the same time, and has a strong feature discovery capability for one-dimensional time series data of earthquake waveform type.

[0096] (2) In the decoder part of the model, a one-dimensional deconvolutional neural network with the same number of layers as the encoder is used, which can restore the feature sequence extracted by the encoder part to the same dimension as the original input waveform data through upsampling, thereby realizing the picking of the phase arrival time.

[0097] (3) In the prediction layer of the model, a combination of fully connected layer and softmax function activation is adopted, which enables the model to distinguish multiple seismic phases at the same time, thereby realizing the identification of multiple seismic phases.

[0098] (4) Combining (1) to (3), compared with the current AI phase picking method which can only pick up two independent phases, P and S, the present invention can realize the overall detection and timely picking of phase sequences composed of multiple phases in seismic waveform data;

[0099] (5) Without affecting the original network structure and prediction accuracy of the model, an interpretable attention neural network for seismic waveform data phase sequence detection was designed. The waveform data regions that the key convolutional layers in the model focus on when contributing to the prediction results are visualized by a class activation map. This not only ensures the performance of the model but also improves the internal transparency of the model, providing interpretability for the model's decision-making behavior.

[0100] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting seismic phase sequences based on interpretable attention neural networks, characterized in that, The method includes the following steps: Step 1: By setting the time window and step size, the acquired raw seismic waveform data is slidably truncated into several windows to be detected, and the truncated data is preprocessed. Step 2: Input the preprocessed seismic waveform data into the model's encoder, and extract the feature sequences of spatial geometry, time series, and global and local attention information in sequence; Step 3: Input the feature sequence extracted in Step 2 into the decoder of the model, upsample it, and finally output the detection and identification of the seismic phase sequence and the time estimation results through the model prediction layer; Step 4: Add gradient class activation maps to the key convolutional layers in the model decoder and visualize them to provide interpretability for the decision-making behavior of the model output seismic phase sequence identification and time estimation results; in, Step four specifically includes: S4.1 First, select the convolutional layer of interest in the model decoder as the key convolutional layer; S4.2, Based on a certain fully connected layer category The corresponding channel output result, where c represents the category number, and the output feature map of the above key convolutional layer. Calculate the inverse gradient matrix k represents the channel number, and the dimension of the gradient matrix is ​​the same as that of the feature map. Consistent, with dimensions of width and height; S4.3, By analyzing the gradient matrix Perform global average pooling to obtain the feature map. weight value for: in, , and They represent feature maps respectively. Width and height; S4.4 Calculate the weight values ​​of all channel feature maps using S4.3, and then use the calculated weight values ​​to weight each channel feature map to obtain the class activation map of category c. for: S4.5 The class activation map calculated in S4.4 is overlaid on the original seismic waveform data in the form of a heat map for visualization. The part with higher heat values ​​in the figure is the waveform region that the model focuses on when the output prediction result is class c. Visualization improves the local transparency of the black box model and provides decision-making basis and necessary explanation for the model's prediction behavior. S4.6 By repeating steps S4.1 to S4.5, calculate the corresponding class activation map for any convolutional layer of interest according to any fully connected layer category and visualize it, thereby improving the interpretability of the model's decision-making behavior while enhancing the transparency of the model.

2. The earthquake phase sequence detection method based on interpretable attention neural network as described in claim 1, characterized in that, The time window set in step one is 8.192 seconds, the step size is 0.1 seconds, and the preprocessing includes delinearization, filtering, and normalization operations on the seismic waveform data after sliding capture.

3. The earthquake phase sequence detection method based on interpretable attention neural network as described in claim 2, characterized in that, The model encoder described in step two consists of a multi-layer convolutional neural network, a recurrent neural network, and an attention neural network. The encoder extracts feature sequences of spatial geometry, time series, and global and local attention information from the seismic waveform data in sequence.

4. The earthquake phase sequence detection method based on interpretable attention neural network as described in claim 3, characterized in that, Step two specifically includes the following steps: S2.1 First, the preprocessed seismic waveform data is input into a one-dimensional convolutional neural network (CNN) consisting of 7 layers of Conv1D and 5 layers of ResCNN to extract spatial geometric features from the original waveform data by downsampling. S2.2 Then, the downsampled geometric feature sequence extracted in step S2.1 is input into a recurrent neural network (RNN) consisting of two layers of Bi-LSTM and one layer of LSTM to further extract time series features. S2.3 Next, the time series features extracted in step S2.2 are input into the attention neural network consisting of two layers of global Transformer, one layer of LSTM and one layer of local Transformer, respectively, to extract global attention and local attention feature information in sequence; the global attention feature information is event-level feature information, and the local attention feature information is phase-level feature information.

5. The earthquake phase sequence detection method based on interpretable attention neural network as described in claim 1, characterized in that, The model decoder described in step three consists of a 7-layer Conv1D one-dimensional deconvolutional neural network, and the model prediction layer consists of a 1-layer fully connected layer and a softmax activation function layer.

6. The earthquake phase sequence detection method based on interpretable attention neural network as described in any one of claims 1-5, characterized in that, In step four, gradient class activation maps are added to and visualized in the key convolutional layers of the model decoder. Without changing the original network structure and model parameters, the backpropagation gradients of the key convolutional layers are extracted to calculate the feature map weights. The gradient class activation maps are obtained through weighted calculation and visualized, providing interpretability for the decision-making behavior of the model output seismic phase sequence identification and time estimation results.

7. The earthquake phase sequence detection method based on interpretable attention neural network as described in any one of claims 1-5, characterized in that, Step three specifically includes: S31. Input the feature sequence extracted by the encoder in step two into the decoder composed of a pre-trained 7-layer Conv1D one-dimensional deconvolutional neural network. By upsampling, restore it to the same dimension as the original input waveform data, thereby realizing the picking of the phase arrival time. S32. Input the deconvolution feature sequence output from S31 into the model prediction layer, which consists of a pre-trained 1-layer fully connected layer and a softmax activation function layer, thereby realizing the identification of different seismic phase categories. S33. Through S31 and S32, the model finally outputs the phase sequence detection and identification and the estimated time of the original seismic waveform data.

8. The earthquake phase sequence detection method based on interpretable attention neural network as described in claim 7, characterized in that, In S32, the seismic phase categories include four types: Pn, Pg, Sn, and Sg. In S33, the detection, identification, and time estimation results of the four types of seismic phases (Pn, Pg, Sn, and Sg) of the original seismic waveform data are output.

9. The earthquake phase sequence detection method based on interpretable attention neural network as described in claim 8, characterized in that, Step four specifically includes the following steps: S41. Select the last convolutional layer of the model decoder as the key convolutional layer; S42. Based on the channel output results corresponding to the Pg phase category in the fully connected layer, analyze the output feature map of the key convolutional layer. Calculate the inverse gradient matrix k represents the channel number, and the dimension of the gradient matrix is ​​the same as that of the feature map. Consistent, with dimensions of width and height; S43, By adjusting the gradient matrix Perform global average pooling to obtain the feature map. weight value for: in, , and They represent feature maps respectively. Width and height; S44. Repeat step S43 to calculate the weight values ​​of all channel feature maps of the key convolutional layer, and use the calculated weight values ​​to weight each channel feature map to obtain the class activation map of the Pg seismic phase category. for: S45. Repeat steps S41 to S44 to obtain the class activation map of the Sg seismic phase category. ; S46. The calculated Pg and Sg phase activation maps are respectively superimposed on the original seismic waveform data in the form of heat maps for visualization. The part with higher heat values ​​in the figure is the waveform area that the model focuses on when the output prediction result is Pg or Sg phase.

Images