Earthquake phase automatic identification method based on dense connection and U-shaped structure combination

Abstract

The application belongs to the field of seismic signal detection and estimation, and is a seismic phase automatic identification method based on the combination of dense connection and U-shaped structure. One-dimensional microseismic signals are saved as an image, target objects in the picture are framed and calibrated according to image calibration information, a data set is obtained, a one-dimensional neural network model for seismic phase picking is designed and trained, the data set is trained, the one-dimensional neural network model is obtained, the seismic signal to be detected is input for intuitive testing, and the seismic occurrence time and position are obtained according to the obtained test results. Therefore, accurate identification and time estimation of seismic phases are realized, and accurate seismic event position and time are obtained.

Description

Technical Field

[0001] This invention belongs to the field of earthquake signal detection and estimation, and is an automatic earthquake phase identification method based on the combination of dense connection and U-shaped structure. Background Technology

[0002] Seismic phase picking is a crucial step in seismic station data processing, involving signal detection, phase identification, and arrival time estimation based on data recorded by seismic monitoring stations. Currently, three main methods are used for seismic phase picking: manual picking, traditional methods, and deep learning methods. Manual picking, where professionals analyze seismic waveforms and identify P-waves, offers the highest accuracy but is extremely inefficient. Traditional methods, such as those based on the Short-Time Average / Long-Time Average Ratio (STA / LTA) method for signal detection, the Autoregressive Akaike Information Criterion (AR-AIC) for accurate arrival time estimation, and high-order statistical methods like skewness and kurtosis, all focus on one or a few features of the seismic phase, failing to encompass all characteristics of the waveform data. Furthermore, these methods require extensive preprocessing to extract necessary features and carefully set detection thresholds, all of which significantly impact the results. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide an automatic earthquake phase identification method based on a combination of dense connections and a U-shaped structure. This enables accurate identification of earthquake phases and estimation of their arrival time, as well as accurate location and time of earthquake events.

[0004] This invention is implemented as follows:

[0005] An automatic seismic facies identification method based on a combination of dense connectivity and U-shaped structure includes the following steps:

[0006] S1: The effective signals in the seismic record are synthesized using the Ricker wavelet to obtain N microseismic signals, each of which contains microseismic events;

[0007] S2: The synthesized seismic dataset is preprocessed using LabelMe, a tool used in the field of computer vision to create labeled image data. The one-dimensional microseismic signal is saved as an image. According to the image labeling information, the target objects in the image are bounded with position boxes and labeled with type. The specific coordinates and category information are saved to a JSON file. The labeling method is to mark the effective signal as 1 and the noise signal as 0. N microseismic signal samples and N noise samples are obtained, for a total of 2N samples in the dataset.

[0008] S3: Design and train a one-dimensional neural network model for seismic phase picking. The earthquake training dataset from step S2 is used for training. The JSON format dataset in S2 is converted into an npy format dataset using Python. The test set and training set are divided into a 2:8 ratio and input into the one-dimensional neural network model for deep learning. The training set data undergoes multiple convolution and pooling processes. The encoded output feature map is continuously upsampled to gradually obtain a segmentation map with the same size as the original input.

[0009] S4: After training is complete, save the one-dimensional neural network model and input the seismic signal to be detected for intuitive testing;

[0010] S5: Obtain the time and location of the earthquake based on the test results obtained in S4.

[0011] Furthermore, the effective signals in the specific Rayleigh wavelet synthesized seismic records in S1 are: the dominant frequency is between 15Hz and 30Hz, and the waveform data is extracted with a window of 1.6S, as shown in the following formula:

[0012]

[0013] Where A represents amplitude, t0 represents start time, and f0 represents dominant frequency.

[0014] Furthermore,

[0015] S3 specifically includes:

[0016] Furthermore, the one-dimensional neural network model includes encoder downsampling, decoder upsampling, and skip paths. Encoder downsampling includes the first convolutional block X passed down sequentially. (0，0) The second convolutional block X (1，0) The third convolutional block X (2，0) And the fourth convolutional block. The first convolutional block defines 16 filters with a height of 32 and a max pooling layer, using the ReLU function as the activation function. In the subsequent convolutional blocks, the number of filters is doubled one by one, while the other parameters remain unchanged. No max pooling layer is connected after the fourth convolutional block.

[0017] Decoder upsampling includes a transposed convolutional layer X passed up sequentially. (2，1) A clipping layer X (1，2) and a convolutional layer X (0，3) The three upsampling layers of the decoder are symmetrical to the convolutional blocks of the first three downsampling layers of the encoder, and the transposed convolutional layer X (2，1) Connect the fourth convolutional block to perform transposed convolution and interconnection operations between left and right symmetric layers;

[0018] The jump path includes dense jump connections formed by three convolutional blocks, forming multiple jump paths between decoder upsampling and encoder downsampling. The first three convolutional blocks in the encoder downsampling jump to the decoder upsampling layer through the convolutional blocks in the dense jump connections.

[0019] Furthermore: In S3, the cross-entropy loss function is set to the loss function used during network training, defined as follows:

[0020]

[0021] L is the loss value, y (i) For label values, These are predicted values.

[0022] Furthermore: the first convolutional block X (0，0) The jumps between the upsampling layer and the convolutional block X are: individual convolutional blocks X that pass through the jump path. (0，1) Convolutional block X via jump path (0，2) and through convolutional block X (0，1) and convolution block X (0，2) Jump to the convolutional layer X of the upsampling layer (0，3) ;

[0023] Second convolutional block X (1，0) The jumps between the upsampling layer and the clipping layer X include: a separate jump to the clipping layer X. (1，2) The convolutional block X via the jump path (1，1) Jump to clipping layer X (1，2) Jump to convolution block X (0，1) convolutional block X (1，1) Can jump to convolution block X (0，2) .

[0024] Third convolutional block X (2，0) The jumps between the upsampling layer and the upsampling layer include: a separate jump to the transposed convolutional layer X. (2 ，1) Jump to convolution block X (1，1) .

[0025] Compared with existing technologies, the advantages of this invention are as follows: This invention can accurately distinguish the locations of different types of signal segments, and after network training, there is no need to manually adjust parameters or set thresholds. The invention's high detection capability in earthquake monitoring data has been verified using synthetic simulation data and real-world earthquake data from the Southern California Earthquake Network. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method for seismic phase picking according to the present invention;

[0027] Figure 2The model shape and parameter diagram provided for embodiments of the present invention;

[0028] Figure 3 The joint loss curve of the dataset and validation set provided in the embodiments of the present invention;

[0029] Figure 4 The signal detection results provided in this embodiment of the invention are as follows: (a) is the detection result of the simulated seismic signal to be detected, and (b) is the detection result of the measured seismic signal. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0031] Please refer to Figure 1 This invention provides an automatic seismic phase identification method based on a combination of dense connectivity and a U-shaped structure, comprising:

[0032] S1: 3000 microseismic signals were obtained by synthesizing the effective signals from the seismic records using the Ricker wavelet, with each signal containing microseismic events. Specifically, the effective signals from the synthesized seismic records in S1 have a dominant frequency between 15Hz and 30Hz, with waveform data extracted in 1.6-second windows, as shown in the following formula:

[0033]

[0034] Where A represents amplitude, t0 represents start time, and f0 represents dominant frequency.

[0035] S2: The synthesized seismic dataset is preprocessed using LabelMe, a tool used in the field of computer vision to create labeled image data. The one-dimensional microseismic signal is saved as an image. According to the image labeling information, the target objects in the image are bounded with position boxes and labeled with type. The specific coordinates and category information are saved to a JSON file. The labeling method is to mark the effective signal as 1 and the noise signal as 0. A training dataset of 6,000 samples is obtained, consisting of 3,000 microseismic signal samples and 3,000 noise samples.

[0036] S3: Design and train a one-dimensional neural network model for seismic phase picking, starting with the seismic training dataset from step S2; design and train the one-dimensional neural network model for seismic phase picking, using the seismic training dataset from step S2. Convert the JSON format dataset from S2 into an .npy format dataset using Python, and divide it into a test set and a training set at a 2:8 ratio. Input the data into the one-dimensional neural network model for deep learning; the training data undergoes multiple convolution and pooling processes, which is the image encoding process and also a downsampling process, followed by decoding. Decoding can be understood as the inverse operation of encoding. The output feature map of the encoding is continuously upsampled to gradually obtain a segmentation map, i.e., the output, with the same size as the original input; set the cross-entropy loss function as the loss function in network training, defined as follows:

[0037]

[0038] L is the loss value, y (i) For label values, These are predicted values.

[0039] S4: After training is complete, save the model, input the waveform data to be detected into the neural network obtained in S3, and obtain the phase category and accurate phase arrival time estimate;

[0040] S5: Determine the time and location of the earthquake based on the picked-up seismic phases.

[0041] Please refer to Figure 2 Following the above method, construct a neural network model with the following structure: Figure 2 As shown. The neural network model consists of encoding and decoding, as well as a series of nested, dense convolutional blocks connecting them. The one-dimensional neural network model includes encoder downsampling, decoder upsampling, and skip paths. The encoder downsampling consists of four convolutional blocks. The first convolutional block defines 16 filters of height 32 and a max pooling layer, using the ReLU function as the activation function. In subsequent convolutional blocks, the number of filters doubles, while the other parameters remain unchanged. No max pooling layer is connected after the fourth convolutional block.

[0042] The decoder upsampling includes a convolutional layer, a cropping layer, and a transposed convolutional layer. The three layers of decoder upsampling are symmetrical to the convolutional blocks of the first three layers of encoder downsampling, and are used to perform transposed convolution and interconnection operations between the left and right symmetrical layers.

[0043] The skip path comprises dense skip connections formed by three convolutional blocks, creating multiple skip paths between decoder upsampling and encoder downsampling. The first three convolutional blocks in the encoder downsampling skip to the decoder upsampling layer via convolutional blocks in the dense skip connections.

[0044] The first half performs feature extraction, which is downsampling, and the second half performs upsampling. This structure is called an encoder-decoder structure.

[0045] The first convolutional block of the encoding part (X) (0,0) The code defines 16 filters (also called feature detectors) with a height of 32 (also known as kernel size) and a max-pooling layer; here, a pooling layer of size 3 is chosen, using the ReLU function as the activation function. In the subsequent convolutional block (X... (1,0) ,X (2,0) ,X (3,0) The number of filters is doubled one by one, while other parameters remain unchanged. It is worth noting that the bottom layer, i.e., the fourth layer (X), is the most important. (3,0) No max pooling operation is performed after convolution.

[0046] The decoding part consists of a transposed convolutional layer X (2，1) A clipping layer X (1，2) A convolutional layer X (0，3) The system consists of three downsampling layers on the left that perform convolution and pooling operations to extract abstract features of seismic phases, which helps solve the problem of seismic phase localization. The three upsampling layers on the right perform transposed convolution and interconnection operations between the left and right symmetrical layers, which can gradually restore the detailed features of the seismic phases, thus helping to solve the problem of seismic phase classification.

[0047] Among them, the first convolutional block X (0，0) The jumps between the upsampling layer and the convolutional block X are: individual convolutional blocks X that pass through the jump path. (0，1) Convolutional block X via jump path (0，2) and through convolutional block X (0，1) and convolution block X (0，2) Jump to the convolutional layer X of the upsampling layer (0，3) ;

[0048] Second convolutional block X (1，0) The jumps between the upsampling layer and the clipping layer X include: a separate jump to the clipping layer X. (1，2) The convolutional block X via the jump path (1，1) Jump to clipping layer X (1，2) Jump to convolution block X (0，1) convolutional block X (1，1) Can jump to convolution block X (0，2) .

[0049] Third convolutional block X (2，0) The jumps between the upsampling layer and the upsampling layer include: a separate jump to the transposed convolutional layer X. (2 ，1) Jump to convolution block X (1，1) .

[0050] The redesigned skip paths bridge the semantic gap between the encoder and decoder sub-paths. This makes the semantic level of the encoded features closer to the semantic level of the feature maps waiting in the decoder, thus making optimization easier when semantically similar feature maps are received. Dense skip connections implement skip paths between the encoder and decoder. The aim is to improve segmentation accuracy and gradient flow. Dense skip connections ensure that all prior feature maps are accumulated and reach the current node through convolutional blocks on each skip path. By ensuring that all feature maps are collected and reach the correct nodes, dense skip connections bridge the semantic gap between the encoder and decoder feature maps, helping to improve gradient flow.

[0051] Figure 3 This shows how the joint loss value decreases with the number of iterations during model training. It can be seen that the loss value of the validation set reaches the minimum value of 0.075 in the 30th iteration, at which point the phase recognition accuracy of the validation set reaches 99.23%. The model was trained on a single NVIDIA TITAN V graphics processing unit, and the training time was approximately 4 hours, 40 minutes, and 11 seconds.

[0052] Figure 4 The image shows the final detection results of the simulated seismic signal to be detected and the measured seismic signal from the Southern California Seismic Network, as shown below. Figure 4 As shown in (a), this is a segmentation map with the same size as the original input. The black line represents a whole segment of seismic signal containing both valid and noise signals, while the light-colored line in the middle represents the valid signal. The range from the rising to the falling line is the range of the valid signal. Figure 4 (b) shows the detection results of measured seismic signals. The ability to accurately distinguish the locations of different signal segments demonstrates the high detection capability of this invention in seismic monitoring data. The performance of seismic phase identification was evaluated using precision and recall metrics. First, the following confusion matrix was defined, as shown in Table 1.

[0053] Table 1 Definition of Confusion Matrix

[0054]

[0055] Define the precision and recall of the P-phase as follows:

[0056]

[0057] The test results of the method of this invention were compared in terms of precision and recall. A high precision indicates a low false positive rate, and a high recall indicates a low false negative rate. Only algorithms with both high precision and recall have practical value. The results are shown in Table 2. It can be seen that the method of this invention far surpasses the STA / LTA algorithm in terms of both precision and recall.

[0058] Table 2

[0059] Method of the present invention STA / LTA Algorithm Precision 97.62％ 76.45％ Recall 97.83％ 45.78％

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatic seismic facies identification based on a combination of dense connectivity and U-shaped structure, characterized in that, Specifically, the following steps are included: S1: The effective signals in the seismic record are synthesized using the Ricker wavelet to obtain N microseismic signals, each of which contains microseismic events; S2: The synthesized seismic dataset is preprocessed using LabelMe, a tool used in the field of computer vision to create labeled image data. The one-dimensional microseismic signal is saved as an image. According to the image labeling information, the target objects in the image are bounded with position boxes and labeled with type. The specific coordinates and category information are saved to a JSON file. The labeling method is to mark the effective signal as 1 and the noise signal as 0. N microseismic signal samples and N noise samples are obtained, for a total of 2N samples in the dataset. S3: Design and train a one-dimensional neural network model for earthquake phase picking. Use the earthquake training dataset from step S2 for training. Convert the JSON format dataset in S2 into an npy file format dataset using Python. Divide the test set and training set into a 2:8 ratio and input them into the one-dimensional neural network model for deep learning. The training set data undergoes multiple convolution and pooling processes, and the encoded output feature map is continuously upsampled to gradually obtain a segmentation map with the same size as the original input. S4: After training is complete, save the one-dimensional neural network model and input the seismic signal to be detected for intuitive testing; S5: Obtain the time and location of the earthquake based on the test results obtained in S4.

2. The method according to claim 1, characterized in that: The effective signals in the specific Ricker wavelet synthesized seismic records in S1 are: the dominant frequency is between 15Hz and 30Hz, and the waveform data is extracted with a window of 1.6S, as shown in the following formula: Where A represents amplitude, t0 represents start time, and f0 represents dominant frequency.

3. The method according to claim 1, characterized in that, S3 specifically includes: A one-dimensional neural network model includes encoder downsampling, decoder upsampling, and skip paths. Encoder downsampling includes the first convolutional block X passed down sequentially. (0，0) The second convolutional block X (1，0) The third convolutional block X (2，0) And the fourth convolutional block. The first convolutional block defines 16 filters with a height of 32 and a max pooling layer, using the ReLU function as the activation function. In the subsequent convolutional blocks, the number of filters is doubled one by one, while the other parameters remain unchanged. No max pooling layer is connected after the fourth convolutional block. Decoder upsampling includes a transposed convolutional layer X passed up sequentially. (2，1) A clipping layer X (2，1) and a convolutional layer X (0，3) The three upsampling layers of the decoder are symmetrical to the convolutional blocks of the first three downsampling layers of the encoder, and the transposed convolutional layer X (2，1) Connect the fourth convolutional block to perform transposed convolution and interconnection operations between left and right symmetric layers; The jump path includes dense jump connections formed by three convolutional blocks, forming multiple jump paths between decoder upsampling and encoder downsampling. The first three convolutional blocks in the encoder downsampling jump to the decoder upsampling layer through the convolutional blocks in the dense jump connections.

4. A method for picking up seismic phases according to claim 3, characterized in that: In S3, the cross-entropy loss function is set to the loss function used during network training, defined as follows: L is the loss value, y (i) For label values, These are predicted values.

5. A method for picking up seismic phases according to claim 3, characterized in that: First convolutional block X (0，0) The jumps between the upsampling layer and the convolutional block X are: individual convolutional blocks X that pass through the jump path. (0，1) Convolutional block X via jump path (0，2) and through convolutional block X (0，1) and convolution block X (0，2) Jump to the convolutional layer X of the upsampling layer (0，3) ; Second convolutional block X (1，0) The jumps between the upsampling layer and the clipping layer X include: a separate jump to the clipping layer X. (1，2) The convolutional block X via the jump path (1，1) Jump to clipping layer X (1，2) Jump to convolution block X (0，1) convolutional block X (1，1) Can jump to convolution block X (0，2) . Third convolutional block X (2，0) The jumps between the upsampling layer and the upsampling layer include: a separate jump to the transposed convolutional layer X. (2，1) Jump to convolution block X (1，1) .

Images