A seismic time-series signal detection method and related device based on attention mechanism

Abstract

This application discloses a seismic time series signal detection method and related apparatus based on an attention mechanism. The method includes monitoring a seismic time series signal with three components; inputting the seismic time series signal into a pre-trained detection network model; and determining the seismic phase, P-wave first arrival time, and S-wave first arrival time of the seismic time series signal through the detection network model. The downsampling module in the detection network model of this application includes a sequence attention subunit for extracting sequence features. The sequence attention subunit can efficiently extract sequence signal features, reduce the computational complexity of the detection model, and improve the detection speed of seismic signals.

Description

Technical Field

[0001] This application relates to the field of earthquake detection technology, and in particular to an attention-based method and related apparatus for detecting earthquake time-series signals. Background Technology

[0002] Due to the widespread deployment of earthquake monitoring stations and seismic sensors, the speed and scale of earthquake records that can now be obtained have increased significantly. This makes it impossible for earthquake analysis experts to process all the acquired data manually. Therefore, automated programs are needed to complete the analysis of earthquake signals. The objectives of earthquake signal analysis include earthquake detection and P / S wave picking. Earthquake detection refers to distinguishing noise from earthquake signals in long-term continuous monitoring signals. P / S wave picking is to identify the first arrival time of the P wave and the first arrival time of the S wave.

[0003] Currently, existing seismic signal analysis methods typically involve artificially designing a function to describe a certain attribute of a seismic signal within a time window, and then setting a threshold. When the attribute value exceeds the threshold, a seismic signal is considered to exist within that time window. However, due to the complexity of seismic signals, factors such as different source types, topography, distance, path fading, environmental noise, and sensor sensitivity all affect the received seismic signals. Therefore, an artificially designed function for a specific condition is difficult to apply to other conditions. It often requires some prior knowledge of the received signal and then readjusting the hyperparameters of the function, which significantly increases the barrier to entry for application.

[0004] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a method and related apparatus for detecting seismic time-series signals based on an attention mechanism, addressing the shortcomings of existing technologies.

[0006] To address the aforementioned technical problems, the first aspect of this application provides a seismic time-series signal detection method based on an attention mechanism, the method comprising:

[0007] The monitoring system consists of three components: east-west, north-south, and vertical.

[0008] The earthquake time series signal is input into a pre-trained detection network model. The detection network model determines the earthquake phase, P-wave first arrival time, and S-wave first arrival time of the earthquake time series signal. The detection network model includes a downsampling module and three parallel upsampling modules. The downsampling module is connected to the three upsampling modules respectively. The downsampling module includes several downsampling units. Each downsampling unit includes a downsampling layer and several sequence attention subunits for extracting sequence features.

[0009] The earthquake time series signal detection method based on the attention mechanism, wherein the three upsampling modules correspond to the earthquake phase, the first arrival time of the P-wave, and the first arrival time of the S-wave, respectively; the determination of the earthquake phase, the first arrival time of the P-wave, and the first arrival time of the S-wave of the earthquake time series signal through the detection network model specifically includes:

[0010] The downsampling module acquires several features corresponding to the seismic time series signal, wherein each feature corresponds to a specific downsampling unit.

[0011] Several features are input into three upsampling modules, and the probability distributions of earthquake phase, P-wave first arrival, and S-wave first arrival are determined by the three upsampling modules respectively.

[0012] Based on the probability distribution of the earthquake phase, the probability distribution of the first arrival of the P-wave, and the probability distribution of the first arrival of the S-wave, the earthquake phase, the first arrival time of the P-wave, and the first arrival time of the S-wave of the earthquake time series signal are determined.

[0013] The seismic time series signal detection method based on the attention mechanism includes an upsampling module comprising several cascaded upsampling units and an activation layer. The upsampling units correspond one-to-one with the downsampling units. The input of the later upsampling unit in two adjacent upsampling units is the output of the earlier upsampling unit and the output of its corresponding downsampling unit. The input of the first upsampling unit is the output of its corresponding downsampling unit. The input of the activation layer is the output of the last upsampling unit.

[0014] The seismic time-series signal detection method based on the attention mechanism, wherein the sequence attention subunit includes several cascaded sequence attention blocks, each sequence attention block includes an attention mechanism layer, and the attention mechanism layer includes a convolutional layer, a dilated convolutional layer, and a cross-channel convolutional layer. The convolutional layer is connected to the dilated convolutional layer, and the dilated convolutional layer is connected to the cross-channel convolutional layer. The convolutional layer and the dilated convolutional layer operate on channels, and the cross-channel convolutional layer is used to operate between channels.

[0015] The seismic time-series signal detection method based on the attention mechanism, wherein the sequence attention block includes a first cross-channel convolutional layer, a first batch of normalization layers, a nonlinear activation function layer, a multiplier, a second cross-channel convolutional layer, a second batch of normalization layers, and an adder. The first cross-channel convolutional layer, the first batch of normalization layers, the nonlinear activation function layer, the attention mechanism layer, the multiplier, the second cross-channel convolutional layer, the second batch of normalization layers, and the adder are connected sequentially. The input terms of the multiplier are the output terms of the attention mechanism layer and the nonlinear activation function layer, and the input terms of the adder are the output terms of the second batch of normalization layers and the input terms of the first cross-channel convolutional layer.

[0016] The seismic time series signal detection method based on the attention mechanism includes a sequence attention subunit that further includes a normalization layer. The first sequence attention block among the plurality of sequence attention blocks is connected to the downsampling layer, and the last sequence attention block is connected to the normalization layer.

[0017] The earthquake time-series signal detection method based on the attention mechanism, wherein, before inputting the earthquake time-series signal into the pre-trained detection network model, the method further includes:

[0018] Each component of the earthquake time series signal is normalized, and the normalized earthquake time series signal is used as the earthquake time series signal.

[0019] A second aspect of this application provides an attention-based seismic time-series signal detection system, the system comprising:

[0020] The monitoring module is used to monitor seismic time-series signals with three components, which correspond to the east-west direction, the north-south direction, and the vertical direction, respectively.

[0021] The control module is used to input the earthquake time series signal into a pre-trained detection network model, and to determine the earthquake phase, P-wave first arrival time, and S-wave first arrival time of the earthquake time series signal through the detection network model. The detection network model includes a downsampling module and an upsampling module, and the downsampling module is connected to the upsampling module. The downsampling module includes several downsampling units, and each downsampling unit includes a downsampling layer and several sequence attention subunits for extracting sequence features.

[0022] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the attention-based seismic time-series signal detection method as described above.

[0023] A fourth aspect of this application provides a terminal device, which includes: a processor, a memory, and a communication bus; the memory stores a computer-readable program that can be executed by the processor;

[0024] The communication bus enables communication between the processor and the memory;

[0025] When the processor executes the computer-readable program, it implements the steps in the attention-based seismic time-series signal detection method described above.

[0026] Beneficial Effects: Compared with existing technologies, this application provides a seismic time series signal detection method and related apparatus based on an attention mechanism. The method includes monitoring a seismic time series signal with three components; inputting the seismic time series signal into a pre-trained detection network model; and determining the seismic phase, P-wave first arrival time, and S-wave first arrival time of the seismic time series signal through the detection network model. The downsampling module in the detection network model of this application includes a sequence attention subunit for extracting sequence features. The sequence attention subunit can efficiently extract sequence signal features, reduce the computational complexity of the detection model, and improve the detection speed of seismic signals. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A flowchart of the attention mechanism-based seismic time-series signal detection method provided in this application.

[0029] Figure 2 This is an example diagram of an earthquake time series signal.

[0030] Figure 3 for Figure 2 The diagram shows the detection results corresponding to the earthquake time series signal.

[0031] Figure 4 This is a schematic diagram of the structure and principle of the network model for testing.

[0032] Figure 5 This is a schematic diagram of the structure of a sequence attention block.

[0033] Figure 6 This is a structural diagram of the attention mechanism layer.

[0034] Figure 7This is a schematic diagram of one component of the seismic time-series signal for the test case.

[0035] Figure 8 for Figure 7 A schematic diagram of the detection results corresponding to one component of the earthquake time series signal.

[0036] Figure 9 This is a comparison chart of the number of parameters and computational complexity of several models.

[0037] Figure 10 The structural principle diagram of the attention mechanism-based seismic time-series signal detection system provided in this application.

[0038] Figure 11 A schematic diagram of the terminal device provided in this application. Detailed Implementation

[0039] This application provides a method and related apparatus for detecting seismic time-series signals based on an attention mechanism. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0040] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0041] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0042] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0043] The inventors discovered through research that due to the widespread deployment of earthquake monitoring stations and vibration sensors, the speed and scale of earthquake records that can now be obtained have increased significantly. This makes it impossible for earthquake analysis experts to process all the collected data manually. Therefore, automated programs are needed to complete the analysis of earthquake signals. The objectives of earthquake signal analysis include earthquake detection and P / S wave picking. Earthquake detection refers to distinguishing noise from earthquake signals in long-term continuous monitoring signals. P / S wave picking is to identify the first arrival time of the P wave and the first arrival time of the S wave.

[0044] Currently, existing seismic signal analysis methods typically involve artificially designing a function to describe a certain attribute of a seismic signal within a time window, and then setting a threshold. When the attribute value exceeds the threshold, a seismic signal is considered to exist within that time window. However, due to the complexity of seismic signals, factors such as different source types, topography, distance, path fading, environmental noise, and sensor sensitivity all affect the received seismic signals. Therefore, an artificially designed function for a specific condition is difficult to apply to other conditions. It often requires some prior knowledge of the received signal and then readjusting the hyperparameters of the function, which significantly increases the barrier to entry for application.

[0045] With the application of deep learning methods in various fields, researchers have begun to apply them to seismic signal detection and phase picking. Commonly used deep learning models include PhaseNet, PickNet, and EQT, whose performance surpasses existing traditional methods, but they also have some limitations. The first two deep learning models directly use convolutional neural networks to extract features from seismic sequences. Due to the local connectivity of convolutional layers, the model has a relatively local receptive field; furthermore, once trained, the weights of the convolutional neural network are fixed and cannot adapt to different inputs. The latter, after extracting features using convolutional layers, further uses Long Short-Term Memory (LSTM) units to learn temporal information, and then uses a self-attention mechanism to learn the global correlation of features. However, because LSTM processes information flow in an autoregressive manner, it cannot compute the sequence input in parallel, increasing the model's inference time; and the self-attention mechanism requires calculating the correlation between every two time points in the feature sequence, with computational complexity proportional to the square of the sequence length.

[0046] To address the aforementioned issues, this embodiment monitors a three-part seismic time-series signal. The seismic time-series signal is then input into a pre-trained detection network model, which determines the seismic phase, P-wave first arrival time, and S-wave first arrival time of the seismic time-series signal. The downsampling module in the detection network model of this application includes a sequence attention subunit for extracting sequence features. This sequence attention subunit efficiently extracts sequence signal features, reduces the computational complexity of the detection model, and improves the detection speed of seismic signals.

[0047] The application content will be further explained below with reference to the accompanying drawings and through the description of the embodiments.

[0048] This embodiment provides an attention-based method for seismic time-series signal detection. The method describes the seismic signal detection and phase extraction tasks as sequence-to-sequence learning, where the input time series is mapped to a probability-output time series, and predictions are made for each time point to obtain the seismic phase, the first arrival time of the P-wave, and the first arrival time of the S-wave. Figure 1 As shown, the method includes:

[0049] S10. Monitor seismic time-series signals with three components.

[0050] Specifically, the earthquake time series signal is obtained through monitoring by pre-set geological instruments, which are used to detect earthquake time and record earthquake signals. In other words, the earthquake time series signal is the earthquake signal recorded by the geological instrument within a preset time period. For example, the earthquake time series signal is the earthquake signal collected continuously for 60 seconds by a geological instrument with a sampling frequency of 100 Hz. That is, the earthquake time series signal has a time span of 60 seconds and a sampling frequency of 100 Hz.

[0051] The three-component seismic time-series signal consists of seismic waveform data in the east-west, north-south, and vertical directions. Based on this, the seismic time-series signal is a sequence of seismic times acquired using a three-component seismograph to obtain seismic waveform data with east-west, north-south, and vertical directions. For example, as... Figure 2 As shown, the monitored seismic time-series signals are divided into three components, and each component contains information about P-waves and S-waves.

[0052] S20. Input the earthquake time series signal into a pre-trained detection network model, and determine the earthquake phase, P-wave first arrival time and S-wave first arrival time of the earthquake time series signal through the detection network model.

[0053] Specifically, the detection network model is used to detect earthquake time series signals to obtain the earthquake phase, P-wave first arrival time, and S-wave first arrival time of the earthquake time series signals. The input of the detection network model is the earthquake time series signal, and the output is the probability distribution of the earthquake phase, the probability distribution of the P-wave first arrival, and the probability distribution of the S-wave first arrival corresponding to the earthquake time series signal. Then, the probability distribution of the earthquake phase, the probability distribution of the P-wave first arrival, and the probability distribution of the S-wave first arrival are analyzed point by point to determine the earthquake phase, the P-wave first arrival time, and the S-wave first arrival time.

[0054] In one implementation, the detection network model can be trained and validated on the STEAD dataset, with training, validation, and test sets comprising 60%, 20%, and 20% of the total dataset, respectively. The detection network model is implemented using the PyTorch deep learning framework, with a smooth-L1 loss function, ADAM optimizer, and 50 epochs of training. The initial learning rate is set to 0.001, and the learning rate is halved every 5 epochs. Furthermore, during both training and use of the detection network model, after acquiring the earthquake time series signal, the signal can be normalized by subtracting the mean and then dividing by the maximum value.

[0055] The detection network model SEA-net includes a downsampling module and three parallel upsampling modules. The downsampling module is connected to the three upsampling modules. The downsampling module is used to acquire several features of the seismic time series signal. The three upsampling modules correspond to the seismic phase, the first arrival time of the P-wave, and the first arrival time of the S-wave, respectively. That is, one upsampling module is used to determine the probability distribution of the seismic phase, one upsampling module is used to determine the probability distribution of the first arrival of the P-wave, and one upsampling module is used to determine the probability distribution of the first arrival of the S-wave.

[0056] Based on this, the determination of the seismic phase, P-wave first arrival time, and S-wave first arrival time of the seismic time series signal through the detection network model specifically includes:

[0057] S21. Obtain several features corresponding to the seismic time series signal through the downsampling module;

[0058] S22. Input several features into three upsampling modules respectively, and determine the probability distribution of earthquake phase, the probability distribution of P-wave first arrival and the probability distribution of S-wave first arrival through the three upsampling modules respectively.

[0059] S23. Based on the probability distribution of the earthquake phase, the probability distribution of the first arrival of the P-wave, and the probability distribution of the first arrival of the S-wave, determine the earthquake phase, the first arrival time of the P-wave, and the first arrival time of the S-wave of the earthquake time series signal.

[0060] Specifically, such as Figure 3 As shown, the probability distributions of the earthquake phase, the first arrival of the P-wave, and the first arrival of the S-wave are all time-series signals, and the time lengths of the probability distributions of the earthquake phase, the first arrival of the P-wave, and the first arrival of the S-wave are the same as the time length of the earthquake time-series signal. For example, if the time length of the earthquake time-series signal is 60 seconds, then the time lengths of the probability distributions of the earthquake phase, the first arrival of the P-wave, and the first arrival of the S-wave are all 60 seconds.

[0061] like Figure 4 As shown, the downsampling module includes several cascaded downsampling units. The output of each downsampling unit is the input of the next downsampling unit and also serves as a feature of the seismic time series signal. Therefore, the number of features is the same as the number of downsampling units, and there is a one-to-one correspondence between the features and the downsampling units, with each feature being the output of its corresponding downsampling unit.

[0062] like Figure 4 As shown, the upsampling module includes several upsampling units, which correspond one-to-one with several downsampling units. The input of the later upsampling unit in two adjacent upsampling units includes the output of the earlier upsampling unit and the output of its corresponding downsampling unit. The input of the first upsampling unit includes the output of its corresponding downsampling unit. The input of the activation layer includes the output of the last upsampling unit. This allows the downsampling and upsampling modules to form a U-net network structure, enabling the detection network model to adopt the U-net encoder-decoder framework. The encoder stage is used to extract the signal features of the seismic time series signal. The decoder stage upsamples the features extracted by the encoder stage to restore the features to the same length as the input, thereby obtaining the probability distribution of the seismic phase, the probability distribution of the P-wave first arrival, and the probability distribution of the S-wave first arrival. In this way, the features are restored by a single independent upsampling model in the decoder stage, which allows the seismic phase detection, P-wave first arrival detection, and S-wave first arrival detection to be processed in parallel, improving the detection speed of the seismic time series signal. Furthermore, by using the encoder-decoder framework, a shortcut connection can be set between the downsampling unit and its corresponding upsampling unit, which can avoid the problem of affecting the temporal resolution due to excessive downsampling in the time dimension.

[0063] like Figure 4As shown, the upsampling module may further include an activation layer. This activation layer restricts the input terms of the last upsampling unit to a range of 0 to 1, thereby obtaining the probability distributions of the seismic phase, the P-wave first arrival, and the S-wave first arrival. Thus, in a series of cascaded upsampling units, the last upsampling unit is connected to an activation layer. The input terms of the activation layer are the output terms of the last upsampling unit. By restricting the output terms of the last upsampling unit, the output of the upsampling model is limited to a range of 0 to 1. In a typical implementation, the activation function layer may be configured with a sigmoid function.

[0064] like Figure 4 As shown, the downsampling unit includes a downsampling layer and several cascaded sequence attention sub-units. Each sequence attention sub-unit is used to extract sequence features from the seismic time series signal. The downsampling layer is connected to the first sequence attention sub-unit, and the output of the downsampling unit is the output of the last sequence attention sub-unit. This embodiment extracts local features from the seismic time series signal through sequence attention sub-units and can learn long-range dependencies, thus exhibiting adaptability to the input. Furthermore, this module has fewer parameters and lower computational complexity, enabling efficient feature extraction from seismic sequences.

[0065] In one implementation, such as Figure 4 As shown, the sequence attention subunit includes several cascaded sequence attention blocks (SEAs) and a normalization layer. The first sequence attention block is connected to the downsampling layer, and the last sequence attention block is connected to the normalization layer. In this implementation, the sequence attention subunit performs attention learning on the seismic time series signal through several cascaded sequence attention blocks to extract effective feature information.

[0066] like Figure 5 As shown, the sequence attention block includes a first cross-channel convolutional layer, a first batch of normalization layers, a non-linear activation function layer, an attention mechanism layer, a multiplier, a second cross-channel convolutional layer, a second batch of normalization layers, and an adder, all cascaded sequentially. The multiplier's input consists of the outputs of the attention mechanism layer and the non-linear activation function layer, while the adder's input consists of the output of the second batch of normalization layers and the input of the first cross-channel convolutional layer. The sequence attention block uses the first cross-channel convolutional layer, the first batch of normalization layers, the non-linear activation function layer, the second cross-channel convolutional layer, and the second batch of normalization layers to demonstrate the expressive power of the extracted features, and uses an adder to prevent the gradient vanishing problem during training due to excessive network depth.

[0067] The attention mechanism layer is used to learn attention weights for each vector element in a feature by studying local features, long-range dependencies between features in different feature channels, and dependencies between features in different channels. These attention weights reflect the usefulness of the information carried by each vector element. In one implementation, such as... Figure 6 As shown, the attention mechanism layer includes convolutional layers, dilated convolutional layers, and cross-channel convolutional layers. The convolutional layers are connected to the dilated convolutional layers, and the dilated convolutional layers are connected to the cross-channel convolutional layers. Both the convolutional and dilated convolutional layers operate on channels. The cross-channel convolutional layer learns the dependency features of different channels by utilizing the local features of the convolutional and dilated convolutional layers and the long-range dependencies of features in each feature channel. Furthermore, attention weights reflect the importance of the corresponding parts input to the attention mechanism layer and have the same shape as the input, allowing the attention weights to be multiplied by the input features. Thus, the sequence attention block multiplies the input and output terms of the attention mechanism layer using a multiplier. In other words, the attention mechanism layer and the sequence attention subunit can be:

[0068]

[0069]

[0070] Where X represents the input term of the attention mechanism layer, A(X) represents the output term of the attention mechanism layer, W1 represents the network module formed by the first cross-channel convolutional layer, the first batch of normalization layers, and the non-linear activation function layer, and W2 represents the network module formed by the second cross-channel convolutional layer and the second batch of normalization layers. Indicates an adder. Indicates a multiplier.

[0071] In summary, this embodiment provides a seismic time series signal detection method based on an attention mechanism. The method includes monitoring a seismic time series signal with three components; inputting the seismic time series signal into a pre-trained detection network model; and determining the seismic phase, P-wave first arrival time, and S-wave first arrival time of the seismic time series signal through the detection network model. The detection network model in this application includes a sequence attention subunit in the downsampling module for extracting sequence features. The sequence attention subunit can efficiently extract sequence signal features, reduce the computational complexity of the detection model, and improve the detection speed of seismic signals.

[0072] To further illustrate the effectiveness of the attention-based seismic time-series signal detection method provided in this embodiment, a specific test example is given.

[0073] Specific test examples are as follows: Figure 7The earthquake time series signal shown serves as the input to the detection network model (only one component is plotted; the actual input consists of three components), along with earthquake information manually annotated by seismologists. The output of the detection network model is as follows: Figure 8 As shown, by determining the arrival time of P / S waves through the peak value of the probability distribution, it can be seen that the detection network model can correctly detect the earthquake phase and pick up the first arrival time of P / S waves. That is, the method provided in this embodiment can correctly detect the earthquake phase and pick up the first arrival time of P / S waves.

[0074] To quantitatively evaluate the model's performance, we compared several popular methods and computed several metrics on the same test set, including accuracy (Pr), recall (Re), F1 score (F1), and the mean of the phase-picking residuals (μ). |Δ| ) and standard deviation (σ) Δ The results are shown in Table 1.

[0075] Table 1 Performance comparison of different methods

[0076]

[0077] The first three methods in Table 1 are traditional methods, while the rest are deep learning methods. It can be seen that deep learning methods outperform traditional methods in all performance metrics. Furthermore, this embodiment compares the learnable parameters and computational complexity of deep learning methods, such as... Figure 9 As shown, the detection method provided in this embodiment has a lower computational complexity than other methods.

[0078] Based on the above-mentioned attention-based seismic time-series signal detection method, this embodiment provides an attention-based seismic time-series signal detection system, such as... Figure 10 As shown, the system includes:

[0079] The monitoring module 100 is used to monitor seismic time-series signals with three components, which correspond to the east-west direction, the north-south direction, and the vertical direction, respectively.

[0080] The control module 200 is used to input the earthquake time series signal into a pre-trained detection network model, and to determine the earthquake phase, P-wave first arrival time, and S-wave first arrival time of the earthquake time series signal through the detection network model. The detection network model includes a downsampling module and an upsampling module, and the downsampling module is connected to the upsampling module. The downsampling module includes several downsampling units, and each downsampling unit includes a downsampling layer and several sequence attention subunits for extracting sequence features.

[0081] Based on the above-described attention-based seismic time-series signal detection method, this embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the attention-based seismic time-series signal detection method described in the above embodiment.

[0082] Based on the above-mentioned attention-based seismic time-series signal detection method, this application also provides a terminal device, such as... Figure 11 As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.

[0083] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0084] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.

[0085] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as transient storage media.

[0086] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A seismic time series signal detection method based on an attention mechanism, characterized in that, The method includes: The monitoring system consists of three components: east-west, north-south, and vertical. The earthquake time series signal is input into a pre-trained detection network model. The detection network model determines the earthquake phase, P-wave first arrival time, and S-wave first arrival time of the earthquake time series signal. The detection network model includes a downsampling module and three parallel upsampling modules. The downsampling module is connected to the three upsampling modules respectively. The downsampling module includes several downsampling units. Each downsampling unit includes a downsampling layer and several sequence attention subunits for extracting sequence features. The sequence attention subunit comprises several cascaded sequence attention blocks. Each sequence attention block includes an attention mechanism layer, a first cross-channel convolutional layer, a first batch normalization layer, a non-linear activation function layer, a multiplier, a second cross-channel convolutional layer, a second batch normalization layer, and an adder. The first cross-channel convolutional layer, the first batch normalization layer, the non-linear activation function layer, the attention mechanism layer, the multiplier, the second cross-channel convolutional layer, the second batch normalization layer, and the adder are connected sequentially. The input to the multiplier is the output of the attention mechanism layer and the non-linear activation function layer. The output of the adder is the output of the second batch normalization layer and the input of the first cross-channel convolutional layer. The attention mechanism layer includes a convolutional layer, a dilated convolutional layer, and a cross-channel convolutional layer. The convolutional layer is connected to the dilated convolutional layer, and the dilated convolutional layer is connected to the cross-channel convolutional layer. The convolutional layer and the dilated convolutional layer operate on channels. The cross-channel convolutional layer is used to operate between channels by learning the local features of the features of the convolutional layer and the dilated convolutional layer, as well as the long-range dependencies of the features in each feature channel. The cross-channel convolutional layer is used to learn the dependency features of different channel features. 2.The method of claim 1, wherein, The three upsampling modules correspond to the earthquake phase, the first arrival time of the P-wave, and the first arrival time of the S-wave, respectively; the determination of the earthquake phase, the first arrival time of the P-wave, and the first arrival time of the S-wave of the earthquake time series signal through the detection network model specifically includes: The downsampling module acquires several features corresponding to the seismic time series signal, wherein each feature corresponds to a specific downsampling unit. Several features are input into three upsampling modules, and the probability distributions of earthquake phase, P-wave first arrival, and S-wave first arrival are determined by the three upsampling modules respectively. Based on the probability distribution of the earthquake phase, the probability distribution of the first arrival of the P-wave, and the probability distribution of the first arrival of the S-wave, the earthquake phase, the first arrival time of the P-wave, and the first arrival time of the S-wave of the earthquake time series signal are determined.

3. The method according to claim 1 or 2, wherein, The upsampling module includes several cascaded upsampling units and an activation layer. The several upsampling units correspond one-to-one with several downsampling units. The input of the upsampling unit located later in two adjacent upsampling units is the output of the upsampling unit located earlier and the output of its corresponding downsampling unit. The input of the upsampling unit located at the very beginning is the output of its corresponding downsampling unit. The input of the activation layer is the output of the upsampling unit located at the very end.

4. The method of claim 1, wherein, The sequence attention subunit further includes a normalization layer. The first sequence attention block in the plurality of cascaded sequence attention blocks is connected to the downsampling layer, and the last sequence attention block is connected to the normalization layer.

5. The method of claim 1, wherein, Before inputting the seismic time series signal into the pre-trained detection network model, the method further includes: Each component of the earthquake time series signal is normalized, and the normalized earthquake time series signal is used as the earthquake time series signal.

6. An attention mechanism based seismic time series signal detection system, characterized in that, The system includes: The monitoring module is used to monitor seismic time-series signals with three components, which correspond to the east-west direction, the north-south direction, and the vertical direction, respectively. The control module is used to input the earthquake time series signal into a pre-trained detection network model, and to determine the earthquake phase, P-wave first arrival time, and S-wave first arrival time of the earthquake time series signal through the detection network model. The detection network model includes a downsampling module and an upsampling module, the downsampling module being connected to the upsampling module. The downsampling module includes several downsampling units, each of which includes a downsampling layer and several sequence attention subunits for extracting sequence features. The sequence attention subunit comprises several cascaded sequence attention blocks. Each sequence attention block includes an attention mechanism layer, a first cross-channel convolutional layer, a first batch normalization layer, a non-linear activation function layer, a multiplier, a second cross-channel convolutional layer, a second batch normalization layer, and an adder. The first cross-channel convolutional layer, the first batch normalization layer, the non-linear activation function layer, the attention mechanism layer, the multiplier, the second cross-channel convolutional layer, the second batch normalization layer, and the adder are connected sequentially. The input to the multiplier is the output of the attention mechanism layer and the non-linear activation function. The output of the layer, the input of the adder is the output of the second batch normalization layer and the input of the first cross-channel convolutional layer, wherein the attention mechanism layer includes a convolutional layer, a dilated convolutional layer and a cross-channel convolutional layer, the convolutional layer is connected to the dilated convolutional layer, the dilated convolutional layer is connected to the cross-channel convolutional layer, wherein the convolutional layer and the dilated convolutional layer operate on channels, and learn the dependency features of different channel features through the local features of the convolutional layer and the dilated convolutional layer and the long-range dependencies of each feature in the feature channel, the cross-channel convolutional layer is used to operate between channels, and learn the dependency features of different channel features through the cross-channel convolutional layer.

7. A computer readable storage medium characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the attention-based seismic time-series signal detection method as described in any one of claims 1-5.

8. A terminal device, characterized in that, include: Processor, memory, and communication bus; the memory stores a computer-readable program that can be executed by the processor; The communication bus enables communication between the processor and the memory; When the processor executes the computer-readable program, it implements the steps in the attention-based seismic time-series signal detection method as described in any one of claims 1-5.

Images