CN113568043B - Three-stage seismic phase picking method based on deep convolutional neural network - Google Patents
Three-stage seismic phase picking method based on deep convolutional neural network Download PDFInfo
- Publication number
- CN113568043B CN113568043B CN202110837608.XA CN202110837608A CN113568043B CN 113568043 B CN113568043 B CN 113568043B CN 202110837608 A CN202110837608 A CN 202110837608A CN 113568043 B CN113568043 B CN 113568043B
- Authority
- CN
- China
- Prior art keywords
- wave
- stage
- time window
- network
- pickup
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
- G01V1/28—Processing seismic data, e.g. analysis, for interpretation, for correction
- G01V1/282—Application of seismic models, synthetic seismograms
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
- G01V1/28—Processing seismic data, e.g. analysis, for interpretation, for correction
- G01V1/30—Analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
A three-stage seismographic pickup method based on a deep convolutional neural network aims to solve the problem that the existing seismographic pickup method depends on initial time window selection and cannot be directly applied to continuous seismic record. The picking method comprises the following steps: establishing a data set of a first stage; establishing a time window detection network in the first stage, wherein the time window detection network comprises 9 convolutional layers and 1 full-connection layer; thirdly, training a time window detection network; fourthly, extracting a time window containing the P wave; fifthly, establishing a data set of the second stage; establishing a P wave seismic phase pickup network at a second stage; seventhly, training a P wave seismic phase pickup network; eighthly, predicting the arrival time of the P wave; ninthly, extracting a time window containing the S wave; establishing a data set of a third stage; eleven, establishing an S-wave seismic phase pickup network in the third stage; twelve, S wave seismic phase picking network training; thirteen, predicting the arrival time of the S wave. The invention adopts a three-stage mode, and a deep convolution network model is trained independently in each stage, so that the precision of seismic phase picking is improved.
Description
Technical Field
The invention belongs to the field of seismic engineering, and particularly relates to a three-stage seismic facies pickup method based on a deep convolutional neural network.
Background
The tectonic earthquake is the earthquake which occurs because the crust movement causes the fracture and the dislocation of the crust rock, and almost all destructive earthquakes in the world are tectonic earthquakes. When the earthquake directly causes casualties, various secondary disasters such as landslide, tsunami, debris flow and the like are also accompanied, and the casualties and the economic losses caused by destructive earthquakes are very disastrous every year. The earthquake emergency response can effectively reduce casualties caused by earthquake damage, and simple and quick earthquake source parameter estimation can provide scientific guidance for the earthquake emergency response in a short time after the earthquake. Seismic source parameter estimation needs seismic phase picking to acquire seismic phase information such as P waves and S waves.
Originally, the seismographic pickup was performed manually by experienced professionals, which took a long time although the accuracy was high, and the pickup accuracy was greatly affected by subjective factors of the pickup personnel. With the increase of earthquake events, the number of earthquake motion records is increased rapidly, and the manual seismographic pickup cannot meet the working requirements, so that students propose an automatic seismographic pickup algorithm based on feature calculation, such as STA/LTA (station-to-array/low-temperature array) and the like. However, the automatic seismographic pickup algorithm usually utilizes shallow features of seismic record to pick up seismographic, such features are easily interfered by noise, and the problems of insufficient robustness, relatively low recall rate and accuracy rate and the like of low signal to noise ratio seismic record seismographic pickup generally exist.
In recent years, with the development of computer technology and deep learning, a deep convolutional neural network obtains excellent performance in the field of feature extraction, so that learners extract deep features related to seismographic pickup in earthquake motion records by using the deep convolutional neural network, the robustness of low signal-to-noise ratio seismographic pickup of the earthquake motion records is greatly improved, and the accuracy and the recall rate of the seismographic pickup are greatly improved. However, such methods are often highly dependent on the selection of an initial time window, and the time window including the arrival time of the seismic phase needs to be preset during application, and cannot be directly used for continuous seismic motion recording.
Disclosure of Invention
The invention aims to solve the technical problems that the existing seismic phase pickup method depends on initial time window selection and cannot be directly applied to continuous seismic motion recording, and provides a novel three-stage seismic phase pickup method based on a deep convolutional neural network.
The three-stage seismographic pickup method based on the deep convolutional neural network is realized according to the following steps:
step one, collecting earthquake motion records and noise records with the same quantity, and establishing a data set of a first stage:
collecting seismic motion records and corresponding P wave arrival time label values and S wave arrival time label values, carrying out length interception on the collected seismic motion records, randomly extracting 3-7S before the P wave arrival time label values and the S wave arrival time label values as initial interception points, and respectively intercepting 10.24S seismic motion records (respectively called P wave arrival time windows and S wave arrival time windows) after the P wave initial interception points and the S wave initial interception points as input data; randomly intercepting 10.24s records (called noise time windows) from the collected noise records to ensure the length of input data to be consistent; the output data is a one-hot vector to obtain a seismic data set;
step two, establishing a time window detection network of the first stage:
the time window detection network comprises 9 convolutional layers and 1 full-link layer, and the full-link layer is activated by using a normalized exponential function (softmax); the convolution operation is carried out on the first 6 layers of convolution layers by using 64 convolution kernels with the size of 3 (one-dimensional convolution), and the convolution operation is carried out on the last 3 layers of convolution layers by using 128 convolution kernels with the size of 3 to obtain a time window detection network model;
step three, time window detection network training:
training a time window detection network model, performing back propagation through a cross entropy loss function (CategoricalCrossentpy) and an Adam self-adaptive optimization function, and obtaining the trained time window detection network model by taking the classification precision as an evaluation index;
step four, extracting a time window containing P waves:
continuously intercepting earthquake motion by using a sliding time window with the length of 10.24s from the moment when t is 0, classifying the input time window by using the time window detection network model trained in the step three, and judging whether the input time window contains P waves until the time window containing the P waves is found;
step five, establishing a data set of a second stage:
using the time window data identified as the P wave time window in the step four as the input data of the second stage; and (4) establishing output data of a second stage according to the arrival time tag value of the P wave collected in the first step, wherein the establishment principle is as follows: the output data has the same length as the input data, the numerical value corresponding to the time tag value of the P wave in the output data is set as 1, the numerical value of +/-10 points near the point linearly decreases to 0, and the numerical values of other points are set as 0;
step six, establishing a P wave seismic phase pickup network at a second stage:
the P wave seismic phase picking network comprises a plurality of down sampling layers and a plurality of up sampling layers, the up sampling step length is consistent with the pooling step length of the corresponding down sampling stage and the number of convolution kernels, and the corresponding up sampling layer and the down sampling layer are connected in a jumping mode with an attention mechanism, so that a P wave seismic phase picking network model is obtained;
step seven, training a P wave seismic phase pickup network:
training a P wave seismic phase picking network model, performing back propagation by adopting a Binary cross entropy loss function (Binary _ cross) and an Adam self-adaptive optimization function, and taking the precision of P wave seismic phase picking errors within 0.1s as an evaluation index to obtain the trained P wave seismic phase picking network model;
step eight, P wave arrival time prediction:
extracting accurate P wave arrival time t in the input time window by utilizing the P wave seismic phase pickup network model trained in the step sevenp；
Step nine, extracting a time window containing S waves:
from t to tpContinuously intercepting the seismic motion by using a sliding time window with the length of 10.24S when the time begins, classifying the input time window by using the time window detection network model trained in the step three, and judging whether the input time window contains S waves until the time window containing the S waves arrives;
step ten, establishing a data set of the third stage:
using the time window data identified as the S-wave time window in the step nine as input data of a third stage; and (4) establishing output data of the third stage according to the S wave arrival time tag value collected in the first step, wherein the establishment principle is as follows: the output data has the same length as the input data, the numerical value corresponding to the time-of-arrival tag value of the S wave in the output data is set as 1, the numerical value of +/-10 points near the point linearly decreases to 0, and the numerical values of other points are set as 0;
step eleven, establishing an S-wave seismic phase pickup network in the third stage:
the S-wave seismic phase pickup network has the same structure as the P-wave seismic phase pickup network and comprises a plurality of down-sampling layers and a plurality of up-sampling layers, the up-sampling step length is consistent with the pooling step length of the corresponding down-sampling stage and the number of convolution kernels, and the corresponding up-sampling layers and the down-sampling layers are connected in a jumping mode with an attention mechanism to obtain an S-wave seismic phase pickup network model;
step twelve, S wave seismic phase pickup network training:
training an S wave seismic phase picking network model, performing back propagation by adopting a Binary cross entropy loss function (Binary _ cross) and an Adam self-adaptive optimization function, and taking the precision of S wave seismic phase picking errors within 0.1S as an evaluation index to obtain the trained S wave seismic phase picking network model;
step thirteen, predicting the arrival time of S wave:
extracting accurate S-wave arrival time t in the input time window by using the S-wave seismic phase pickup network model trained in the step twelves。
The invention provides a three-stage seismographic pickup method based on a deep convolutional neural network, which comprises the following steps: the first stage is to construct a time window detection network and extract a time window containing P wave arrival time in the continuous earthquake motion record; constructing a P wave seismic phase pickup network at the second stage, and accurately judging the arrival time of the P wave in the time window extracted at the first stage; in the third stage, firstly, the earthquake motion record after the arrival of the P wave is intercepted by utilizing the time of arrival of the P wave judged in the second stage, then, the time window detection network constructed in the first stage is utilized to extract the time window containing the arrival time of the S wave from the earthquake motion record obtained after interception, and finally, the S wave earthquake phase pickup network is constructed to accurately judge the arrival time of the S wave.
The invention mainly selects more than 300000 earthquake motion records from American and European regions (STEAD databases), adopts a three-stage mode, trains a deep convolution network model in each stage independently, solves different tasks in sequence, and can improve the accuracy of earthquake phase picking and ensure the applicability of each deep neural network model.
Compared with the traditional automatic seismic facies picking algorithm and the early deep neural network model, the three-stage seismic facies picking method based on the deep convolutional neural network has the advantages of higher precision, better applicability and more convenience and rapidness in actual use.
Drawings
FIG. 1 is a flowchart of an overall framework of a three-stage facies seismic picking method based on a deep convolutional neural network according to an embodiment;
FIG. 2 is a diagram illustrating a network structure of a first stage time window detection network model in an embodiment, where Δ represents a max pooling layer,. smallcircle represents a linear rectification function,. lambada represents a global max pooling layer;
FIG. 3 is a network structure diagram of a facies seismic picking network model for extracting P-wave or S-wave facies at the second stage in the embodiment, in which a represents a residual module, b represents a bidirectional gated recurrent neural network, c represents an attention mechanism module, 1 represents Maxpool1D, 2 represents Up-Conv1D, and 3 represents full connection;
FIG. 4 is a test chart of P-wave seismic phase pickup using PhaseNet;
FIG. 5 is a test chart of S-wave seismic phase picking using PhaseNet;
FIG. 6 is a test chart of P-wave seismic phase picking using AR-AIC + STA/LTA;
FIG. 7 is a test chart of S-wave seismic phase picking using AR-AIC + STA/LTA;
FIG. 8 is a test chart of P-wave seismic facies pickup using an embodiment;
FIG. 9 is a test chart of S-wave seismic facies pickup using the example;
FIG. 10 is a graph comparing the error of P-seismogram phase picking for examples, PhaseNet and AR-AIC + STA/LTA, wherein 1 represents an embodiment of the invention, 2 represents AR-AIC + STA/LTA, and 3 represents PhaseNet;
FIG. 11 is a graph comparing the error of S-seismogram pickup of examples, PhaseNet and AR-AIC + STA/LTA, wherein 1 represents an embodiment of the present invention, 2 represents AR-AIC + STA/LTA, and 3 represents PhaseNet.
Detailed Description
The first embodiment is as follows: the three-stage seismographic pickup method based on the deep convolutional neural network is implemented according to the following steps:
step one, collecting earthquake motion records and noise records with the same quantity, and establishing a data set of a first stage:
collecting seismic motion records and corresponding P wave arrival time label values and S wave arrival time label values, carrying out length interception on the collected seismic motion records, randomly extracting 3-7S before the P wave arrival time label values and the S wave arrival time label values as initial interception points, and respectively intercepting 10.24S seismic motion records (respectively called P wave arrival time windows and S wave arrival time windows) after the P wave initial interception points and the S wave initial interception points as input data; randomly intercepting 10.24s records (called noise time windows) from the collected noise records to ensure the length of input data to be consistent; the output data is a one-hot vector to obtain a seismic data set;
step two, establishing a time window detection network in the first stage:
the time window detection network comprises 9 convolutional layers and 1 full-link layer, and the full-link layer is activated by using a normalized exponential function (softmax); carrying out convolution operation on the first 6 layers of convolution layers by using 64 convolution kernels with the size of 3, and carrying out convolution operation on the last 3 layers of convolution layers by using 128 convolution kernels with the size of 3 to obtain a time window detection network model;
step three, time window detection network training:
training a time window detection network model, performing back propagation through a cross entropy loss function (CategoricalCrossentpy) and an Adam self-adaptive optimization function, and obtaining the trained time window detection network model by taking the classified precision as an evaluation index;
step four, extracting a time window containing P waves:
continuously intercepting earthquake motion by using a sliding time window with the length of 10.24s from the moment when t is 0, classifying the input time window by using the time window detection network model trained in the step three, and judging whether the input time window contains P waves until the time window containing the P waves is found;
step five, establishing a data set of a second stage:
using the time window data identified as the P wave time window in the step four as the input data of the second stage; and (3) establishing output data of a second stage according to the P wave arrival time tag value collected in the first step, wherein the establishment principle is as follows: the output data has the same length as the input data, the numerical value corresponding to the time tag value of the P wave in the output data is set as 1, the numerical value of +/-10 points near the point linearly decreases to 0, and the numerical values of other points are set as 0;
step six, establishing a P wave seismic phase pickup network at the second stage:
the P wave seismic phase picking network comprises a plurality of down sampling layers and a plurality of up sampling layers, the up sampling step length is consistent with the pooling step length of the corresponding down sampling stage and the number of convolution kernels, and the corresponding up sampling layer and the down sampling layer are connected in a jumping mode with an attention mechanism, so that a P wave seismic phase picking network model is obtained;
step seven, P wave seismic phase pickup network training:
training a P wave seismic phase picking network model, performing back propagation by adopting a Binary cross entropy loss function (Binary _ cross) and an Adam self-adaptive optimization function, and taking the precision of P wave seismic phase picking errors within 0.1s as an evaluation index to obtain the trained P wave seismic phase picking network model;
step eight, P wave arrival time prediction:
extracting accurate P wave arrival time t in the input time window by utilizing the P wave seismic phase pickup network model trained in the step sevenp；
Step nine, extracting a time window containing S waves:
from t to tpContinuously intercepting the seismic motion by using a sliding time window with the length of 10.24S when the time begins, classifying the input time window by using the time window detection network model trained in the step three, and judging whether the input time window contains S waves until the time window containing the S waves arrives;
step ten, establishing a data set of the third stage:
using the time window data identified as the S-wave time window in the step nine as input data of a third stage; and (3) establishing output data of a third stage according to the S wave arrival time tag value collected in the first stage, wherein the establishment principle is as follows: the output data has the same length as the input data, the numerical value corresponding to the time-of-arrival tag value of the S wave in the output data is set as 1, the numerical value of +/-10 points near the point linearly decreases to 0, and the numerical values of other points are set as 0;
step eleven, establishing an S-wave seismic phase pickup network in the third stage:
the S-wave seismic phase pickup network has the same structure as the P-wave seismic phase pickup network and comprises a plurality of down-sampling layers and a plurality of up-sampling layers, the up-sampling step length is consistent with the pooling step length of the corresponding down-sampling stage and the number of convolution kernels, and the corresponding up-sampling layers and the down-sampling layers are connected in a jumping mode with an attention mechanism to obtain an S-wave seismic phase pickup network model;
step twelve, S wave seismic phase pickup network training:
training an S wave seismic phase picking network model, performing back propagation by adopting a Binary cross entropy loss function (Binary _ cross) and an Adam self-adaptive optimization function, and taking the precision of S wave seismic phase picking errors within 0.1S as an evaluation index to obtain the trained S wave seismic phase picking network model;
step thirteen, predicting the arrival time of S wave:
extracting accurate S wave arrival time t in the input time window by using the S wave seismic phase pickup network model trained in the step twelves。
The second embodiment is as follows: this embodiment differs from the embodiment in that the seismic and noise recordings in step one are selected from the STEAD database.
The third concrete implementation mode: the difference between the first embodiment and the second embodiment is that in the first step, input data are located between-1 and 1 through a standardization method for each seismic motion record, and output data are not subjected to standardization processing.
The fourth concrete implementation mode: the difference between this embodiment and one of the first to third embodiments is that, in addition to the first two convolutional layers in the two-step time window detection network being closely connected, other convolutional layers all include a batch normalization layer (batch normalization layer), a linear rectification function (ReLU), and a maximum pooling layer (maxpooling layer), and the last convolutional layer and the full-connection layer include a linear rectification function (ReLU) and a global maximum pooling layer (global maxpooling layer).
In the time window detection network of the embodiment, the 9 convolutional layers are sequentially a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, an eighth convolutional layer and a ninth convolutional layer, the first convolutional layer and the second convolutional layer are closely connected, adjacent convolutional layers in the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer, the seventh convolutional layer and the eighth convolutional layer all comprise a batch standard layer, a linear rectification function and a maximum pooling layer, and the ninth convolutional layer and the full-connection layer comprise a linear rectification function and a global maximum pooling layer.
The fifth concrete implementation mode: the difference between this embodiment and one of the first to the fourth embodiments is that the expression of the cross entropy loss function described in step three is as follows:
wherein: p is a radical ofic-probability that sample i belongs to class c; y isic-sign function (0 or 1), taking 1 if sample i belongs to class c, otherwise taking 0; m-number of classes; n-the amount of data.
The sixth specific implementation mode: the difference between the present embodiment and one of the first to fifth embodiments is that the earthquake motion is continuously intercepted in step four, and the sliding distance of the time window each time is 5.12 s.
The seventh embodiment: the difference between the present embodiment and one of the first to sixth embodiments is that in the sixth step, after the down-sampling is finished, a 4-layer residual error module and a 3-layer bidirectional gated recurrent neural network (GRU) are added to further pick up features related to P-wave vibration.
The specific implementation mode is eight: the difference between this embodiment and one of the first to seventh embodiments is that the expression of the Binary cross entropy loss function (Binary _ cross) described in step seven and step twelve is as follows:
wherein: p is a radical ofi-probability of a sample prediction being positive; y isi-the label value of sample i, positive class 1, negative class 0; n-number of dataAmount of the compound (A).
The specific implementation method nine: the difference between this embodiment and the first to eighth embodiments is that the algorithm of Adam adaptive optimization function is as follows:
(1) calculating a first moment estimate and a second moment estimate of the gradient by the following formula:
mt＝β1*mt-1+(1-β1)*gt，vt＝β2*vt-1+(1-β2)*gt 2；
in the formula, gtIs a gradient in which mtIs the mean value of the gradient at time t, vtIs the non-central variance value, m, at time t of the gradientt-1Is the mean value at time t-1 of the gradient, Vt-1The exponential decay rate beta of the moment estimate, which is the non-central variance value at time t-1 of the gradient1And beta2Within the interval [0,1 ], beta1Take 0.9, beta2Taking 0.999;
(2) correcting the first order moment estimate and the second order moment estimate by calculating the formula:
(3) the final formula for parameter update is:
in the formula, thetatFor updated parameters, η is the learning rate, ε is a small constant for numerical stability, ε is taken to be 10-8。
The detailed implementation mode is ten: the present embodiment is different from the first to ninth embodiments in that the batch size (batch size) of the time window detection network model after the training in the third step is 256, the training round (Epoch) is 7, and the learning rate is 0.0005.
The concrete implementation mode eleven: the present embodiment is different from the first to tenth embodiments in that the batch size (batch size) of the P-wave seismic phase picking network model after the training in the seventh step is 300, the training round (Epoch) is 30, and the learning rate is 0.0001.
The specific implementation mode twelve: the present embodiment is different from the first to eleventh embodiments in that the batch size (batch size) of the S-wave seismic phase picking network model after the training in the twelfth step is 300, the training round (Epoch) is 30, and the learning rate is 0.0001.
Example (b): the three-stage epicenter phase picking method based on the deep convolutional neural network is implemented according to the following steps:
step one, collecting earthquake motion records, and establishing a data set of a first stage:
selecting about 15 thousands of earthquake motion records and 15 thousands of noise records from an STEAD database, selecting corresponding P wave arrival time tag values and S wave arrival time tag values, randomly extracting the P wave arrival time tag values and 3-7S before the S wave arrival time tag values as initial interception points, and respectively intercepting 10.24S earthquake motion records behind the initial interception points, which are respectively called as P wave arrival time windows and S wave arrival time windows, as input data; randomly intercepting 10.24s records (called noise time windows) of the collected noise records to ensure that the lengths of input data are consistent; the output data is a one-hot vector, and the label values corresponding to the P wave arrival time window, the S wave arrival time window and the noise time window are (1,0,0), (0,1,0) and (0,0,1) respectively; respectively enabling input data to be located between-1 and 1 through a standardization method for each seismic motion record, and enabling output data not to be subjected to standardization processing to obtain a data set of a first stage;
the formula of the normalization method is as follows:
wherein: x (t)': recording the standardized earthquake motion; x is the number ofmean: recording the average value of each data point by single earthquake motion; x is the number ofmin: recording the standard deviation of each data point by single earthquake motion; x (t): recording the earthquake motion before standardization;
step two, dividing the data set of the first stage;
randomly dividing the data set of the first stage into a training set, a verification set and a test set, wherein the proportion of the training set, the verification set and the test set is 8: 1: 1;
step three, establishing a time window detection network in the first stage:
establishing a time window detection network consisting of 9 convolutional layers and 1 fully-connected Layer, wherein except that the former two convolutional layers are closely connected, other convolutional layers comprise Batch Normalization layers (Batch), linear rectification functions (ReLU) and maximum Pooling layers (Max Pooling layers), the last convolutional Layer and the fully-connected Layer comprise the linear rectification functions (ReLU) and the Global maximum Pooling layers (Global Max Pooling layers), and the fully-connected Layer is activated by using the Normalization exponential function (softmax); the convolution operation is carried out on the first 6 layers of convolution layers by using 64 convolution kernels with the size of 3, and the convolution operation is carried out on the last 3 layers of convolution layers by using 128 convolution kernels with the size of 3; performing back propagation by adopting a cross entropy loss function (category cross) and an Adam adaptive optimization function, and obtaining a time window detection network model by taking the classified precision as an evaluation index;
step four, time window detection network training:
training the time window detection network model, ensuring the training precision through a cross entropy loss function and the classification precision, enabling an attenuation curve to smoothly descend, and obtaining the trained time window detection network model, wherein the batch size (Batchsize) of the trained deep neural network model is 256, the training round (Epoch) is 7, and the learning rate is 0.0005;
step five, extracting a time window containing P waves:
continuously intercepting earthquake motion by using a sliding time window with the length of 10.24s from the moment t being 0, wherein the sliding distance of the time window every time is 5.12s, classifying the input time window by using the time window detection network model trained in the step three, and judging whether the input time window contains P waves until the time window containing the P waves is found;
step six, establishing a data set of a second stage:
using the time window data identified as the P wave time window in the step four as the input data of the second stage; and (3) establishing output data of a second stage according to the P wave arrival time tag value collected in the first step, wherein the establishment principle is as follows: the output data has the same length as the input data, the numerical value corresponding to the time tag value of the P wave in the output data is set as 1, the numerical value of +/-10 points near the point linearly decreases to 0, and the numerical values of other points are set as 0;
step seven, establishing a P wave seismic phase pickup network at the second stage:
the P wave seismic phase picking network comprises a plurality of down sampling layers and up sampling layers, the step length of each up sampling layer is consistent with the pooling step length of the corresponding down sampling stage and the number of convolution kernels, jump connection with a attention mechanism is adopted between the corresponding up sampling layer and the corresponding down sampling layer, and after down sampling is finished, 4 layers of residual error modules and 3 layers of bidirectional gated cyclic neural networks (GRUs) are added to further pick up characteristics related to P wave seismic phase; carrying out back propagation by adopting a Binary cross entropy loss function (Binary _ cross) and an Adam self-adaptive optimization function, and obtaining a P wave seismic phase pickup network model by taking the precision of P wave seismic phase pickup errors within 0.1s as an evaluation index;
step eight, training a P wave seismic phase pickup network:
training the P wave seismic phase pickup network model, ensuring the training precision through a binary cross entropy loss function and the classification precision, enabling an attenuation curve to smoothly descend to obtain the trained P wave seismic phase pickup network model, wherein the batch size (Batchsize) of the trained deep neural network model is 300, the training round (Epoch) is 30, and the learning rate is 0.0001;
step nine, P wave arrival time prediction:
extracting accurate P wave arrival time t in the input time window by utilizing the P wave seismic phase pickup network model trained in the step sevenp；
Step ten, extracting a time window containing S waves:
from t to tpAnd (4) intercepting the earthquake motion continuously by using a sliding time window with the length of 10.24S at the beginning of the moment, wherein the sliding distance of each time of the time window is 5.12S, classifying the input time window by using the time window detection network model trained in the step three, and judging whether the input time window contains S waves until the time when the input time window contains the S waves is found outA time window of time;
step eleven, establishing a data set of the third stage:
using the time window data identified as the S-wave time window in the step nine as input data of a third stage; and (3) establishing output data of a second stage according to the S wave arrival time tag value collected in the first step, wherein the establishment principle is as follows: the output data has the same length as the input data, the numerical value corresponding to the time-of-arrival tag value of the S wave in the output data is set as 1, the numerical value of +/-10 points near the point linearly decreases to 0, and the numerical values of other points are set as 0;
step twelve, establishing an S-wave seismic phase pickup network in the third stage:
the S-wave seismographic pickup network and the P-wave seismographic pickup network have the same structure and comprise a plurality of down-sampling layers and up-sampling layers, the up-sampling step length is consistent with the pooling step length and the convolution kernel number of the corresponding down-sampling stage, the corresponding up-sampling layers and the down-sampling layers are connected in a jumping mode with an attention mechanism, and after the down-sampling is finished, 4 layers of residual error modules and 3 layers of bidirectional gated recurrent neural networks (GRUs) are added to further pick up the characteristics related to the S-wave seismogram; carrying out back propagation by adopting a Binary cross entropy loss function (Binary _ cross) and an Adam self-adaptive optimization function, and obtaining an S wave seismic phase pickup network model by taking the precision of S wave seismic phase pickup errors within 0.1S as an evaluation index;
step thirteen, S wave seismic phase pickup network training:
training the S-wave seismic phase pickup network model, ensuring the training precision through a binary cross entropy loss function and the classification precision, and enabling an attenuation curve to be smoothly reduced to obtain the trained S-wave seismic phase pickup network model, wherein the batch size (Batchsize) of the trained deep neural network model is 300, the training round (Epoch) is 30, and the learning rate is 0.0001;
step fourteen, predicting the arrival time of S waves:
extracting accurate S wave arrival time t in the input time window by using the S wave seismic phase pickup network model trained in the step twelvesThus, the three-stage epicenter phase picking method based on the deep convolutional neural network is completed.
The seismographic pickup network combines U-Net + + and Attention U-Net, a deep supervision mechanism is removed on the basis of a U-Net + + network structure, a self-Attention mechanism in the Attention U-Net network is added, and 4 residual modules containing batch standardization and a 3-layer bidirectional gating cyclic neural network (GRU) are added to a bottleneck. The seismic phase picking network comprises 4 up-sampling layers and 4 down-sampling layers, the number of convolution kernels of the 4 up-sampling layers is 8,16,32 and 64 in sequence, the sizes of the convolution kernels are 11,9,7 and 5 in sequence, the pooling step length is kept to be 2, and the parameters of the 4 down-sampling layers are kept consistent with the parameters of the corresponding up-sampling layers. And a fusion module is also arranged between the connected down-sampling layers, the output of the upper layer and the up-sampling output of the lower layer are fused, a new round of information fusion is carried out by the fusion module between the fusion modules, and finally the fused information is output to the up-sampling layer through the self-attention module.
The traditional seismographic pickup method relies on an automatic seismographic pickup algorithm based on AR-AIC and STA/LTA feature calculation, but the automatic seismographic pickup algorithm usually utilizes shallow features of seismic motion records to pick up seismographic phases, the features are easily interfered by noise, and the problems of insufficient robustness, relatively low recall rate and accuracy rate and the like of seismographic motion record seismographic pickup with low signal-to-noise ratio generally exist. With the development of computer technology, the amplification of data sets, the development of neural networks towards deep layers and the development of neurons towards high orders, and the method has the advantages that a learner proposes PhaseNET to perform feature extraction, so that great progress is achieved, but the method still has the defect that the PhaseNET is highly dependent on the selection of an initial time window and cannot be directly used for continuous earthquake motion recording. Therefore, the present invention selects AR-AIC + STA/LTA and PhaseNET for comparison.
The results are shown in Table 1 and Table 2, and the comparative graphs are shown in FIGS. 4 to 11. The three-stage epicenter picking method provided by the embodiment performs best through comparison of the epicenter picking precision, recall rate, calculation comparison of F1 scores and observation comparison of picking error distribution.
TABLE 1 comparison of the present invention with the AIC + STA/LTA and PhaseNET methods P wave seismic phase pickup results
TABLE 2 comparison of S-wave seismic phase picking results of the present invention and AIC + STA/LTA and PhaseNET methods
Claims (10)
1. The three-stage seismograph pickup method based on the deep convolutional neural network is characterized by being realized according to the following steps:
step one, collecting earthquake motion records and noise records with the same quantity, and establishing a data set of a first stage:
collecting seismic motion records and corresponding P wave arrival time label values and S wave arrival time label values, carrying out length interception on the collected seismic motion records, randomly extracting the P wave arrival time label values and 3-7S before the S wave arrival time label values as initial interception points, respectively intercepting the seismic motion records 10.24S after the P wave initial interception points and the S wave initial interception points, and taking the seismic motion records as input data; randomly intercepting 10.24s records of the collected noise records, and ensuring the lengths of input data to be consistent; the output data is a one-hot vector to obtain a seismic data set;
step two, establishing a time window detection network in the first stage:
the time window detection network comprises 9 convolutional layers and 1 full-link layer, and the full-link layer is activated by using a normalized exponential function; the convolution operation is carried out on the first 6 layers of convolution layers by using 64 convolution kernels with the size of 3, and the convolution operation is carried out on the last 3 layers of convolution layers by using 128 convolution kernels with the size of 3 to obtain a time window detection network model;
step three, time window detection network training:
training a time window detection network model, performing back propagation through a cross entropy loss function and an Adam self-adaptive optimization function, and taking the classification precision as an evaluation index to obtain the trained time window detection network model;
step four, extracting a time window containing P waves:
continuously intercepting earthquake motion by using a sliding time window with the length of 10.24s from the moment when t is 0, classifying the input time window by using the time window detection network model trained in the step three, and judging whether the input time window contains P waves until the time window containing the P waves is found;
step five, establishing a data set of a second stage:
using the time window data identified as the P wave time window in the step four as the input data of the second stage; and (4) establishing output data of a second stage according to the arrival time tag value of the P wave collected in the first step, wherein the establishment principle is as follows: the output data has the same length as the input data, the numerical value corresponding to the time tag value of the P wave in the output data is set as 1, the numerical value of +/-10 points near the point linearly decreases to 0, and the numerical values of other points are set as 0;
step six, establishing a P wave seismic phase pickup network at the second stage:
the P wave seismic phase picking network comprises a plurality of down sampling layers and a plurality of up sampling layers, the up sampling step length is consistent with the pooling step length of the corresponding down sampling stage and the number of convolution kernels, and the corresponding up sampling layer and the down sampling layer are connected in a jumping mode with an attention mechanism, so that a P wave seismic phase picking network model is obtained;
step seven, P wave seismic phase pickup network training:
training a P wave seismic phase pickup network model, performing back propagation by adopting a binary cross entropy loss function and an Adam self-adaptive optimization function, and obtaining the trained P wave seismic phase pickup network model by taking the precision of P wave seismic phase pickup errors within 0.1s as an evaluation index;
step eight, P wave arrival time prediction:
extracting accurate P wave arrival time t in the input time window by utilizing the P wave seismic phase pickup network model trained in the step sevenp；
Step nine, extracting a time window containing S waves:
from t to tpAnd (4) intercepting the earthquake motion continuously by using a sliding time window with the length of 10.24s from the beginning of the moment, classifying the input time window by using the time window detection network model trained in the step three, and judging the input time windowWhether the arrival time of the S wave is contained or not is judged until a time window containing the arrival time of the S wave is found;
step ten, establishing a data set of the third stage:
using the time window data identified as the S-wave time window in the step nine as input data of a third stage; and (3) establishing output data of a third stage according to the S wave arrival time tag value collected in the first stage, wherein the establishment principle is as follows: the output data has the same length as the input data, the numerical value corresponding to the time-of-arrival tag value of the S wave in the output data is set as 1, the numerical value of +/-10 points near the point linearly decreases to 0, and the numerical values of other points are set as 0;
step eleven, establishing an S-wave seismic phase pickup network in the third stage:
the S-wave seismic phase picking network has the same structure as the P-wave seismic phase picking network and comprises a plurality of down-sampling layers and a plurality of up-sampling layers, the up-sampling step length is consistent with the pooling step length of the corresponding down-sampling stage and the number of convolution kernels, and the corresponding up-sampling layers and the down-sampling layers are connected in a jumping mode with an attention mechanism to obtain an S-wave seismic phase picking network model;
step twelve, S wave seismic phase pickup network training:
training an S-wave seismic phase pickup network model, performing back propagation by adopting a binary cross entropy loss function and an Adam self-adaptive optimization function, and obtaining the trained S-wave seismic phase pickup network model by taking the precision of S-wave seismic phase pickup errors within 0.1S as an evaluation index;
step thirteen, predicting the arrival time of S wave:
extracting accurate S wave arrival time t in the input time window by using the S wave seismic phase pickup network model trained in the step twelves。
2. The three-stage seismographic pick-up method based on the deep convolutional neural network as claimed in claim 1, wherein the seismic motion recording and the noise recording in the first step are selected from STEAD database.
3. The three-stage seismographic pickup method based on the deep convolutional neural network as claimed in claim 1, wherein in the step one, input data is located between-1 and 1 through a standardization method for each seismic motion record, and output data is not subjected to standardization processing.
4. The three-stage seismographic pickup method based on the deep convolutional neural network as claimed in claim 1, wherein the two time window detection networks in the step two are characterized in that except that the first two convolutional layers are closely connected, other convolutional layers comprise a batch normalization layer, a linear rectification function and a maximum pooling layer, and the last convolutional layer and the full-connection layer comprise a linear rectification function and a global maximum pooling layer.
5. The three-stage epicenter picking method based on deep convolutional neural network as claimed in claim 1, wherein the expression of the cross entropy loss function in step three is as follows:
wherein: p is a radical ofic-probability that sample i belongs to class c; y isic-sign function, taking 1 if sample i belongs to class c, otherwise taking 0; m-number of classes; n-the amount of data.
6. The three-stage seismographic pick-up method based on the deep convolutional neural network as claimed in claim 1, wherein the seismic motion is continuously intercepted in the fourth step, and the sliding distance of the time window is 5.12s each time.
7. The three-stage seismograph pickup method based on the deep convolutional neural network as claimed in claim 1, wherein in step six, 4 layers of residual error modules and 3 layers of bidirectional gated cyclic neural network are added after the down-sampling is finished to further pick up the features related to the P wave seismograph.
8. The three-stage epicenter picking method based on the deep convolutional neural network as claimed in claim 1, wherein the expression of the binary cross entropy loss function in the seventh step and the twelfth step is as follows:
wherein: p is a radical ofi-probability of a sample prediction being positive; y isi-the label value of sample i, positive class 1, negative class 0; n-the amount of data.
9. The three-stage epicenter picking method based on the deep convolutional neural network as claimed in claim 1, wherein the batch size of the time window detection network model after the training in the step three is 256, the training round is 7, and the learning rate is 0.0005.
10. The three-stage epicenter picking method based on the deep convolutional neural network as claimed in claim 1, wherein the batch size of the P wave epicenter picking network model after seven training is 300, the training round is 30, and the learning rate is 0.0001.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110837608.XA CN113568043B (en) | 2021-07-23 | 2021-07-23 | Three-stage seismic phase picking method based on deep convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110837608.XA CN113568043B (en) | 2021-07-23 | 2021-07-23 | Three-stage seismic phase picking method based on deep convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113568043A CN113568043A (en) | 2021-10-29 |
| CN113568043B true CN113568043B (en) | 2022-05-24 |
Family
ID=78166758
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110837608.XA Active CN113568043B (en) | 2021-07-23 | 2021-07-23 | Three-stage seismic phase picking method based on deep convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113568043B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113468466B (en) * | 2021-07-23 | 2022-04-15 | 哈尔滨工业大学 | One-dimensional wave equation solving method based on neural network |
| CN114509811B (en) * | 2022-01-28 | 2022-12-02 | 中国地震局地球物理研究所 | Single station rear azimuth estimation method and device based on deep learning |
| CN115099523B (en) * | 2022-07-20 | 2023-10-24 | 哈尔滨工业大学 | Earthquake motion parameter prediction method based on artificial intelligence and earthquake motion experience model mixing |
| CN115310488A (en) * | 2022-08-16 | 2022-11-08 | 哈尔滨工业大学 | Seismic oscillation recording filtering method based on generating type antagonistic neural network |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6418378B1 (en) * | 2000-06-26 | 2002-07-09 | Westerngeco, L.L.C. | Neural net prediction of seismic streamer shape |
| CN103995290B (en) * | 2014-06-03 | 2016-11-02 | 山东科技大学 | A kind of microseism P ripple seismic phase first arrival automatic pick method |
| CN109212597B (en) * | 2018-10-16 | 2020-04-17 | 中国矿业大学(北京) | Micro seismic source automatic positioning method based on deep belief network and scanning superposition |
| WO2020122915A1 (en) * | 2018-12-13 | 2020-06-18 | Halliburton Energy Services, Inc. | Seismic multiples mitigation in borehole seismic data using inversion |
| CN110032975B (en) * | 2019-04-15 | 2021-09-07 | 禁核试北京国家数据中心 | Seismic facies picking method |
| CN110501742A (en) * | 2019-08-22 | 2019-11-26 | 东北大学 | A method of seismic events are distinguished using Boosting Ensemble Learning Algorithms |
| CN110516740A (en) * | 2019-08-28 | 2019-11-29 | 电子科技大学 | A kind of fault recognizing method based on Unet++ convolutional neural networks |
| CN111580151B (en) * | 2020-05-13 | 2021-04-20 | 浙江大学 | SSNet model-based earthquake event time-of-arrival identification method |
-
2021
- 2021-07-23 CN CN202110837608.XA patent/CN113568043B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN113568043A (en) | 2021-10-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113568043B (en) | Three-stage seismic phase picking method based on deep convolutional neural network | |
| Mousavi et al. | STanford EArthquake Dataset (STEAD): A global data set of seismic signals for AI | |
| CN108805269B (en) | Method for picking seismic facies arrival time based on LSTM recurrent neural network | |
| Zhu et al. | Deep learning for seismic phase detection and picking in the aftershock zone of 2008 Mw7. 9 Wenchuan Earthquake | |
| CN110032975B (en) | Seismic facies picking method | |
| McBrearty et al. | Pairwise association of seismic arrivals with convolutional neural networks | |
| Guo et al. | AEnet: Automatic picking of P-wave first arrivals using deep learning | |
| CN115758212A (en) | Mechanical equipment fault diagnosis method based on parallel network and transfer learning | |
| CN109613610B (en) | Automatic picking method for microseism signal arrival time difference | |
| CN111580151B (en) | SSNet model-based earthquake event time-of-arrival identification method | |
| CN111458746B (en) | Tunnel microseismic waveform arrival time picking method based on U-Net neural network | |
| CN109063687A (en) | A kind of microseism P wave recognition methods and system based on depth convolutional neural networks | |
| CN111027452A (en) | Microseismic signal arrival time and seismic phase identification method and system based on deep neural network | |
| CN114387520A (en) | Precision detection method and system for intensive plums picked by robot | |
| CN115327616B (en) | Automatic positioning method for mine microseism focus driven by massive data | |
| CN114509811B (en) | Single station rear azimuth estimation method and device based on deep learning | |
| CN111983676A (en) | Earthquake monitoring method and device based on deep learning | |
| Rojas et al. | Artificial neural networks as emerging tools for earthquake detection | |
| CN113009553A (en) | Interpretable seismic motion parameter probability density distribution prediction method based on NGboost and SHAP values | |
| Saad et al. | EQCCT: A production-ready EarthQuake detection and phase picking method using the Compact Convolutional Transformer | |
| CN115308793A (en) | Intelligent prediction method for shallow surface natural gas hydrate enrichment area | |
| CN105787045A (en) | Precision enhancing method for visual media semantic indexing | |
| CN111505706A (en) | Microseism P wave first arrival pickup method and device based on depth T-Net network | |
| León et al. | Small-layered Feed-Forward and Convolutional neural networks for efficient P wave earthquake detection | |
| Luo et al. | Detecting and Locating Aftershocks for the 2020 M w 6.5 Stanley, Idaho, Earthquake Using Convolutional Neural Networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |