Construction method and application method of seismic phase arrival time discrimination model

Abstract

The invention discloses a construction method and application method of a seismic phase arrival time discrimination model, and the method comprises the steps: obtaining seismic phase arrival time discrimination related data, and building a sample set; the seismic phase arrival time judgment related data comprises seismic event waveform data, corresponding seismic phase arrival time related information and a correct seismic phase arrival time type; a label set in the sample set is a correct seismic phase arrival time type; establishing a neural network; the neural network comprises a network input layer, a convolution group, a multistage feature fusion layer, a multi-modal fusion layer, a full connection layer and an output layer; the network input layer loads seismic event waveform data, the multi-modal fusion layer loads seismic phase arrival time associated information, and the output layer loads a correct seismic phase arrival time type; and performing model training on the neural network, and constructing a seismic phase arrival time discrimination model. According to the invention, the method can optimize the recognition accuracy of the seismic phase recognition process, improves the completeness and accuracy of waveform data feature extraction, and improves the constraint capability and prediction precision of a network.

Description

Technical Field

[0001] The present invention relates to the field of earthquake monitoring technology, and in particular to a method for constructing and applying a seismic phase arrival time discrimination model. Background Art

[0002] The arrival time of a seismic phase refers to the expected time for a seismic wave to propagate inside the earth and reach a seismic station. It plays an indispensable role in constructing the internal structure of the earth, obtaining the location of the earthquake source, and exploring geodynamics. The accuracy of the arrival time of the seismic phase determines the validity of the result. In specific applications, manual picking can guarantee a high accuracy rate, but it relies on experience and takes a lot of time. In order to improve the efficiency of arrival time picking, some automatic picking algorithms have been proposed, such as the long-short time window energy ratio method, Akaike information criterion, Kalman filtering, polarization analysis, fractal dimension, wavelet transform, etc. However, these methods focus on the first arrival picking, ignore the subsequent seismic phases, and the picking effect is also very sensitive to noise.

[0003] In recent years, machine learning has been introduced into seismology and widely used in earthquake detection and arrival picking. With the continuous increase in data volume, deep learning, as the latest development of machine learning, has shown better results and provided new tools for efficient and high-precision phase arrival picking, such as PhaseNet, PickNet, PpkNet, EQtransform, etc., and has been widely used in earthquake detection and positioning, tomography, focal mechanism solutions (such as focal mechanism solutions), etc. These networks free researchers from repetitive picking work and provide a basis for the rapid advancement of subsequent research.

[0004] However, these algorithm networks are still found to have some problems during use, such as the picking results still need manual proofreading, and cannot accurately identify the types of two seismic phases with similar distances. The usual processing method is to limit the arrival time residual to a very narrow range, or set a minimum magnitude, or remove the epicenter distance range where the arrival time of two seismic phases is too close. The quality of the picking results has been improved to a certain extent, but these strategies also ignore a large amount of correct arrival time data, thereby limiting the resolution and accuracy of subsequent work.

[0005] Therefore, a technical solution is needed that can make full use of waveform data with various characteristics to distinguish the arrival types of different seismic phases. Summary of the invention

[0006] To achieve the above purpose, the present application provides a method for constructing a seismic phase arrival time discrimination model, comprising the following steps:

[0007] Obtaining phase arrival time discrimination related data, establishing a sample set, and generating a training set, a validation set, and a test set; wherein the phase arrival time discrimination related data includes: earthquake event waveform data, corresponding phase arrival time associated information, and correct phase arrival time type; the label set in the sample set is the correct phase arrival time type;

[0008] Establish a neural network; the neural network includes: a network input layer, a convolution group, a multi-level feature fusion layer, a multi-modal fusion layer, a fully connected layer and an output layer; wherein the network input layer loads the earthquake event waveform data, the multi-modal fusion layer loads the seismic phase arrival time correlation information, and the output layer loads the correct seismic phase arrival time type;

[0009] The neural network is trained through the sample set to construct a seismic phase arrival time discrimination model.

[0010] Among them, before establishing the sample set, data preprocessing is performed, including data enhancement and data balancing;

[0011] Data enhancement is performed on earthquake event waveform data, including: data window shifting, data mirroring, and data filtering;

[0012] Data balancing is an operation to ensure that there will be no deviation caused by imbalanced training data during the phase identification process. Data balancing means: classifying the data evenly according to different phase types to ensure that the amount of data under different phase types is equal.

[0013] Furthermore, the earthquake event waveform strictly corresponds to the corresponding earthquake phase arrival time correlation information.

[0014] Among them, the waveform data is four channels, the first three channels are the Z, N, and E component data of the waveform of the passing earthquake event, and the fourth channel is represented by a sharp pulse as the time position of the earthquake phase.

[0015] Among them, the label of the correct phase arrival type is composed of a one-dimensional vector of size 4. The positions of the vector elements correspond to the four phase types Pg, Pn, Sg, and Sn from front to back. The value of the vector element represents the probability of each phase type. The position with a value of 1 represents the correct phase type corresponding to each phase arrival in the label data set.

[0016] The phase arrival time correlation information includes: location, epicentral distance, focal depth, and magnitude; after the multimodal fusion layer loads the phase arrival time correlation information, the phase arrival time correlation information is multimodally fused with the characteristics of the earthquake event waveform data, which is reflected as follows:

[0017]

[0018] Among them, z wave is the multi-level feature encoding vector of earthquake event waveform data, zdist 、z mag 、z depth are the encoding vectors of focal depth, epicenter distance and magnitude, respectively, w (i) is the i-th low-rank factor, symbol Represents the element-wise product of tensors.

[0019] Furthermore, the method for implementing data balancing is as follows:

[0020] Determine a first epicentral distance range corresponding to Pg and Sg and a second epicentral distance range corresponding to Pg, Pn, Sg, Sn;

[0021] The earthquake phase arrival time discrimination related data are divided into two parts according to the first epicentral distance range and the second epicentral distance range; the boundary epicentral distance between the first epicentral distance range and the second epicentral distance range is obtained according to calculation, the minimum boundary of the first epicentral distance range is the minimum epicentral distance in the sample set, and the maximum boundary of the second epicentral distance range is the maximum epicentral distance in the sample set;

[0022] The seismic phase types that can be generated in each epicentral distance range are extracted from the two parts; when extracting, the amount of data of different seismic phase types in each epicentral distance range is equal.

[0023] The calculation method of the boundary epicentral distance of two epicentral distance ranges is:

[0024] Among them, Δ is the boundary epicenter distance, H is the average Moho depth of the study area, h is the average station elevation, d is the average focal depth, are the average velocities of P and S waves in the crust of the study area, are the average velocities of P and S waves at the top of the upper mantle in the study area, respectively; aver refers to the average value.

[0025] On the other hand, the present invention provides an application method of a seismic phase arrival time discrimination model, wherein the seismic phase arrival time discrimination model is constructed by the construction method provided by the present invention, and the application method comprises the following steps:

[0026] Inputting earthquake event waveform data into the input layer of the earthquake phase arrival time discrimination model;

[0027] Inputting the seismic phase arrival time correlation information into the multimodal fusion layer of the seismic phase arrival time discrimination model;

[0028] After the operation of the neural network, the result obtained in the output layer is the probability of each seismic phase type, and the correct seismic phase arrival type is determined based on the probability.

[0029] Among them, the probability judgment of the correct seismic phase arrival type includes: the seismic phase type with the highest probability is the correct seismic phase arrival type corresponding to the arrival position; if the probability of all seismic phases is less than 50%, it is determined that the arrival position does not belong to any seismic phase.

[0030] According to the present invention, the seismic phase identification process can be optimized, so that the identification result has a higher accuracy, the comprehensiveness and accuracy of waveform data feature extraction are improved, and the constraint ability and prediction accuracy of the network are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a step diagram of a method for constructing a seismic phase arrival time discrimination model based on deep learning according to an embodiment of the present invention;

[0032] Figure 2 is a schematic diagram of a neural network architecture provided according to an embodiment of the present invention;

[0033] Figure 3 2. It is a schematic diagram of ASFF adaptive spatial feature fusion according to an embodiment of the present invention;

[0034] Figure 4 is a schematic diagram of low-rank multimodal fusion provided according to an embodiment of the present invention;

[0035] Figure 5 It is a schematic diagram of the position of the seismic phase arrival time and the waveform data of the earthquake event provided by an embodiment of the present invention;

[0036] Figure 6 is a neural network training effect curve diagram provided according to an embodiment of the present invention;

[0037] Figure 7 This is an actual data test diagram of the seismic phase arrival time discrimination model provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention provides a method for constructing a seismic phase arrival time discrimination model based on deep learning. Through convolutional neural networks, multi-level feature fusion and multimodal fusion modules, it can make full use of a variety of earthquake-related data to determine the type of seismic phase arrival time.

[0039] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings.

[0040] Figure 1 The steps of building a seismic phase arrival time discrimination model based on deep learning are provided, as shown in the figure, including:

[0041] Step S100: Acquire seismic phase arrival time discrimination related data, establish a sample set, and generate a training set, a validation set, and a test set; wherein the seismic phase arrival time discrimination related data includes: S101 earthquake event waveform data, S102 corresponding seismic phase arrival time association information, and correct seismic phase arrival time type;

[0042] The S101 earthquake event waveform data consists of four channels. The first three channels are the Z, N, and E component data of the earthquake event waveform, and the fourth channel is represented by a sharp pulse as the time position of the earthquake phase.

[0043] S102 phase arrival-time correlation information includes: location, epicenter distance, focal depth, magnitude, etc. When collecting phase arrival-time correlation information, other information is also obtained, such as: average Moho depth, average station elevation, average focal depth, crustal P-wave velocity, S-wave velocity, upper mantle top P-wave velocity, S-wave velocity, etc.;

[0044] The label set in the sample set is the S103 correct earthquake phase arrival type.

[0045] At the same time, the waveform of the S101 earthquake event strictly corresponds to the corresponding seismic phase arrival time information of S102.

[0046] Before establishing the sample set, data preprocessing is performed, including normalizing, removing the mean, removing the linear trend, and performing data enhancement operations on the earthquake event waveform data.

[0047] The data enhancement includes: data window shifting, data mirroring, data filtering and data balancing;

[0048] Among them, data window shifting refers to the process of randomly moving the data cropping time window forward or backward in the process of cropping the waveform data to a set number of sampling points; data mirroring refers to the process of multiplying the waveform data by -1, and data filtering refers to the process of filtering the waveform data in different frequency bands.

[0049] The types of seismic phases that can be produced within different epicentral distance ranges must be divided into near and far epicentral distance ranges according to the specific geographical area of ​​application. There are differences in the types of seismic phases that can be distinguished within different epicentral distance ranges. Therefore, in the present invention, data enhancement processing also needs to increase data balancing.

[0050] Data balancing means: in order to ensure that there will be no deviation caused by imbalanced training data during the seismic phase identification process, the data is balanced and classified according to different seismic phase types, thereby ensuring that the amount of data under different seismic phase types is equal.

[0051] In the present invention, the boundaries of different epicentral distance ranges are calculated based on the velocity, Moho depth and ray theory formula of the specific research area; at the same time, in the data sets within different epicentral distance ranges, the data volume between each seismic phase type needs to be consistent, and the data volume between each seismic phase type within different magnitude ranges must also be kept equal. Based on this, the method for determining data balance is:

[0052] Determine a first epicentral distance range corresponding to Pg and Sg and a second epicentral distance range corresponding to Pg, Pn, Sg, Sn;

[0053] The earthquake phase arrival time discrimination related data are divided into two parts according to the first epicentral distance range and the second epicentral distance range; wherein the boundary epicentral distance between the first epicentral distance range and the second epicentral distance range is obtained according to calculation, the minimum boundary of the first epicentral distance range is the minimum epicentral distance in the sample set, and the maximum boundary of the second epicentral distance range is the maximum epicentral distance in the sample set;

[0054] The calculation method of the boundary epicentral distance of two epicentral distance ranges is:

[0055] Among them, Δ is the boundary epicenter distance, H is the average Moho depth of the study area, h is the average station elevation, d is the average focal depth, are the average velocities of P and S waves in the crust of the study area, are the average velocities of P and S waves at the top of the upper mantle in the study area, respectively; aver refers to the average value.

[0056] The seismic phase types that can be generated in each epicentral distance range are extracted from the two divided data parts respectively; when extracting, the amount of data of different seismic phase types in each epicentral distance range is equal.

[0057] The preprocessed earthquake event waveform data (including three components: Z, N, and E) and the corresponding phase arrival positions and other earthquake information (such as epicentral distance, focal depth, magnitude, etc.) are used as feature data sets, and the corresponding correct phase arrival types (such as Pg, Pn, Sg, Sn) are used as label data sets.

[0058] Among them, the label composed of the correct seismic phase arrival type is composed of a one-dimensional vector of size 4. The positions of the vector elements correspond to the four seismic phase types Pg, Pn, Sg, and Sn from front to back. The value of the vector element represents the probability of each seismic phase type. The position with a value of 1 represents the correct seismic phase type corresponding to each seismic phase arrival time in the label data set.

[0059] The present invention provides a specific implementation case: taking the eastern region of China as an example, waveform data of near-earthquake earthquake events within 500 km of the epicenter of fixed stations in the eastern region of China and corresponding seismic phase arrival type data and related information data are collected as a training data set, the length of the earthquake event waveform data is 10,000 sampling points, the sampling frequency is 100 Hz, and 2-8 Hz Butterworth bandpass filtering, de-meaning, de-linear trending and normalization operations are performed on the data, and the corresponding seismic phase types include Pg, Pn, Sg, and Sn.

[0060] In order to ensure that the earthquake event waveform and the type of phase arrival time are strictly corresponding, the phase arrival time position is integrated with the earthquake event waveform data, and the waveform data is expanded into four channels, such as Figure 5 As shown in the figure, the first three channels are the normalized Z, N, and E component data of the earthquake event waveform, and the fourth channel is represented by a sharp pulse as the time position of the earthquake phase, as shown in the figure. Figure 5 In the fourth channel, the horizontal axis arrival time position is a sharp pulse with a value of 3, and the values ​​of other positions on the horizontal axis are all 0. The data of the four channels are combined with the corresponding related information data (such as epicentral distance, focal depth, magnitude, etc.) as a feature data set.

[0061] For label processing: a one-dimensional vector of size 4 is used as a label data. The positions of the vector elements correspond to the four types of seismic phases Pg, Pn, Sg, and Sn from front to back, and the size of the vector elements themselves represents the probability of each type of seismic phase. The position with a value of 1 represents the correct seismic phase type corresponding to each phase arrival time in the label data set. For example, (1, 0, 0, 0) represents that the current phase arrival time position corresponds to the Pg phase, or (0, 0, 0, 1) represents that the current phase arrival time position corresponds to the Sn phase.

[0062] At the same time, the earthquake event waveform data is subjected to data window shifting, data mirroring and data filtering to further increase the richness of the training data and prevent overfitting. The specific processing contents include:

[0063] 1) Data window shifting: randomly move the data cropping time window forward or backward in the process of cropping the waveform data into 10,000 sampling points. The maximum moving range set in this implementation is ±5s;

[0064] 2) Data mirroring: Multiply the waveform data by -1 to form mirror data;

[0065] 3) Data filtering: The waveform data is further filtered based on the 2-8 Hz bandpass filtering. This implementation also performs Butterworth bandpass filtering operations in three frequency bands: 2-5 Hz, 5-8 Hz, and 3-6 Hz. At the same time, to ensure that the position of the seismic phase remains unchanged, the number of channels of the Butterworth filter is 1.

[0066] The waveform data before and after enhancement are used together to construct a training feature set, and the corresponding phase arrival time, phase type and other earthquake information are kept consistent with the data before enhancement.

[0067] In the sample data of the implementation case of the present invention, the maximum epicentral distance is 500km, that is, the second epicentral distance range is 500km; according to the boundary epicentral distance calculation method of the two epicentral distance ranges, the boundary epicentral distance is calculated to be 110km. The data is divided into two parts according to the epicentral distance range of 0-110km (corresponding to Pg, Sg) and 110-500km (corresponding to Pg, Pn, Sg, Sn), so that the data amount of different seismic phase types in each epicentral distance range is equal, corresponding to the seismic phase types that can be generated in each epicentral distance range.

[0068] On the premise of satisfying the above data classification, the data set is further divided into training set, validation set and test set, for example, the ratio is 8:1:1, where the training set data is used for training the network model, the validation set data is used for adjusting the training parameters during the training process, and the test set data is used for testing the accuracy of the model after training.

[0069] Step S110: establishing a neural network;

[0070] In this step, a deep learning network based on a multi-layer convolutional neural network is established, the ASFF module is used to perform multi-level feature fusion of seismic waveform data, and the low-rank fusion module is combined to realize the multimodal fusion between seismic waveform features and seismic information. The cross entropy is used as the loss function of the network.

[0071] Specifically, the neural network includes: a network input layer, a convolution group, a multi-level feature fusion layer, a multimodal fusion layer, a fully connected layer and an output layer.

[0072] The S111 network input layer loads the earthquake event waveform data; the waveform data and the arrival position together form a four-channel one-dimensional sequence, where the first three channels are the Z, N, and E components of the earthquake waveform data and the data is normalized, and the fourth channel is the arrival position channel, and the fourth channel has a pulse with a value of 3 only at the arrival position, and the rest of the position data are 0.

[0073] In the implementation case, the network input layer size is 10000×4, and the input data is the preprocessed earthquake event waveform data in the feature data set and the corresponding earthquake phase arrival time position.

[0074] S112 convolution layer: The first two convolution groups each include two convolution layers and one maximum pooling layer. The last three convolution groups each include three convolution layers and one maximum pooling layer. Each convolution group extracts features once, and a total of five layers of features are obtained, which are fused through a multi-level feature fusion layer.

[0075] In the implementation case, the convolution layer consists of five convolution groups, which are used to extract the feature information of the data transmitted from the input layer. The first two convolution groups each include two convolution layers and one maximum pooling layer, and the last three convolution groups each include three convolution layers and one maximum pooling layer. The convolution kernel size is 3×1, the convolution kernel sliding step is 1, and the edge padding size is 1. The extracted features are retained in the last maximum pooling layer of each convolution group, and a total of five features with different depths are obtained. The calculation principle of the convolution layer is as follows:

[0076] b n+1 =h n *K n +b n

[0077] Among them, b n and h n+1 are the input features and output features of the convolutional layer, K n is the convolution kernel, b n is the bias term. The relationship between the input data and the output features is:

[0078] W n+1 =(W n -F+2×P) / S+1

[0079] H n+1 =(H n -F+2×P) / S+1

[0080] D n+1 =K

[0081] Among them, W n and W n+1 is the width of the feature map before and after convolution, H n and H n+1 It is the height of the feature map before and after convolution, F is the convolution kernel size, P is the edge padding size, S is the convolution kernel sliding step size, and K is the number of convolution kernels.

[0082] S113: In the multi-level feature fusion layer, the ASFF (adaptive spatial feature fusion) module is used to fuse the feature information of different depths extracted from each convolution group. After the fused features are expanded, subsequent multimodal fusion operations are performed with other input earthquake information (such as epicenter distance, magnitude, focal depth, etc.).

[0083] The ASFF module fuses feature information of different depths extracted from different convolution groups through scale unification and adaptive fusion.

[0084] The ASFF calculation principle is as follows:

[0085]

[0086] in, It is the feature vector of the feature map of different depths and scales unified to the feature map of the lth layer at position (i, j). is the output feature map y l The (i,j)th vector along the channel. In this implementation, the spatial importance weights of the feature maps of five different depths can be obtained through network adaptive learning. The scale unification is performed by combining upsampling with convolution and downsampling with maximum pooling. In this implementation, l = 4. Figure 3 shown.

[0087] The S114 multimodal fusion layer loads the earthquake phase arrival time correlation information; in the multimodal fusion layer, a low-rank fusion method is used to input multiple earthquake-related information including epicenter distance, magnitude, focal depth, etc., and fuse the data with the incoming earthquake waveform feature information, and then connect with the subsequent fully connected layer.

[0088] In this implementation case, the multimodal fusion part consists of a flattening layer and a low-rank fusion module. The flattening layer is used to expand the multi-level fusion features extracted from the convolution part into a one-dimensional vector, so as to perform multimodal fusion with the earthquake-related information in the input feature data set (such as epicenter distance, focal depth, magnitude, etc.).

[0089] The calculation principle of low-rank fusion is as follows:

[0090]

[0091] Among them, z wave is the waveform multi-level feature encoding vector of the earthquake event waveform data extracted by the convolution part, z dist 、z mag 、z depth are the coding vectors of the additional input focal depth, epicenter distance and magnitude, which are composed of two parts. The first part is the waveform feature or seismic information data, and the second part is an element with a value of "1" added after the information data. (i) is the i-th low-rank factor, which can be obtained through network adaptive learning. In this implementation, i=75 is selected. Symbol Represents the element-wise product of a tensor, as Figure 4 shown.

[0092] S115 Fully connected layer: The neural network of the present invention includes three fully connected layers. The first two fully connected layers each include 512 nodes. The last fully connected layer is the output layer, which includes 4 nodes. Each node corresponds to a seismic phase type. The value of the node indicates the probability corresponding to each seismic phase identified, corresponding to the label data set in the training process and the output result in the prediction process. The sum of the probabilities of each node is 1. The seismic phase type with the largest probability is set as the correct seismic phase arrival type corresponding to the arrival position. If the probability of all seismic phases is less than 50%, it is determined that the arrival position does not belong to any seismic phase.

[0093] In the embodiment of the present invention, a softmax classifier is used;

[0094] This implementation selects cross entropy as the loss function of the network. The cross entropy calculation principle is as follows:

[0095]

[0096] Where M represents the number of samples in the training set, p is the expected output, i.e. the training label, and q is the actual output, i.e. the prediction result. When p and q are completely equal, the cross entropy is 0.

[0097] The predicted result is the output layer content, which is the seismic phase discrimination result and the corresponding probability:

[0098] Step S120: Using the sample set, the neural network is trained to construct a seismic phase arrival time discrimination model.

[0099] During the model training process, the earthquake event waveform data and the corresponding seismic phase arrival time positions contained in the training set in the feature data set are loaded into the network input layer, and the corresponding earthquake event information (such as epicenter distance, focal depth, and magnitude) is loaded into the multimodal fusion layer. The correct seismic phase type contained in the training set in the label data set is loaded into the output layer. Then, the network model is trained in combination with the testing of the training set and the validation set to construct a seismic phase arrival time discrimination model.

[0100] After the phase arrival time discrimination model is completed, the phase arrival time data is combined with the corresponding seismic waveform and input into the input layer of the phase arrival time discrimination model, and the relevant information of the earthquake event (such as epicenter distance, focal depth, magnitude) is input into the multimodal fusion layer of the network model, so that the correct discrimination result for the phase type can be obtained at the output layer through network calculation.

[0101] In controlling the number of training samples, the ADAM optimizer is used, and the number of batches can be calculated based on the actual amount of training set data and the storage size of the computing device.

[0102] In the implementation case, the ADAM optimizer with adaptively adjustable learning rate was used, and the training batch size was set to 512, that is, each batch of data input to the network consisted of 512 four-channel data and corresponding seismic phase types and earthquake information. The initial learning rate was set to 0.001, and the early stopping strategy was used to prevent overfitting. The training was terminated when the residual data of the validation set did not decrease significantly within 5 training rounds. At this time, the model construction was completed when the seismic phase arrived.

[0103] In this implementation, the number of training rounds is 109, and the cross entropy residual finally converges to 0.234. The accuracy of the model validation set reaches 91.47%, where the accuracy refers to the proportion of correct phase category judgments in all samples. The recall rate reaches 90.97%, where the recall rate is also called the recall rate, which refers to the proportion of correct phase category judgments among all results that should be judged as a certain phase. The precision rate reaches 91.87%, where the precision rate is also called the precision rate, which refers to the proportion of correct phase category judgments among all results that have been judged as a certain phase. Figure 6 As shown, similar results were obtained in the test set, with the model test set accuracy being 91.32%, the recall being 90.33%, and the precision being 92.02%.

[0104] When applying the phase arrival time discrimination model, the earthquake event waveform data of the phase arrival time that needs to be discriminated and corrected is input into the input layer of the phase arrival time discrimination model, and the phase arrival time related information (such as epicentral distance, focal depth, magnitude, etc.) is input into the multimodal fusion layer. After the operation of the neural network, the result obtained in the output layer is the probability of belonging to each phase type. The sum of all probabilities is 1. The phase type with the largest probability is set as the correct phase arrival type corresponding to the arrival position. If the probabilities of all phases are less than 50%, it is determined that the arrival position does not belong to any phase.

[0105] To further demonstrate the effectiveness and robustness of the phase arrival time discrimination model, the present invention also selected waveform data collected from fixed stations of the China Earthquake Networks Network, manually picked up the phases to obtain the correct phase arrival time positions, performed various preprocessing operations on the data, and finally input them into the trained network model for prediction, and obtained the final discrimination results as shown in the figure. Figure 7 As shown. It can be seen that after being processed by the network, this implementation has achieved good recognition and discrimination effects. The network correctly discriminates the type of seismic phase corresponding to the arrival position, and has a stable discrimination effect at different stations and different epicentral distances.

[0106] The present invention adopts the seismic waveform data enhancement operation to enrich the training set data, prevent overfitting in the neural network model training, and ensure a higher recognition accuracy in the subsequent seismic phase identification process; at the same time, the data set classification processing strategy enables the seismic phase type data of the training set of the present invention to be strictly balanced under different epicentral distances, different magnitudes, etc., to ensure that there will be no recognition deviation caused by imbalanced training data in the subsequent seismic phase identification process; the four-channel data input mode is specifically adopted, so that the present invention can strictly combine the seismic waveform information with the seismic phase arrival position, so as to realize the direct fusion of the waveform sequence and the arrival position feature in the subsequent feature extraction process; the ASFF module is adopted in the neural network, so that the present invention can extract multi-level feature information of the seismic waveform, thereby further improving the comprehensiveness and accuracy of the waveform data feature extraction; and the multimodal fusion layer enables the present invention to add other forms of seismic information while reading the waveform sequence information, ultimately improving the constraint ability and prediction accuracy of the network.

[0107] The above disclosures are only several specific embodiments of the present invention; however, the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for constructing a seismic phase arrival time discrimination model, characterized in that: The following steps are involved: Acquire phase arrival time discrimination related data, establish a sample set, and generate a training set, a validation set, and a test set; wherein the phase arrival time discrimination related data includes: earthquake event waveform data, corresponding phase arrival time associated information, and correct phase arrival time type; the label set in the sample set is the correct phase arrival time type; Establishing a neural network; the neural network comprises: a network input layer, a convolution group, a multi-level feature fusion layer, a multi-modal fusion layer, a fully connected layer and an output layer; wherein the network input layer loads earthquake event waveform data, the multi-modal fusion layer loads seismic phase arrival time association information, and the output layer loads the correct seismic phase arrival time type; The neural network is trained by using the sample set to construct a seismic phase arrival time discrimination model.

2. The construction method according to claim 1, characterized in that: Before establishing the sample set, data preprocessing is performed, including data enhancement and data balancing; The data enhancement is performed on the seismic event waveform data, and the data enhancement includes: data window shifting, data mirroring, and data filtering; The data balancing is an operation to ensure that there is no deviation caused by imbalanced training data during the phase identification process. The data balancing refers to: classifying the data in a balanced manner according to different phase types to ensure that the amount of data under different phase types is equal.

3. The construction method according to claim 1, characterized in that: The earthquake event waveform strictly corresponds to the corresponding earthquake phase arrival time correlation information; Among them, the waveform data is four channels, the first three channels are the Z, N, and E component data of the waveform of the passing earthquake event, and the fourth channel is represented by a sharp pulse as the time position of the earthquake phase.

4. The construction method according to claim 1, characterized in that: The label composed of the correct seismic phase arrival type is composed of a one-dimensional vector of size 4. The positions of the vector elements correspond to the four seismic phase types Pg, Pn, Sg, and Sn from front to back. The value of the vector element represents the probability of each seismic phase type. The position with a value of 1 represents the correct seismic phase type corresponding to each seismic phase arrival time in the label data set.

5. The construction method according to claim 1, characterized in that: The earthquake phase arrival time associated information includes: location, epicenter distance, focal depth, and magnitude; After the multimodal fusion layer loads the seismic phase arrival time correlation information, the seismic phase arrival time correlation information is multimodally fused with the characteristics of the earthquake event waveform data, which is reflected as follows: Among them, z wave is the multi-level feature encoding vector of earthquake event waveform data, z dist 、z mag 、z depth are the encoding vectors of focal depth, epicenter distance and magnitude, respectively, w (i) is the i-th low-rank factor, and the symbol "°" represents the element-wise product of the tensor.

6. The construction method according to claim 2, characterized in that: The method for implementing the data balancing is as follows: Determine a first epicentral distance range corresponding to Pg and Sg and a second epicentral distance range corresponding to Pg, Pn, Sg, Sn; The seismic phase arrival time discrimination related data are divided into two parts according to the first epicentral distance range and the second epicentral distance range; wherein the boundary epicentral distance between the first epicentral distance range and the second epicentral distance range is obtained according to calculation, the minimum boundary of the first epicentral distance range is the minimum epicentral distance in the sample set, and the maximum boundary of the second epicentral distance range is the maximum epicentral distance in the sample set; The seismic phase types that can be generated in each epicentral distance range are extracted from the two parts; when extracting, the data amounts of different seismic phase types in each epicentral distance range are equal.

7. The construction method according to claim 6, characterized in that: The calculation method of the boundary epicentral distance between the first epicentral distance range and the second epicentral distance range is: Among them, Δ is the boundary epicenter distance, H is the average Moho depth of the study area, h is the average station elevation, d is the average focal depth, are the average velocities of P and S waves in the crust of the study area, are the average velocities of P and S waves at the top of the upper mantle in the study area, respectively; aver refers to the average value.

8. An application method of a seismic phase arrival time discrimination model, characterized in that: The seismic phase arrival time discrimination model is constructed by the construction method according to claims 1 to 7, and the application method comprises the following steps: Inputting earthquake event waveform data into the input layer of the earthquake phase arrival time discrimination model; Inputting the seismic phase arrival time correlation information into the multimodal fusion layer of the seismic phase arrival time discrimination model; After the operation of the neural network, the result obtained in the output layer is the probability of each seismic phase type, and the correct seismic phase arrival type is determined based on the probability.

9. The application method according to claim 8, characterized in that: Judging the correct seismic phase arrival type based on probability includes: the seismic phase type with the greatest probability is the correct seismic phase arrival type corresponding to the arrival position.

10. The application method according to claim 9, characterized in that: Judging the correct type of earthquake phase arrival time according to probability also includes: if the probability of all earthquake phases is less than 50%, it is judged that the arrival position does not belong to any earthquake phase.

Images