CN114624768B - Method and device for training earthquake first arrival pickup model - Google Patents
Method and device for training earthquake first arrival pickup model Download PDFInfo
- Publication number
- CN114624768B CN114624768B CN202011471662.9A CN202011471662A CN114624768B CN 114624768 B CN114624768 B CN 114624768B CN 202011471662 A CN202011471662 A CN 202011471662A CN 114624768 B CN114624768 B CN 114624768B
- Authority
- CN
- China
- Prior art keywords
- data
- model
- training
- arrival
- type
- 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
- 238000012549 training Methods 0.000 title claims abstract description 92
- 238000000034 method Methods 0.000 title claims abstract description 55
- 230000000694 effects Effects 0.000 claims abstract description 16
- 238000007781 pre-processing Methods 0.000 claims abstract description 13
- 230000001360 synchronised effect Effects 0.000 claims abstract description 11
- 238000005070 sampling Methods 0.000 claims description 13
- 238000004364 calculation method Methods 0.000 claims description 12
- 238000012545 processing Methods 0.000 claims description 12
- 230000008447 perception Effects 0.000 abstract description 2
- 238000003672 processing method Methods 0.000 description 11
- 238000010586 diagram Methods 0.000 description 6
- 230000009466 transformation Effects 0.000 description 6
- 238000013501 data transformation Methods 0.000 description 5
- 238000000605 extraction Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000010606 normalization Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000008602 contraction Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000010339 dilation Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
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. for interpretation or for event detection
- G01V1/30—Analysis
- G01V1/307—Analysis for determining seismic attributes, e.g. amplitude, instantaneous phase or frequency, reflection strength or polarity
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V2210/00—Details of seismic processing or analysis
- G01V2210/60—Analysis
- G01V2210/63—Seismic attributes, e.g. amplitude, polarity, instant phase
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
- Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping
Landscapes
- Engineering & Computer Science (AREA)
- Remote Sensing (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Acoustics & Sound (AREA)
- Environmental & Geological Engineering (AREA)
- Geology (AREA)
- General Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Geophysics (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention provides a method and a device for training a first-arrival earthquake pickup model, wherein the method for training the first-arrival earthquake pickup model comprises the following steps: step 1, preprocessing to obtain a plurality of groups of data, wherein the plurality of groups of data comprise first-class data and second-class data; step 2, selecting a training arrangement mode, wherein the training comprises a first type of training and a second type of training; step 3, generating data arrangement of each round of training; and 4, training the model until convergence. The method and the device for training the first arrival pickup model of the earthquake can integrate more data features in one model, and the perception of the synchronous arrangement and the asynchronous arrangement of the cross training arrangement on the first arrival is more focused and accurate, so that the effect of 1+1>2 is achieved.
Description
Technical Field
The invention relates to the field of petroleum geophysical exploration data processing, in particular to a method and a device for training a first-arrival earthquake pickup model.
Background
In seismic exploration, a seismic wave excited from a bubble point and first reaching a detector is called a first arrival wave, and this arrival time is called a first arrival time of the seismic, abbreviated as a first arrival of the seismic. The first-arrival seismic pick-up is directly related to the quality of the static correction and near-surface velocity tomographic inversion. Especially for the detection areas of complex earth surfaces such as deserts, mountain lands and the like, the accuracy of first arrival pickup is directly related to success and failure of earthquake processing projects.
The most classical method for picking up the first arrival of an earthquake is a long and short time window average energy ratio method (STA/LTA), and is characterized in that the first arrival position is calculated by taking a single channel as a unit and by channel. And parameters such as the length of the time window, the threshold value of the energy ratio and the like are manually defined. The method and a plurality of improved methods thereof have better picking effect in areas with high signal to noise ratio, but have generally unsatisfactory application effect in areas with medium and low signal to noise ratio.
On the other hand, along with the application of artificial intelligence technology, more and more research and analysis section picture segmentation, energy distribution, target feature detection and other application methods are adopted in first arrival pickup. However, the known researches are based on a certain kind of data, and after modeling, the fitting model parameters are repeatedly and iteratively trained.
Whether manually set or automatically learned by training model parameters, the feature analysis is often single and cannot adapt to various, abnormal and medium-low signal-to-noise ratio production data. Therefore, the key problem to be solved is to organize and arrange multiple training until convergence under one model frame aiming at different seismic data processing methods or first arrival tag codes.
In application number: in CN201711381206.3, a method and apparatus for picking up the first arrival of seismic waves are disclosed. The method comprises the following steps: determining a first arrival set of the plurality of seismic trace data meeting specified conditions; fitting the first arrival speeds respectively corresponding to the first arrivals and the offsets respectively corresponding to the plurality of seismic channel data in the first arrival sets, determining target time window ranges respectively corresponding to the plurality of seismic channel data according to fitting results, determining a second first arrival set of the plurality of seismic channel data, and iteratively determining the target time window ranges and the second first arrival set until the absolute value of the first arrival difference of adjacent seismic channel data in the second first arrival set is smaller than a preset time difference threshold; determining final time window ranges corresponding to the plurality of seismic channel data respectively, and determining target first-arrival sets of the plurality of seismic channel data so that the target first-arrival sets meet specified conditions, and the absolute value of the difference between the first arrivals of adjacent seismic channel data is smaller than a preset time difference threshold. The patented claim of this patent application may be regarded as a synchronous arrangement in which each iteration uses the same data encoding and decision conditions until the model convergence meets the set threshold. It is well known that first arrival picking is affected by superposition of different geological types, excitation modes and noise types, and a single model training mode can restrict accuracy of time window estimation, offset fitting and target identification. While multitasking requires not only flexible training organization and orchestration, but also multiple types of data processing and encoding, which are not considered by this patent application.
In application number: in CN201710326043.2, a first arrival picking method and device based on deep learning are related, and the method includes: acquiring seismic data; respectively extracting a seismic data training set and a seismic data testing set from the seismic data; training a specific deep learning initial model based on the seismic data training set to obtain a first arrival pickup model; and carrying out first-arrival picking on the data in the seismic data test set according to the first-arrival picking model to obtain a first-arrival picking result. The patency of this patent application is directed to data set partitioning and usage patterns, which are common knowledge of the art about artificial intelligence, and are not novel or inventive.
In application number: in the chinese patent application CN201910193634.6, a first arrival picking method and apparatus are related, the method includes: extracting characteristic attributes of the seismic data test samples, inputting the characteristic attributes of the seismic data test samples into a trained first-arrival pickup model, and acquiring a first-arrival of the seismic data test samples, wherein the first-arrival pickup model is a first-arrival pickup model based on a U-net full-convolution neural network. The contraction path of the first arrival pickup model based on the U-net full convolution neural network comprises two units, wherein each unit comprises a convolution layer and a downsampling layer; the dilation path of the first arrival picking model based on the U-net full convolution neural network comprises two units, and each unit comprises an up-sampling layer and a convolution layer. The independent claim of this patent application is an application mode of a trained model, in particular picking up first arrivals based on a trained UNet model, but does not relate to a training mode and a data coding mode of the UNet network.
Therefore, the invention discloses a novel method and a device for training the first-arrival pickup model of the earthquake, and solves the technical problems.
Disclosure of Invention
The invention aims to provide a method and a device for training a first-arrival picking model of an earthquake, which can replace manual picking or a traditional method and improve the precision and the intelligent degree of first-arrival picking.
The aim of the invention can be achieved by the following technical measures: a method of training a seismic first-arrival pickup model, the method of training a seismic first-arrival pickup model comprising: step 1, preprocessing to obtain a plurality of groups of data, wherein the plurality of groups of data comprise first-class data and second-class data; step 2, selecting a training arrangement mode, wherein the training comprises a first type of training and a second type of training; step 3, generating data arrangement of each round of training; and 4, training the model until convergence.
The aim of the invention can be achieved by the following technical measures:
in step 1, each set of data in the first type of data includes: the preprocessed seismic data and the tags of the first arrival before and after arrival; each set of data in the second class of data comprises: the preprocessed seismic data and the first arrival time tags.
In step 1, the preprocessed seismic data in the first type data and the second type data are obtained by using the same preprocessing method as a model training input, or are obtained by using different preprocessing methods as a model training input.
In step 1, the labels of the first-arrival "before arrival and after arrival" in the first-class data refer to marking the sampling time before arrival of the first arrival and the sampling time after arrival of the first arrival as two different classifications according to the manually marked first-arrival time.
In step 1, the "when arriving" label in the second class data refers to marking the sampling time of first arrival and the sampling time before or after the first arrival as different classifications according to the manually marked first arrival time.
In step 2, the first type of training comprises: training a model by using the first type data and converging; the second type of training comprises: the model is trained and converged using the second class data.
In step 2, the training arrangement comprises a synchronous arrangement and an asynchronous arrangement.
In step 2, the synchronous arrangement means that in each iteration, the same preprocessed seismic data in the first type data and the second type data are used as model input, corresponding labels in the first type data and the second type data are used for comparing with model prediction results, training effects are comprehensively evaluated, and parameter distribution of the whole model is updated.
In step 2, the asynchronous arrangement means that in the first type of iteration, the preprocessed seismic data in the first type of data is used as model input, the corresponding label in the first type of data is used for comparing with a model prediction result, the training effect is evaluated, and the local parameter distribution of the model is updated; in the second class iteration, the preprocessed seismic data in the second class data are used as model input, corresponding labels in the second class data are used for comparing with model prediction results, training effects are evaluated, and local parameter distribution of the model is updated.
In step 3, the data arrangement method includes: sequential, random, or dynamic.
In step 3, the sequential arrangement refers to generating an arrangement result of the first-type iteration and the second-type iteration according to a preset rule.
In step 3, the random permutation refers to the permutation result executed by randomly generating the first class iteration and the second class iteration.
In step 3, the dynamic arrangement refers to dynamically adjusting the arrangement result executed by the subsequent first-class iteration and the second-class iteration according to the effect of the previous iteration in the training process.
The aim of the invention can be achieved by the following technical measures: apparatus for training a seismic first-arrival pickup model, the apparatus for training a seismic first-arrival pickup model comprising: a model first section, a model second section, a model third section, and a difference calculation section;
the first model part is used as a shared module, and shares model parameters and is used for receiving and processing first type data and second type data;
the second part of the model is used as a local module, the output of the first type of data in the first part of the model is used as input, and a prediction result is output;
the third part of the model is used as a local module, the output of the second class data in the first part of the model is used as input, and a prediction result is output;
the difference calculation section includes: and quantitatively calculating the difference between the output of the second part of the model and the first class data label, and quantitatively calculating the difference between the output of the third part of the model and the second class data label.
The method and the device for training the first arrival picking model of the earthquake can integrate more data features in one model, focus and accurately sense the first arrival, and improve the accuracy of the first arrival picking. Aiming at different seismic data processing methods or first arrival tag codes, the method organizes and arranges training until convergence under a model frame. The method and the device for training the first arrival pickup model of the earthquake can integrate more data features in one model, and the perception of the synchronous arrangement and the asynchronous arrangement of the cross training arrangement on the first arrival is more focused and accurate, so that the effect of 1+1>2 is achieved.
Drawings
FIG. 1 is a schematic diagram of a synchronous orchestration first-arrival picking model according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of an asynchronous orchestration first-arrival pick model based on a sequential arrangement according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an asynchronous orchestration first-arrival pick model based on dynamic alignment according to an embodiment of the present invention;
FIG. 4 is a flow chart of a method of training a first arrival pickup model in accordance with an embodiment of the present invention.
Detailed Description
The foregoing and other objects, features and advantages of the invention will be apparent from the following more particular description of preferred embodiments, as illustrated in the accompanying drawings.
As shown in FIG. 1, FIG. 1 is a flow chart of a method of training a first arrival pickup model of an earthquake according to the present invention. The method for training the first arrival pickup model of the earthquake comprises the following steps:
and step 1, preprocessing to obtain a plurality of groups of data. Wherein the plurality of sets of data includes a first type of data and a second type of data;
each set of data in the first class of data includes: the preprocessed seismic data and the tags of the first arrival before and after arrival;
each set of data in the second class of data comprises: the preprocessed seismic data and the label of first arrival time;
the preprocessed seismic data in the first type data and the second type data may be obtained by using the same preprocessing method as a model training input, or may be obtained by using different preprocessing methods as a model training input.
The labels of the first-arrival before and after arrival in the first-class data refer to marking the sampling time before the first-arrival and the sampling time after the first-arrival into two different classifications according to the manually marked first-arrival time. The label of the first arrival time in the second class data refers to marking the sampling time of the first arrival and the sampling time before or after the first arrival as different classifications according to the manually marked first arrival time.
And 2, selecting a training arrangement mode.
The training comprises a first type of training and a second type of training;
the first type of training comprises: training a model by using the first type data and converging;
the second type of training comprises: training a model by using the second class data and converging;
the training schedule includes synchronous schedule and asynchronous schedule.
The synchronous arrangement means that in each iteration, the same preprocessed seismic data in the first type data and the second type data are used as model input, corresponding labels in the first type data and the second type data are used for comparing with model prediction results, training effects are comprehensively evaluated, and parameter distribution of the whole model is updated. The asynchronous arrangement means that in the first type of iteration, the preprocessed seismic data in the first type of data is used as model input, corresponding labels in the first type of data are used for comparing with model prediction results, training effects are evaluated, and local parameter distribution of a model is updated. In the second class iteration, the preprocessed seismic data in the second class data are used as model input, corresponding labels in the second class data are used for comparing with model prediction results, training effects are evaluated, and local parameter distribution of the model is updated.
And step 3, generating data arrangement of each round of training.
The arrangement mode of the first type iteration and the second type iteration of the asynchronous arrangement comprises the following steps: sequential, random, or dynamic. The sequential arrangement refers to generating an arrangement result of the first-class iteration and the second-class iteration according to a preset rule. The random arrangement refers to the arrangement result of the first-type iteration and the second-type iteration which are randomly generated. The dynamic arrangement refers to the arrangement result of the subsequent first-class iteration and the second-class iteration which are dynamically adjusted according to the effect of the previous iteration in the training process.
And 4, training the model until convergence.
The following are several specific examples of the application of the present invention.
Embodiment one:
referring to FIG. 1, a schematic diagram of a synchronous orchestration first-arrival pick-up model includes steps 101-107. In this embodiment, the first type of data includes: the normalized or normalized amplitude data is shown in the inset of step 101; the sample points of the first arrival "before arrival and" after arrival "are labeled as categories 0 and 1, respectively, as labels, as shown in the inset of step 104. The second type of data includes: the amplitude data after normalization or normalization is the same as the preprocessing method of the first type of data, as shown in the inset of step 101; the first arrival "at" sample points are labeled as class 1, the other sample points are labeled as class 0, as a label, as shown in the inset at step 104.
In step 101, pre-processed amplitude data is acquired as a model input.
In step 102, a first portion of the model processes the input data, with possible processing methods including: downsampling, feature extraction, data transformation, etc.
In step 103, the second part of the model receives the output of step 102, and further processing obtains a classification prediction for each sample point, and possible processing methods are: upsampling, interpolation, inverse transformation, etc.
In step 105, the third part of the model receives the output of step 102, and further processing obtains a classification prediction for each sample point, and possible processing methods are: upsampling, interpolation, inverse transformation, etc.
In step 107, the difference between the model prediction in step 103 and the first class data tag obtained in step 104 is calculated, the difference between the model prediction in step 105 and the second class data tag obtained in step 106 is calculated, and after the difference between the predicted value and the labeling value of the two classes of data is fused, all parameters in the first part, the second part and the third part of the model are updated as a unified gradient. Thus, one iteration is completed. In step 107, a possible variance calculating method includes: absolute error, relative error, cross entropy, cross-over ratio, etc. Possible fusion methods include: weighting, addition, multiplication, etc.
Embodiment two:
referring to FIG. 2, a schematic diagram of an asynchronous orchestration first-arrival pick model based on a sequential arrangement includes steps 201-209. In this embodiment, the first type of data includes: the normalized or normalized amplitude data is shown in the inset of step 202; the sample points of the first arrival "before arrival and" after arrival "are labeled as categories 0 and 1, respectively, as labels, as shown in the inset of step 205. The second type of data includes: the amplitude data after normalization or normalization is the same as the preprocessing method of the first type of data, as shown in the inset of step 202; the first arrival "at sample points are labeled as class 1, the other sample points are labeled as class 0, as a label, as shown in the inset of step 207.
In step 201, the sequence generator sets the type of the next iteration according to a preset rule, and forms an iteration list. The embodiment provides a method for setting a fixed interval N, namely, after N iterations, switching to an iteration type. When n=1, iterate list: { first class iteration, second class iteration, first class iteration, second class iteration.
When the first type of iteration is selected, the following is performed:
in step 202, the preprocessed amplitude data is acquired as a model input.
In step 203, the first part of the model processes the input data, possible processing methods include: downsampling, feature extraction, data transformation, etc.
In step 204, the second part of the model receives the output of step 203, and further processing obtains a classification prediction for each sample point, and possible processing methods are: upsampling, interpolation, inverse transformation, etc.
In step 208, the differences between the model predictions of step 204 and the first type data tags obtained in step 205 are calculated, and as a unified gradient, all parameters in the first and second parts of the model are updated. Thus, one iteration is completed. In step 208, a possible variance calculation method includes: absolute error, relative error, cross entropy, cross-over ratio, etc.
When the second type of iteration is selected, the following is performed:
in step 202, the preprocessed amplitude data is acquired as a model input. In step 203, the first part of the model processes the input data, possible processing methods include: downsampling, feature extraction, data transformation, etc. In step 206, the second part of the model receives the output of step 203, and further processing obtains a classification prediction for each sample point, and possible processing methods are: upsampling, interpolation, inverse transformation, etc.
In step 209, the difference between the model prediction in step 206 and the second class data tag acquired in step 207 is calculated, and as a unified gradient, all parameters in the first part and the third part of the model are updated. Thus, one iteration is completed. In step 209, a possible variance calculation method includes: absolute error, relative error, cross entropy, cross-over ratio, etc.
Embodiment III:
referring to FIG. 3, a schematic diagram of an asynchronous orchestration first-arrival pick model based on a sequential arrangement includes steps 301-310. In this embodiment, the first type of data includes: the normalized or normalized amplitude data is shown in the inset of step 302; the sample points of the first arrival "before arrival and" after arrival "are labeled as categories 0 and 1, respectively, as labels, as illustrated in the inset of step 306. The second type of data includes: the normalized or normalized energy attribute is different from the preprocessing method of the first type of data, as shown in the inset of step 303; the first arrival "at sample points are labeled as class 1, the other sample points are labeled as class 0, as a label, as shown in the inset of step 308.
In step 301, the gradient values of the previous iteration are recorded in a gradient list, the previous iteration comprising: a first type of iteration and a second type of iteration. The dynamic arrangement generator sets the type of the next iteration according to the gradient list to form an iteration list. The embodiment provides a dynamic selection method for comparing average gradients, namely, average gradient values of first-class iteration and second-class iteration in a gradient list are calculated respectively, and the type with larger average value is selected as the type of the next iteration.
When the first type of iteration is selected, the following is performed:
in step 302, the preprocessed amplitude data is acquired as a model input. In step 304, a first portion of the model processes the input data, with possible processing methods including: downsampling, feature extraction, data transformation, etc. In step 305, the second part of the model receives the output of step 304 and further processing yields a classification prediction for each sample point, possibly with: upsampling, interpolation, inverse transformation, etc.
In step 309, the differences between the model predictions of step 305 and the first type data tags obtained in step 306 are calculated, all parameters in the first and second parts of the model are updated as uniform gradients, and a gradient list is recorded to the dynamic arrangement generator. Thus, one iteration is completed. In step 309, a possible variance calculation method includes: absolute error, relative error, cross entropy, cross-over ratio, etc.
When the second type of iteration is selected, the following is performed:
in step 303, the preprocessed amplitude data is acquired as a model input. In step 304, a first portion of the model processes the input data, with possible processing methods including: downsampling, feature extraction, data transformation, etc. In step 307, the second part of the model receives the output of step 304 and further processing yields a classification prediction for each sample point, possibly with: upsampling, interpolation, inverse transformation, etc.
In step 310, the differences between the model predictions of step 307 and the second class data tags obtained in step 308 are calculated, and as a unified gradient, all parameters in the first and third parts of the model are updated and recorded into the gradient list of the dynamic arrangement generator. Thus, one iteration is completed. In step 310, a possible variance calculation method includes: absolute error, relative error, cross entropy, cross-over ratio, etc.
The invention also relates to a device for training the first-arrival earthquake pickup model, which comprises: a model first part, a model second part, a model third part and a difference calculation part.
The first model part is used as a shared module, and shares model parameters and is used for receiving and processing first type data and second type data; the second part of the model is used as a local module, the output of the first type of data in the first part of the model is used as input, and a prediction result is output; the third part of the model is used as a local module, the output of the second class data in the first part of the model is used as input, and a prediction result is output; the difference calculation section includes: the second part of the quantitative calculation model outputs a difference from the first class data tag, and the third part of the quantitative calculation model outputs a difference from the second class data tag.
Claims (6)
1. A method of training a seismic first arrival pickup model, the method comprising:
step 1, preprocessing to obtain a plurality of groups of data, wherein the plurality of groups of data comprise first-class data and second-class data;
step 2, selecting a training arrangement mode, wherein the training comprises a first type of training and a second type of training;
step 3, generating data arrangement of each round of training;
step 4, training the model until convergence;
in step 1, each set of data in the first type of data includes: the preprocessed seismic data and tags before and after the first arrival; each set of data in the second class of data comprises: the preprocessed seismic data and the tags when the first arrival arrives;
in step 1, the preprocessed seismic data in the first type data and the second type data are obtained by using the same preprocessing method as a model training input, or obtained by using different preprocessing methods as a model training input;
in step 1, the labels before and after the first arrival in the first type of data refer to marking the sampling time before the first arrival and the sampling time after the first arrival as two different classifications according to the manually marked first arrival time;
in step 1, the label when the first arrival arrives in the second type data refers to marking the sampling time when the first arrival arrives and the sampling time before or after the first arrival as different classifications according to the manually marked first arrival time;
in step 2, the first type of training comprises: training a model by using the first type data and converging; the second type of training comprises: training a model by using the second class data and converging; the training arrangement comprises synchronous arrangement and asynchronous arrangement, wherein the synchronous arrangement means that in each iteration, the same preprocessed seismic data in the first type data and the second type data are used as model input, corresponding labels in the first type data and the second type data are used for comparing with model prediction results, the training effect is comprehensively evaluated, and the overall parameter distribution of the model is updated; the asynchronous arrangement means that in the first type of iteration, the preprocessed seismic data in the first type of data is used as model input, corresponding labels in the first type of data are used for comparing with model prediction results, training effects are evaluated, and local parameter distribution of a model is updated; in the second class iteration, the preprocessed seismic data in the second class data are used as model input, corresponding labels in the second class data are used for comparing with model prediction results, training effects are evaluated, and local parameter distribution of the model is updated.
2. The method of training a seismic first arrival pickup model according to claim 1, wherein in step 3, the data arrangement comprises: sequential, random, or dynamic.
3. The method for training a first arrival picking model according to claim 2, wherein in step 3, the sequential arrangement means that an arrangement result of the first-type iteration and the second-type iteration is generated according to a preset rule.
4. The method of training a first arrival picking model of an earthquake of claim 2, wherein in step 3, the random permutation is a permutation result of randomly generating first class iterations and second class iterations.
5. The method for training a first-arrival picking model according to claim 2, wherein in step 3, the dynamic arrangement refers to dynamically adjusting the arrangement result executed by the subsequent first-type iteration and the second-type iteration according to the effect of the previous iteration in the training process.
6. An apparatus for use in a method of training a first arrival picking model of an earthquake as claimed in claim 1 wherein the apparatus for training the first arrival picking model comprises: a model first section, a model second section, a model third section, and a difference calculation section;
the first model part is used as a shared module, and shares model parameters and is used for receiving and processing first type data and second type data;
the second part of the model is used as a local module, the output of the first type of data in the first part of the model is used as input, and a prediction result is output;
the third part of the model is used as a local module, the output of the second class data in the first part of the model is used as input, and a prediction result is output;
the difference calculation section includes: and quantitatively calculating the difference between the output of the second part of the model and the first class data label, and quantitatively calculating the difference between the output of the third part of the model and the second class data label.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011471662.9A CN114624768B (en) | 2020-12-14 | 2020-12-14 | Method and device for training earthquake first arrival pickup model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011471662.9A CN114624768B (en) | 2020-12-14 | 2020-12-14 | Method and device for training earthquake first arrival pickup model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114624768A CN114624768A (en) | 2022-06-14 |
| CN114624768B true CN114624768B (en) | 2023-06-30 |
Family
ID=81896987
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011471662.9A Active CN114624768B (en) | 2020-12-14 | 2020-12-14 | Method and device for training earthquake first arrival pickup model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114624768B (en) |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111368710A (en) * | 2020-02-27 | 2020-07-03 | 东北石油大学 | Seismic data random noise suppression method combined with deep learning |
Family Cites Families (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2010014118A1 (en) * | 2008-08-01 | 2010-02-04 | Ikelle Luc T | Statistical decoding and imaging of multishot and single-shot seismic data |
| CN102081167B (en) * | 2009-12-01 | 2012-08-15 | 中国石油天然气集团公司 | Picking method of three-dimensional VSP data first-motion wave |
| WO2017035104A1 (en) * | 2015-08-24 | 2017-03-02 | Schlumberger Technology Corporation | Velocity model seismic static correction |
| CN110869813A (en) * | 2017-06-08 | 2020-03-06 | 道达尔公司 | Method for acquiring a seismic dataset over a region of interest and related system |
| US11372124B2 (en) * | 2018-04-18 | 2022-06-28 | GeoTomo, LLC | First-break picking of seismic data and generating a velocity model |
| CN110954941B (en) * | 2018-09-26 | 2021-08-24 | 中国石油化工股份有限公司 | Automatic first arrival picking method and system |
| CN110163234B (en) * | 2018-10-10 | 2023-04-18 | 腾讯科技（深圳）有限公司 | Model training method and device and storage medium |
| CN111638551A (en) * | 2019-03-01 | 2020-09-08 | 中国石油天然气集团有限公司 | Seismic first-motion wave travel time chromatography method and device |
| CN110308484A (en) * | 2019-06-11 | 2019-10-08 | 中国石油大学(北京) | A kind of chromatography conversion method and system based on deep learning intelligent screening first arrival |
| CN111190227B (en) * | 2020-01-09 | 2021-06-04 | 吉林大学 | Low signal-to-noise ratio seismic data denoising method based on residual convolution generation countermeasure model |
| CN111695413B (en) * | 2020-04-28 | 2023-06-20 | 长江大学 | Signal first arrival pickup method and device combining U-Net and Temporal Ensembling |
-
2020
- 2020-12-14 CN CN202011471662.9A patent/CN114624768B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111368710A (en) * | 2020-02-27 | 2020-07-03 | 东北石油大学 | Seismic data random noise suppression method combined with deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114624768A (en) | 2022-06-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhou et al. | Hybrid event detection and phase‐picking algorithm using convolutional and recurrent neural networks | |
| Zhu et al. | PhaseNet: a deep-neural-network-based seismic arrival-time picking method | |
| Ross et al. | P wave arrival picking and first‐motion polarity determination with deep learning | |
| Wu et al. | Semiautomatic first-arrival picking of microseismic events by using the pixel-wise convolutional image segmentation method | |
| Hammer et al. | A seismic‐event spotting system for volcano fast‐response systems | |
| Yang et al. | Simultaneous earthquake detection on multiple stations via a convolutional neural network | |
| Bergen et al. | Preface to the focus section on machine learning in seismology | |
| CN111723329B (en) | Seismic phase feature recognition waveform inversion method based on full convolution neural network | |
| CN106054155B (en) | Radar High Range Resolution target identification method based on warp factor analysis model | |
| Duan et al. | Multitrace first-break picking using an integrated seismic and machine learning method | |
| CN110619264B (en) | Method and device for identifying microseism effective signals based on UNet++, and method and device for identifying microseism effective signals based on UNet++ | |
| Liao et al. | ARRU phase picker: Attention recurrent‐residual U‐Net for picking seismic P‐and S‐phase arrivals | |
| Shen et al. | Array‐based convolutional neural networks for automatic detection and 4D localization of earthquakes in Hawai ‘i | |
| Liu et al. | Microseismic first-arrival picking using fine-tuning feature pyramid networks | |
| Zheng et al. | SC-PSNET: A deep neural network for automatic P-and S-phase detection and arrival-time picker using 1C recordings | |
| Wang et al. | Seismic stratum segmentation using an encoder–decoder convolutional neural network | |
| CN109558898A (en) | A kind of more options learning method of the high confidence level based on deep neural network | |
| Timmer et al. | Applying deep learning for identifying bioturbation from core photographs | |
| CN117056680A (en) | Data noise reduction and signal detection method, device and system and storage medium | |
| Smith et al. | Robust deep learning-based seismic inversion workflow using temporal convolutional networks | |
| CN114624768B (en) | Method and device for training earthquake first arrival pickup model | |
| CN113325480A (en) | Seismic lithology identification method based on integrated deep learning | |
| Ma et al. | Machine learning-assisted processing workflow for multi-fiber DAS microseismic data | |
| Dhara et al. | Seismic image registration using multiscale convolutional neural networks | |
| CN115345322A (en) | Small sample radar target identification method based on hierarchical element migration |
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 |