Seismic phase pickup method based on time-frequency analysis and deep learning

Abstract

The invention relates to a seismic phase pickup method based on time-frequency analysis and deep learning, and belongs to the field of seismology, a deep neural network comprising an empirical mode decomposition module, a residual shrinkage module, an LSTM layer, a transformer layer, an up-sampling layer and a convolutional layer is established, a linear rectification function is adopted as an activation function, an Adam adaptive optimization function is adopted for back propagation, and a seismic phase is obtained. And taking a binary cross entropy loss function as a loss function, and taking a mean absolute error function, an error standard deviation, an accuracy rate, a recall rate and an F1 score as evaluation indexes to obtain a deep neural network model. When the frequencies of the noise, the P wave and the S wave are different, empirical mode decomposition can effectively separate the noise, the P wave and the S wave, so that the network is helped to better identify the P wave and the S wave. Even for broadband noise, the noise can be decomposed into different intrinsic mode functions, so that the noise content in the intrinsic mode functions where P waves and S waves are located is reduced. The time-frequency analysis method can effectively improve the anti-noise capability of the network.

Description

Technical Field

[0001] The present invention relates to the field of seismology, and in particular to a seismic phase picking method based on time-frequency analysis and deep learning. Background Art

[0002] Earthquake phase picking is one of the most basic components in earthquake monitoring and earthquake early warning. The accurate picking of the arrival time of P-waves and S-waves not only improves the reliability of earthquake catalogs, but also provides strong support for the establishment of velocity models of P-waves and S-waves. In addition, the arrival time significantly affects the accuracy of earthquake source location. Even a 0.01 second error in picking up a P-wave can cause an error of tens of meters in earthquake source location.

[0003] At present, traditional P-wave and S-wave picking relies on manual picking by human analysts, but with the increase in the number of seismic stations, manual picking is difficult to cope with the huge amount of seismic data. Researchers have developed some automated algorithms, such as the short-time average / long-time average ratio method (STA / LTA) and the autoregression-Akaike information criterion method (AR-AIC). This type of automated algorithm identifies the arrival time of P-waves and S-waves based on the sudden changes in amplitude, energy and statistical characteristics in seismic records. However, this type of algorithm is easily affected by noise and requires setting the optimal initial parameters for different ground motions, which makes it inconvenient to use.

[0004] In recent years, with the development of computer technology and artificial intelligence, deep learning has been widely used in phase picking, but the complex and ubiquitous noise still greatly reduces the picking accuracy of the model. Therefore, based on the advantage of time-frequency analysis in effectively separating noise, a deep learning model based on empirical mode decomposition is proposed to pick up the arrival time of P waves and S waves. Summary of the invention

[0005] The purpose of this application is to solve the technical problems of low precision and insufficient noise resistance of the phase picking method, and to provide a deep learning method based on empirical mode decomposition.

[0006] The seismic phase picking method based on time-frequency analysis and deep learning includes the following steps:

[0007] Step 1: Collect earthquake records and build a data set:

[0008] Select known seismic motion data, pick up the P wave and S wave in the seismic motion data as the label value corresponding to the earthquake record, and obtain the seismic phase picking data set;

[0009] Step 2: Establish a deep neural network model combined with the empirical mode decomposition module:

[0010] The deep neural network consists of an encoder and a decoder. The first part of the encoder is an empirical mode decomposition module, which decomposes the earthquake motion into multiple intrinsic mode functions, each of which represents components of different frequencies in the earthquake motion. The second part is composed of multiple residual shrinkage modules in series. The residual shrinkage module adds a soft threshold function on the basis of the traditional residual module to filter out background noise. The third part is two long and short-term memory layers, which are used to extract temporal context information. The fourth part is a transformer module, which assigns weights to different positions in the sequence based on the attention mechanism to amplify the key features.

[0011] The decoder consists of multiple upsampling layers and convolutional layers to expand the feature map to the same length as the input;

[0012] Step 3: Deep neural network model training:

[0013] The deep neural network is trained, and the loss function and mean absolute error function are used to ensure the training accuracy, so that the loss function gradually decreases, and the trained deep neural network model is obtained;

[0014] Step 4: Seismic phase picking:

[0015] Input the three-component seismic motion record into the deep neural network model trained in step 3, output the probability sequence corresponding to the P wave and S wave, find the position of the maximum value of the probability sequence, and thus realize the seismic phase picking.

[0016] In step 2, the empirical mode decomposition of the encoder includes the following steps:

[0017] 1) Identify all local maxima and local minima in the signal;

[0018] 2) Interpolate local maxima to form an upper envelope, and interpolate local minima to form a lower envelope;

[0019] 3) Calculate the mean of the upper and lower envelopes;

[0020] 4) Subtract the mean from the signal to generate a new signal; if the new signal satisfies the conditions for being an IMF, it is considered an IMF; an IMF is a special type of signal that satisfies the number of extreme values ​​and the number of zero crossings, and the mean of the upper and lower envelopes must be close to zero;

[0021] 5) Subtract this IMF from the original signal and repeat the process for the remaining signal to continue extracting the next IMF.

[0022] Four maximum pooling layers are added in the encoder to reduce the input length; the encoder includes one empirical mode decomposition module, eight residual shrinkage modules, four maximum pooling layers, two LSTM layers and one transformer layer; the empirical mode decomposition module, the first residual shrinkage module, the first maximum pooling layer, the second residual shrinkage module, the second maximum pooling layer, the third residual shrinkage module, the third maximum pooling layer, the fourth residual shrinkage module, the fourth maximum pooling layer, the fifth residual shrinkage module, the sixth residual shrinkage module, the seventh residual shrinkage module, the eighth residual shrinkage module, the first LSTM layer, the second LSTM layer and the transformer layer are connected in sequence.

[0023] The decoder includes 4 upsampling layers and 5 1D convolution layers. The first upsampling layer, the first 1D convolution layer, the second upsampling layer, the second 1D convolution layer, the third upsampling layer, the third 1D convolution layer, the fourth upsampling layer, the fourth 1D convolution layer and the fifth 1D convolution layer are connected in sequence; a batch normalization layer is added after each convolution layer, and a linear rectification function is used as the activation function.

[0024] The loss function in step 3 is the binary cross entropy loss function, as shown in formula (2):

[0025]

[0026] It can measure the difference between the probability distribution of the model output and the label value; BCELoss is the binary cross entropy loss function; N is the number of sampling points in the probability sequence; y i is the probability value of the i-th position of the true probability sequence label; p i The probability value at the i-th position of the probability sequence predicted by the neural network.

[0027] In step 3, the deep neural network model was trained and the training data was augmented, including random shifting of the records and adding Gaussian white noise. The data augmentation method included random shifting of the records with a probability of 99% and adding Gaussian white noise with zero mean and standard deviation of 0.01 to 0.15 times the maximum amplitude of the records with a probability of 50%.

[0028] The batch size of the trained deep neural network model is 64, the training rounds are 25, the initial learning rate is 0.01, and when the loss of the validation set remains unchanged for 3 consecutive epochs, the learning rate is reduced to 1 / 10 of the original.

[0029] In step 1, a standardization operation is performed on each channel of each seismic data to serve as the final network input data;

[0030] The standardization operation is shown in formula (1):

[0031]

[0032] Where: x * is the data after standardization; μ is the mean of the data; σ is the standard deviation of the data; x is the data before standardization.

[0033] The beneficial effects of the present invention are:

[0034] The present invention proposes a deep neural network based on time-frequency analysis for picking up the arrival time of P waves and S waves. The network consists of an encoder and a decoder, and an empirical mode decomposition module is added to convert the seismic motion into an intrinsic mode function. At the same time, multiple residual shrinkage modules, LSTM and transformer modules are used to improve the network's learning ability for spatial and temporal information in seismic motion. The main inventive point of the present invention is to add an empirical mode decomposition module to the architecture of the traditional neural network. This module expresses the time-frequency characteristics of seismic motion by decomposing the seismic motion into intrinsic mode functions of different frequencies. When there are differences in the frequencies of noise, P waves and S waves, empirical mode decomposition can effectively separate the three, thereby helping the network to better identify P waves and S waves. Even for broadband noise, the noise will be decomposed into different intrinsic mode functions, thereby reducing the noise content in the intrinsic mode functions where P waves and S waves are located. This time-frequency analysis method can effectively improve the anti-noise ability of the network.

[0035] Compared with the current deep learning picking method, the seismic phase picking method based on time-frequency analysis and deep learning of the present invention has higher picking accuracy and better picking effect on noisy records with low signal-to-noise ratio. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is the distribution of selected earthquake epicenters, where the red circle represents the epicenter of the earthquake.

[0037] Figure 2 The distribution of magnitude and signal-to-noise ratio in the data set.

[0038] Figure 3 Schematic diagram of the residual shrinkage module of the network in the embodiment.

[0039] Figure 4 Schematic diagram of a deep neural network in an embodiment.

[0040] Figure 5 This is a comparison of the P-wave and S-wave picking results of the deep neural network model in the embodiment and the current optimal picking method EQTransformer on different signal-to-noise ratio records on the test set. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The seismic phase picking method based on time-frequency analysis and deep learning includes the following steps:

[0043] Step 1: Collect earthquake records and build a data set:

[0044] The P-wave and S-wave picking times given in the database are used as label values ​​corresponding to the earthquake records to obtain the earthquake phase picking data set.

[0045] Step 2: Establish a deep neural network model combined with the empirical mode decomposition module:

[0046] The deep neural network is based on a fully convolutional network architecture and consists of an encoder and a decoder. The encoder extracts features from the three-component seismic input and generates a high-dimensional feature map. The decoder converts the high-dimensional feature map into a one-dimensional probability sequence. The probability value of each point indicates the probability of the P wave and S wave arriving at that point.

[0047] The first part of the encoder is an empirical mode decomposition module, which decomposes the ground motion into multiple intrinsic mode functions, each of which represents a component of different frequencies in the ground motion. The second part is composed of multiple residual shrinkage modules in series. The residual shrinkage module adds a soft threshold function on the basis of the traditional residual module to filter out background noise. The third part is two long short-term memory layers (LSTM) for extracting temporal context information. The fourth part is a transformer module, which assigns weights to different positions in the sequence based on the attention mechanism to amplify the key part features. Four maximum pooling layers are added to the encoder to reduce the length of the input.

[0048] The decoder consists of multiple upsampling layers and convolutional layers to expand the feature map to the same length as the input. The network uses the linear rectifier function (ReLU) as the activation function, and the convolutional layers are all 1D convolutions. A batch normalization layer is added after each convolutional layer.

[0049] The loss function uses the binary cross entropy loss function, which can measure the difference between the probability distribution of the model output and the label value.

[0050] The empirical mode decomposition of the encoder can be achieved by the following steps:

[0051] 1) Identify all local maxima and local minima in the signal.

[0052] 2) Interpolate local maxima to form an upper envelope, and interpolate local minima to form a lower envelope.

[0053] 3) Calculate the mean of the upper and lower envelopes.

[0054] 4) Subtract the mean from the signal to generate a new signal. If this new signal meets the conditions to be an IMF, it is considered an IMF. An IMF is a special type of signal that satisfies the number of extreme values ​​and zero crossings, and the average of the upper and lower envelopes must be close to zero.

[0055] 5) Subtract this IMF from the original signal and repeat the process for the remaining signal to extract the next IMF.

[0056] Finally, the empirical mode decomposition decomposes the signal into five intrinsic mode functions and a residual seismic signal component.

[0057] Step 3: Deep neural network model training:

[0058] The deep neural network was trained using the seismic phase picking dataset, and the training data was enhanced, including recording random translation and adding Gaussian white noise. The loss function and mean absolute error (MAE) function were used to ensure the training accuracy, so that the loss function gradually decreased and the trained deep neural network model was obtained.

[0059] Step 4: Seismic phase picking:

[0060] Input the three-component seismic motion record into the deep neural network model trained in step 3, output the probability sequence corresponding to the P wave and S wave, find the position of the maximum value of the probability sequence, and thus realize the seismic phase picking.

[0061] Embodiment 1

[0062] The seismic phase picking method based on time-frequency analysis and deep learning in this embodiment is implemented according to the following steps:

[0063] Step 1: Collect earthquake records and build a data set:

[0064] 1,029,421 three-component ground motion records were selected from the STEAD database, and the epicenters were distributed as follows: Figure 1 As shown, the magnitude and signal-to-noise ratio Figure 2As shown. The P-wave and S-wave pickup times given in the database are used as the label values ​​corresponding to the seismic data. Based on the label values, a one-dimensional probability sequence with the same length as the recording time is generated. The probability of the true arrival point is set to 1 and linearly decreases to 0 within the range of 10 surrounding sampling points. The seismic phase picking data set is obtained, and a standardization operation is performed on each channel of each seismic data as the final network input data;

[0065] The standardization operation is shown in formula (1):

[0066]

[0067] Where: x * is the data after standardization; μ is the mean of the data; σ is the standard deviation of the data; x is the data before standardization.

[0068] The earthquake motion dataset is randomly divided into training set, validation set and test set, and the ratio of training set, validation set and test set is 17:1:2.

[0069] Step 2: Establish a deep neural network combined with time-frequency analysis:

[0070] like Figure 4 The deep neural network shown includes an encoder and a decoder, and the encoder includes 1 empirical mode decomposition module, 8 residual shrinkage modules, 4 maximum pooling layers, 2 LSTM layers and 1 transformer layer. The empirical mode decomposition module, the first residual shrinkage module, the first maximum pooling layer, the second residual shrinkage module, the second maximum pooling layer, the third residual shrinkage module, the third maximum pooling layer, the fourth residual shrinkage module, the fourth maximum pooling layer, the fifth residual shrinkage module, the sixth residual shrinkage module, the seventh residual shrinkage module, the eighth residual shrinkage module, the first LSTM layer, the second LSTM layer and the transformer layer are connected in sequence. The decoder includes 4 upsampling layers and 5 1D convolutional layers, and the first upsampling layer, the first 1D convolutional layer, the second upsampling layer, the second 1D convolutional layer, the third upsampling layer, the third 1D convolutional layer, the fourth upsampling layer, the fourth 1D convolutional layer and the fifth 1D convolutional layer are connected in sequence. A batch normalization layer is added after each convolutional layer, and the ReLU function is used as the activation function, as shown in Figure 3 shown.

[0071] Step 3: Deep neural network model training:

[0072] The deep neural network model was trained and the training data was augmented, including random shifting of the records with a probability of 99% and adding Gaussian white noise with zero mean and standard deviation of 0.01 to 0.15 times the maximum amplitude of the records to the records with a probability of 50%.

[0073] The loss function and mean absolute error function are used to ensure training accuracy. The batch size (Batchsize) of the network model is 64, the training round (Epoch) is 25, the initial learning rate is 0.01, and when the loss of the validation set remains unchanged for 3 consecutive epochs, the learning rate is reduced to 1 / 10 of the original, and a trained neural network model based on time-frequency analysis is obtained.

[0074] The loss function is shown in formula (2):

[0075]

[0076] Where: BCELoss is the binary cross entropy loss function; N is the number of sampling points in the probability sequence; y i is the probability value of the i-th position of the true probability sequence label; p i The probability value at the i-th position of the probability sequence predicted by the neural network.

[0077] Step 4: Seismic phase picking:

[0078] The three-component seismic records are input into the deep neural network model trained in step 4, and the P-wave and S-wave arrival time probability sequences are output. The network uses the Adam adaptive optimization function for back propagation, takes the binary cross entropy loss function as the loss function, and uses the mean absolute error function (MAE), error standard deviation (SD), precision, recall rate and F1 score as evaluation indicators to identify the location of the maximum value of the probability sequence, thereby realizing the P-wave and S-wave arrival time picking.

[0079] With the development of computer technology and deep learning, seismic data has gradually increased, the number of neural network layers has gradually increased, and the nonlinear fitting ability has become stronger and stronger. Various network models have been used for seismic phase picking, but these neural network methods usually face the problem of insufficient noise resistance and difficulty in dealing with low signal-to-noise ratio records. The present invention compares the current deep learning model, and the comparison results are as follows Figure 5 As shown in Table 1, the deep neural network model after training in this embodiment performs better. MAE and SD in the table represent the mean absolute error and standard deviation of phase picking, respectively.

[0080] Table 1 Comparison of prediction results between the present invention and the current deep learning model

[0081]

[0082] In the table, EQTransformer represents the deep learning model mentioned in the paper published in 2020 titled Earthquake transformer—anattentive deep-learning model for simultaneous earthquake detection and phasepicking.

[0083] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered as exemplary and non-restrictive in all respects.

Claims

1. A seismic phase picking method based on time-frequency analysis and deep learning, characterized in that: The following steps are involved: Step 1: Collect earthquake records and build a data set: Select known seismic motion data, pick up the P wave and S wave in the seismic motion data as the label value corresponding to the earthquake record, and obtain the seismic phase picking data set; Step 2: Establish a deep neural network model combined with the empirical mode decomposition module: The deep neural network consists of an encoder and a decoder. The first part of the encoder is the empirical mode decomposition module, which decomposes the ground motion into multiple intrinsic mode functions, each of which represents a component of different frequencies in the ground motion. The second part is composed of multiple residual shrinkage modules connected in series. The residual shrinkage module adds a soft threshold function on the basis of the traditional residual module to filter out background noise. The third part is two long short-term memory layers, which are used to extract temporal context information. The fourth part is the transformer module, which assigns weights to different positions in the sequence based on the attention mechanism to amplify key features. The decoder consists of multiple upsampling layers and convolutional layers to expand the feature map to the same length as the input; Step 3: Deep neural network model training: The deep neural network is trained, and the training accuracy is ensured by using the loss function and the mean absolute error function, so that the loss function gradually decreases, and the trained deep neural network model is obtained; Step 4: Seismic phase picking: Input the three-component seismic motion record into the deep neural network model trained in step 3, output the probability sequence corresponding to the P wave and S wave, find the position of the maximum value of the probability sequence, and thus realize the seismic phase picking.

2. The seismic phase picking method based on time-frequency analysis and deep learning according to claim 1, characterized in that: In step 2, the empirical mode decomposition of the encoder includes the following steps: 1) Identify all local maxima and local minima in the signal; 2) Interpolate local maxima to form an upper envelope, and interpolate local minima to form a lower envelope; 3) Calculate the mean of the upper and lower envelopes; 4) Subtract the mean from the signal to generate a new signal; if the new signal satisfies the conditions for being an IMF, it is considered an IMF; an IMF is a special type of signal that satisfies the number of extreme values ​​and the number of zero crossings, and the mean of the upper and lower envelopes must be close to zero; 5) Subtract this IMF from the original signal and repeat the process for the remaining signal to extract the next IMF.

3. The seismic phase picking method based on time-frequency analysis and deep learning according to claim 1, characterized in that: In step 2, 4 maximum pooling layers are added to the encoder to reduce the length of the input; the encoder includes 1 empirical mode decomposition module, 8 residual shrinkage modules, 4 maximum pooling layers, 2 LSTM layers and 1 transformer layer; the empirical mode decomposition module, the first residual shrinkage module, the first maximum pooling layer, the second residual shrinkage module, the second maximum pooling layer, the third residual shrinkage module, the third maximum pooling layer, the fourth residual shrinkage module, the fourth maximum pooling layer, the fifth residual shrinkage module, the sixth residual shrinkage module, the seventh residual shrinkage module, the eighth residual shrinkage module, the first LSTM layer, the second LSTM layer and the transformer layer are connected in sequence.

4. The seismic phase picking method based on time-frequency analysis and deep learning according to claim 1, characterized in that: In step 2, the decoder includes 4 upsampling layers and 5 1D convolution layers. The first upsampling layer, the first 1D convolution layer, the second upsampling layer, the second 1D convolution layer, the third upsampling layer, the third 1D convolution layer, the fourth upsampling layer, the fourth 1D convolution layer and the fifth 1D convolution layer are connected in sequence; a batch normalization layer is added after each convolution layer, and a linear rectification function is used as the activation function.

5. The seismic phase picking method based on time-frequency analysis and deep learning according to claim 1, characterized in that: The loss function in step 3 is the binary cross entropy loss function, as shown in formula (2): It can measure the difference between the probability distribution of the model output and the label value; BCELoss is the binary cross entropy loss function; N is the number of sampling points in the probability sequence; y i is the probability value of the i-th position of the true probability sequence label; p i The probability value at the i-th position of the probability sequence predicted by the neural network.

6. The seismic phase picking method based on time-frequency analysis and deep learning according to claim 1, characterized in that: In step three, the deep neural network model is trained and the training data is augmented, including recording random shifts and adding Gaussian white noise.

7. The seismic phase picking method based on time-frequency analysis and deep learning according to claim 6, characterized in that: The data enhancement method includes randomly shifting the records with a probability of 99%, and adding Gaussian white noise with zero mean and standard deviation of 0.01 to 0.15 times the maximum amplitude of the records to the records with a probability of 50%.

8. The seismic phase picking method based on time-frequency analysis and deep learning according to claim 1, characterized in that: The batch size of the deep neural network model after training in step 3 is 64, the training rounds are 25, the initial learning rate is 0.01, and when the loss of the validation set remains unchanged for three consecutive epochs, the learning rate is reduced to 1 / 10 of the original.

9. The seismic phase picking method based on time-frequency analysis and deep learning according to claim 1, characterized in that: In step 1, a standardization operation is performed on each channel of each seismic data to serve as the final network input data; The standardization operation is shown in formula (1): Where: x * is the data after standardization; μ is the mean of the data; σ is the standard deviation of the data; x is the data before standardization.

Images