CN113514881A - Method for predicting speed model based on two-dimensional or three-dimensional VSP data - Google Patents
Method for predicting speed model based on two-dimensional or three-dimensional VSP data Download PDFInfo
- Publication number
- CN113514881A CN113514881A CN202110611434.5A CN202110611434A CN113514881A CN 113514881 A CN113514881 A CN 113514881A CN 202110611434 A CN202110611434 A CN 202110611434A CN 113514881 A CN113514881 A CN 113514881A
- Authority
- CN
- China
- Prior art keywords
- dimensional
- data
- velocity model
- training
- neural network
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 72
- 238000012549 training Methods 0.000 claims abstract description 89
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 59
- 238000012360 testing method Methods 0.000 claims description 23
- 238000012795 verification Methods 0.000 claims description 20
- 238000001514 detection method Methods 0.000 claims description 18
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000010586 diagram Methods 0.000 description 9
- 238000004364 calculation method Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 7
- 230000006870 function Effects 0.000 description 4
- 238000003384 imaging method Methods 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000013508 migration Methods 0.000 description 2
- 230000005012 migration Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000001737 promoting effect Effects 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008602 contraction Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000005553 drilling Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 150000003839 salts Chemical class 0.000 description 1
- 238000003325 tomography Methods 0.000 description 1
- 238000012546 transfer Methods 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
- G01V1/303—Analysis for determining velocity profiles or travel times
-
- 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
- G01V2210/62—Physical property of subsurface
- G01V2210/622—Velocity, density or impedance
- G01V2210/6222—Velocity; travel time
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 present disclosure provides a method for predicting a velocity model based on two-dimensional or three-dimensional VSP data, the method for predicting a velocity model based on two-dimensional VSP data comprising: acquiring a mass two-dimensional velocity model, and acquiring two-dimensional VSP seismic data generated by forward modeling of seismic waveforms through a two-dimensional observation system; inputting each two-dimensional velocity model and two-dimensional VSP seismic data into a convolutional neural network for training; and acquiring two-dimensional VSP seismic data to be predicted, and inputting the two-dimensional VSP seismic data into a convolutional neural network to obtain a corresponding two-dimensional velocity model. The method provided by the disclosure can be used for predicting other new seismic data by adopting the network model trained by theoretical data, and can efficiently, reliably and accurately obtain the velocity model.
Description
Technical Field
The disclosure relates to the technical field of seismic monitoring, in particular to a method for predicting a velocity model based on two-dimensional or three-dimensional VSP data.
Background
Seismic wave velocity is one of the most important parameters in seismic exploration, and velocity modeling is also an indispensable step in seismic exploration, and is performed throughout the whole process of seismic data acquisition, processing and interpretation. Accurate velocity models are a key prerequisite for implementing reverse time migration and other high resolution seismic imaging techniques. The high-resolution velocity model can be used for seismic data interpretation, can provide a velocity field with higher precision for migration imaging, and can also provide strong evidence for reservoir identification. Conventional methods for obtaining velocity models are tomography and Full Waveform Inversion (FWI), but such methods for reconstructing velocity models by stepwise iteration are time consuming and computationally expensive, and require an initial velocity model to be given for Inversion. Therefore, there is a need for an efficient, reliable, and accurate velocity model inversion technique.
Many efforts have been made by predecessors to invert velocity models directly from pre-stack seismic data using Machine Learning (ML) techniques. For example, Mosser et al proposed in 2018 that a generative countermeasure network with period constraints was used to represent the problem as a domain transfer problem for seismic inversion, and the mapping relationship between the post-stack seismic traces and the longitudinal wave velocity model was approximated by using this learning method. Before training the network, the seismic traces are converted from the time domain to the depth domain according to a velocity model, so that the input and output of the training are both the depth domain. Yan Chang Shu proposed a full convolution neural network (FCN) in 2019 for salt dome detection and reconstruction of subsurface velocity models for raw seismic data. The method takes the seismic signals received by 29 detectors placed on the earth surface as the input of a network, directly outputs a two-dimensional velocity model, tests theoretical data on the method, compares the prediction result of the network with the inversion result of the FWI, and is considerable. Wanglong proposed a VMB network in 2020 for estimating P-wave velocity during the inter-well oil recovery process, presenting good prediction results.
However, no research has been proposed to use the artificial intelligence method for Vertical Seismic Profile (VSP) velocity modeling. VSP plays an important role in reservoir prediction and description. The VSP data has the characteristics of accurate depth and near reservoir observation, can provide accurate speed information and time-depth relation, improves the precision of surface data imaging processing, and provides reliable geological body position and depth for a drilling target. Therefore, it is of great significance to reconstruct the subsurface velocity model using VSP data.
Disclosure of Invention
Technical problem to be solved
In order to solve the problems, the method for predicting the speed model based on the two-dimensional or three-dimensional VSP data is used for at least partially solving the technical problems that the traditional speed model inversion is large in calculation amount, time is consumed, an initial speed model is depended on, manual intervention is needed and the like.
(II) technical scheme
One aspect of the present disclosure provides a method for predicting a velocity model based on two-dimensional VSP data, including: acquiring a mass two-dimensional velocity model, and acquiring two-dimensional VSP seismic data generated by forward modeling of seismic waveforms through a two-dimensional observation system; inputting each two-dimensional velocity model and two-dimensional VSP seismic data into a convolutional neural network for training; and acquiring two-dimensional VSP seismic data to be predicted, and inputting the two-dimensional VSP seismic data into a convolutional neural network to obtain a corresponding two-dimensional velocity model.
Further, acquiring two-dimensional VSP seismic data generated by forward modeling of seismic waveforms through a two-dimensional observation system further includes establishing a two-dimensional observation system, including: placing at least one detector in a vertical well, and sequentially setting the detectors according to a fixed distance; and placing the shot points on the ground surface, wherein the shot points comprise one shot and a plurality of shots.
Further, before each two-dimensional velocity model is input into the convolutional neural network in combination with the two-dimensional VSP seismic data for training, the method further comprises: and setting parameters of the convolutional neural network according to the two-dimensional seismic data and the matrix size of the corresponding two-dimensional velocity model.
Further, the training of each two-dimensional velocity model in combination with the input of the two-dimensional VSP seismic data into the convolutional neural network further comprises: dividing the two-dimensional seismic data and the corresponding two-dimensional velocity model into a training set, a verification set and a test set, wherein the three data sets are different; training the convolutional neural network by using a training set and a verification set; the convolutional neural network is predicted using a test set.
Further, the shape of the massive two-dimensional velocity model includes: the horizontal layer shape, the inclined layer shape and the bent layer shape, the number of layers is 5-13, the speed is 2000-5000 m/s, and the speed is increased along with the increase of the depth.
Another aspect of the present disclosure provides a method for predicting a velocity model based on three-dimensional VSP data, comprising: acquiring a mass three-dimensional velocity model, acquiring three-dimensional VSP seismic data generated by forward modeling of seismic waveforms through a three-dimensional observation system, and extracting common detection point data according to the three-dimensional VSP seismic data; inputting each three-dimensional speed model and the common detection point data into a convolution neural network for training; acquiring three-dimensional VSP seismic data to be predicted, extracting common detection point data and inputting the common detection point data into a convolutional neural network to obtain a corresponding three-dimensional velocity model.
Further, acquiring three-dimensional VSP seismic data generated by forward modeling of seismic waveforms through a three-dimensional observation system includes establishing a three-dimensional observation system, including: placing at least one detector in a vertical well, and sequentially setting the detectors according to a fixed distance; and placing the shot points on the ground surface, wherein the shot points are uniformly distributed in a circle by taking the wellhead of the vertical well as the center of the circle.
Further, before inputting each three-dimensional velocity model into the convolutional neural network in combination with the common-detector-point data for training, the method further comprises: and setting parameters of the convolutional neural network according to the co-detection point data and the matrix size of the corresponding three-dimensional velocity model.
Further, the training of inputting each three-dimensional velocity model into the convolutional neural network in combination with the common detector point data further comprises: dividing the three-dimensional seismic data and the corresponding three-dimensional velocity model into a training set, a verification set and a test set, wherein the three data sets are different; training the convolutional neural network by using a training set and a verification set; the convolutional neural network is predicted using a test set.
Further, the shape of the mass three-dimensional velocity model to the X-Z plane comprises: the horizontal layer shape, the inclined layer shape and the bent layer shape, the number of layers is 3-8, the speed is 1500-3500 m/s, and the speed is increased along with the increase of the depth.
(III) advantageous effects
According to the method for predicting the velocity model based on the two-dimensional or three-dimensional VSP data, the two-dimensional or three-dimensional VSP pre-stack seismic data are utilized, the underground velocity model is inverted through the convolutional neural network, the two-dimensional or three-dimensional velocity model can be rapidly and accurately predicted, and high accuracy is achieved. The method improves the problems that the traditional method is large in calculation amount, needs an initial velocity model and the like, is an efficient, reliable and accurate velocity model inversion technology, and is beneficial to promoting the development of the velocity model inversion technology.
Drawings
FIG. 1 schematically illustrates a flow chart of a method of predicting a velocity model based on two-dimensional VSP data according to an embodiment of the disclosure;
FIG. 2 schematically illustrates a schematic diagram of 4 two-dimensional velocity models in accordance with an embodiment of the disclosure;
FIG. 3 schematically illustrates two-dimensional VSP seismic data corresponding to the 4 two-dimensional velocity models of FIG. 2;
FIG. 4 is a block diagram schematically illustrating a convolutional neural network constructed by inversion of a velocity model according to an embodiment of the present disclosure;
FIG. 5 is a graph schematically illustrating the predicted results of a trained network model versus a horizontal laminar velocity model according to an embodiment of the present disclosure;
FIG. 6 is a diagram schematically illustrating a result of a prediction of a trained network model versus an upward-sloping laminar velocity model according to an embodiment of the present disclosure;
FIG. 7 is a diagram schematically illustrating a prediction result of a trained network model on a downward-inclined laminar velocity model according to an embodiment of the present disclosure;
FIG. 8 schematically illustrates a flow chart of a method of predicting a velocity model based on three dimensional VSP data in an embodiment in accordance with the present disclosure;
FIG. 9 schematically illustrates a schematic diagram of 3 three-dimensional velocity models in an embodiment in accordance with the present disclosure;
FIG. 10 schematically illustrates a 3D observation system for generating 3D seismic data builds in accordance with embodiments of the disclosure;
FIG. 11 schematically illustrates 3D co-detector point data corresponding to the 3 three-dimensional velocity models of FIG. 9;
FIG. 12 schematically shows a schematic diagram of extracting 3D co-detector point data according to an embodiment of the disclosure;
FIG. 13 schematically shows a flow chart of a method of predicting a velocity model based on two-dimensional VSP data according to embodiment 1 of the present disclosure;
fig. 14 schematically shows a flowchart of a method of predicting a velocity model based on three-dimensional VSP data according to embodiment 2 of the present disclosure.
Detailed Description
For the purpose of promoting a better understanding of the objects, aspects and advantages of the present disclosure, reference is made to the following detailed description taken in conjunction with the accompanying drawings.
Based on the problems in the prior art, the purpose of the present disclosure is to provide an efficient, reliable and accurate velocity model inversion technique, which utilizes an artificial intelligence method to perform VSP velocity modeling, and can solve the problems of large calculation amount, time consumption, dependence on an initial velocity model, need of manual intervention and the like in the conventional velocity model inversion.
FIG. 1 schematically shows a flow chart of a method of predicting a velocity model based on two-dimensional VSP data according to an embodiment of the disclosure.
As shown in fig. 1, the method of predicting a velocity model based on two-dimensional VSP data includes:
in operation S11, a mass two-dimensional velocity model is obtained, and two-dimensional VSP seismic data generated by forward evolution of seismic waveforms is obtained through a two-dimensional observation system.
For 2DVSP velocity modeling, one-shot or multi-shot seismic data is used as input, and a two-dimensional velocity model is output. Firstly, a 2D velocity model and a 2D observation system need to be established, and proper parameters are selected for waveform forward modeling to obtain 2D seismic data for subsequent network training. Fig. 2 schematically shows a schematic diagram of 4 two-dimensional velocity models, including a horizontal layer a, an obliquely downward layer b, an obliquely downward layer c, and a curved layer d, although the velocity model is not limited to these 4 shapes, and may be any other shape. Fig. 3 schematically shows 2DVSP seismic data corresponding to the aforementioned 4 two-dimensional velocity models.
In operation S12, each of the two-dimensional velocity models is input into a convolutional neural network in conjunction with the two-dimensional VSP seismic data.
Then, according to the generated 2DVSP seismic data, a network for velocity model inversion is built; and using the prepared training data for the constructed convolutional neural network to train so as to obtain a network capable of predicting the speed model.
In operation S13, two-dimensional VSP seismic data to be predicted is obtained, and the two-dimensional VSP seismic data is input to the convolutional neural network to obtain a corresponding two-dimensional velocity model.
And inputting the seismic data to be predicted into the trained network model, so that the prediction of the velocity model can be carried out. Although the training phase of the network takes time, the network training is only needed once, once the training is completed, the cost of predicting the speed model is negligible, and the prediction can be completed in only a few seconds. The network model trained by the theoretical data can be used for predicting other new seismic data and is not limited to the same set of seismic data. The method well solves the problems of large calculation amount, time consumption, dependence on an initial velocity model and need of manual intervention in the conventional velocity model inversion, and is an efficient, reliable and accurate velocity model inversion technology.
On the basis of the above embodiment, acquiring the two-dimensional VSP seismic data generated by forward modeling of the seismic waveform by using the two-dimensional observation system further includes establishing the two-dimensional observation system, including: placing at least one detector in a vertical well, and sequentially setting the detectors according to a fixed distance; and placing the shot points on the ground surface, wherein the shot points comprise one shot and a plurality of shots.
Establishing a 2D survey system, for example, involves placing 150 receivers in a vertical well, one receiver every 15 meters from 10 meters, and a shot point at the surface, either one shot or multiple shots.
On the basis of the above embodiment, before the training of inputting each two-dimensional velocity model into the convolutional neural network in combination with two-dimensional VSP seismic data, the method further includes: and setting parameters of the convolutional neural network according to the two-dimensional seismic data and the matrix size of the corresponding two-dimensional velocity model.
After enough 2D velocity models in any shapes and 2D observation systems are built, Finite Difference method (FD) is used for forward modeling of waveforms to generate 2D seismic data (forward modeling parameters refer to table 1), and enough data for network training and corresponding labels (velocity models) can be generated to ensure the follow-up work. The training network for building the speed modeling comprises the following steps: according to the 2D seismic data and the matrix size of the corresponding training label, the parameters of the initial convolutional neural network are modified so as to be suitable for the prediction of the DVSP velocity model in the embodiment 2.
| Number of seismic sources | Sampling interval | Dominant frequency | Duration of |
| 1 | 1ms | 20Hz | 2s |
TABLE 1
On the basis of the above embodiment, the training of each two-dimensional velocity model in combination with the input of two-dimensional VSP seismic data into the convolutional neural network further comprises: dividing the two-dimensional seismic data and the corresponding two-dimensional velocity model into a training set, a verification set and a test set, wherein the three data sets are respectively trained on the convolutional neural network by using the training set and the verification set; the convolutional neural network is predicted using a test set.
Fig. 4 schematically shows the structure of the convolutional neural network inversely built by the velocity model in this embodiment. Training the convolutional neural network includes:
the generated data sets are distributed into a training set, a verification set and a test set, the three data sets are generated by similar speed models and the same observation system, and the three data sets are different from each other;
the network training is carried out in a GPU workstation;
and inputting the test set into the trained network model to obtain a corresponding prediction result.
Wherein we use a loss function to represent the difference between the true velocity model and the predicted velocity model. The loss function used in the network is defined as the squared difference (L2 norm) between the predicted velocity model Vp and the ground true velocity model Vt:
in the formula, nx and nz are the numbers of grid points in the horizontal direction and the vertical direction, respectively. We give the real speed information Vt during the training process, but hide it during the prediction phase. Note that this loss function differs from conventional FWI, which computes the squared difference between observed and modeled seismic data.
Fig. 5, 6 and 7 show the prediction results of the trained network model on the horizontal laminar velocity model, the upward-inclined laminar velocity model and the downward-inclined laminar velocity model.
On the basis of the above embodiment, the shape of the massive two-dimensional velocity model includes: the horizontal layer shape, the inclined layer shape and the bent layer shape, the number of layers is 5-13, the speed is 2000-5000 m/s, and the speed is increased along with the increase of the depth.
Establishing enough 2D speed models in any shapes, wherein the speed models can be horizontal layers, inclined layers, bent layers and the like, the number of layers is limited to 5-13, the speed range is 2000-5000 m/s, and the speed is increased along with the increase of the depth.
FIG. 8 schematically illustrates a flow chart of a method of predicting a velocity model based on three dimensional VSP data according to an embodiment of the disclosure.
As shown in fig. 8, the method of predicting a velocity model based on three-dimensional VSP data includes:
in operation S21, a mass three-dimensional velocity model is obtained, three-dimensional VSP seismic data generated by forward evolution of seismic waveforms is obtained through the three-dimensional observation system, and common geophone point data is extracted from the three-dimensional VSP seismic data.
For 3DVSP velocity modeling, firstly, a 3D velocity model and a 3D observation system need to be established, proper parameters are selected for waveform forward modeling, 3D seismic data are obtained, and common detector point data are extracted to be used as the input of a network. Fig. 9 schematically shows a schematic diagram of 3 three-dimensional velocity models, including a horizontal layer a, an obliquely downward layer b, and a curved layer c, but the velocity model is not limited to these 3 shapes, and may be any other shape. Fig. 11 schematically shows 3D co-detector point data corresponding to the aforementioned 3 three-dimensional velocity models.
In operation S22, each three-dimensional velocity model is input into a convolutional neural network in combination with the common detector point data for training.
Then, according to the generated common detector point data, a network for velocity model inversion is built; and using the prepared training data for the constructed network to train so as to obtain the network capable of predicting the three-dimensional speed model.
In operation S23, three-dimensional VSP seismic data to be predicted is obtained, common geophone point data is extracted, and the common geophone point data is input to a convolutional neural network, so as to obtain a corresponding three-dimensional velocity model.
And extracting common geophone point data by using seismic data to be predicted, and inputting the common geophone point data into the trained network model to predict the velocity model. Fig. 12 schematically shows a schematic diagram of extracting 3D common-probe-point data, selecting a geophone from which the common-probe-point data is to be extracted, and screening out seismic signals of all shot points received by the geophone to obtain the common-probe-point data of the geophone. The method for reconstructing the underground velocity model from the 3DVSP data can solve the problems of large calculation amount, time consumption, dependence on an initial velocity model and the like in the conventional velocity model inversion.
On the basis of the above embodiment, acquiring three-dimensional VSP seismic data generated by forward evolution of a seismic waveform through a three-dimensional observation system includes establishing a three-dimensional observation system, including: placing at least one detector in a vertical well, and sequentially setting the detectors according to a fixed distance; and placing the shot points on the ground surface, wherein the shot points are uniformly distributed in a circle by taking the wellhead of the vertical well as the center of the circle.
FIG. 10 schematically illustrates a 3D observation system for generating 3D seismic data builds.
Establishing a 3D observation system, for example, includes placing 10 receivers (the inverted triangle in the middle of fig. 10 a) in a vertical well, starting from 800 meters, placing shot points (small dots uniformly distributed in fig. 10 b) on the ground surface, and taking the well (the vertical line in fig. 10 a) as the center of a circle and the distance from the outermost shot point (the small dot in fig. 10 b) to the well head (the large dot in the middle of fig. 10 b) as the radius to form a circle.
On the basis of the above embodiment, before inputting each three-dimensional velocity model into the convolutional neural network in combination with the common detector point data for training, the method further includes: and setting parameters of the convolutional neural network according to the co-detection point data and the matrix size of the corresponding three-dimensional velocity model.
After enough 3D velocity models in any shapes and 3D observation systems are built, waveform forward modeling is carried out by using a finite difference method to generate 3D seismic data (forward modeling parameters refer to a table 1), common detection point data are extracted to be used as input of a network, and enough data used for network training and corresponding labels (velocity models) can be generated to ensure the follow-up work. The training network for building the speed modeling comprises the following steps: according to the 3D seismic data and the matrix size of the corresponding training label, the parameters of the initial convolutional neural network are modified so as to be suitable for the prediction of the 3DVSP velocity model in the embodiment.
On the basis of the above embodiment, inputting each three-dimensional velocity model into the convolutional neural network in combination with the common detector point data for training further includes: dividing the three-dimensional seismic data and the corresponding three-dimensional velocity model into a training set, a verification set and a test set, wherein the three data sets are different; training the convolutional neural network by using a training set and a verification set; the convolutional neural network is predicted using a test set.
The network training and predicting the generated training data comprises the following steps:
the generated data sets are distributed into a training set, a verification set and a test set, the three data sets are generated by similar speed models and the same observation system, and the three data sets are different from each other;
the network training is carried out in a GPU workstation;
and inputting the test set into the trained network model to obtain a prediction result of the 3D speed model.
On the basis of the above embodiment, the shape of the mass three-dimensional velocity model for the X-Z plane includes: the horizontal layer shape, the inclined layer shape and the bent layer shape, the number of layers is 3-8, the speed is 1500-3500 m/s, and the speed is increased along with the increase of the depth.
Referring to fig. 9, the shape of the three-dimensional velocity model in the X-Z and Y-Z cross sections is horizontal layer, inclined layer, or curved layer, and has a well-defined layer structure, and the shape, number of layers, and velocity range of the three-dimensional velocity model can be adjusted according to actual situations.
The method of the present disclosure for predicting a velocity model based on two-dimensional or three-dimensional VSP data is further illustrated in two specific embodiments below.
Example 1: two-dimensional VSP data prediction speed model
and step 4, inputting new seismic data into the trained network model, and predicting the velocity model, which is equivalent to step S13.
Referring to fig. 13, in step 1 of the method, the method for generating 2DVSP training data includes: firstly, 3000 2D speed models are established, wherein 500 horizontal layered speed models, 800 inclined layered speed models and 1700 other irregular-shape speed models are established, as shown in FIG. 2, all the speed models are limited to 5-13 layers, the speed range is 2000-5000 m/s, and the speed is increased along with the increase of the depth; then, a 2D observation system is established, 150 detectors are placed in a vertical well (black inverted triangle in figure 2 a), every 15 meters of detectors are placed from 10 meters to 2245 meters, the cannon is placed on the ground surface, can be a cannon, is placed at a detection distance of 1000 meters (a hollow circular point at the upper right corner in figure 2 a), or can be multiple cannons, and every 100 meters of cannon points are placed from a detection distance of 100 meters; as shown in fig. 3, the waveform forward modeling is finally performed by using the FD method to obtain 2D seismic data, where the dominant frequency of the rake wavelet is 20 hz, the sampling interval is 1 ms, and the duration is two seconds. Although it takes time to generate training data in this process, it is only necessary to generate training data once and for subsequent network training, and once the training data is ready, the network training is completed, and the work does not need to be repeated.
Referring to fig. 4, the principle of the convolutional neural network for reconstructing the subsurface velocity model in step 2 is: first, 2D seismic data is input into a network consisting of a contracted path (left side, i.e., down-sampled portion) for extracting geological features and an expanded path (right side, i.e., up-sampled portion) for accurate velocity estimation, wherein each step of the left contracted path contains two 3 x 3 convolutional layers followed by a ReLU activation function and 2 x 2 max pooling operation with a step size of 2, and each step of the right expanded path contains one 2 x 2 up-convolutional layer and two 3 x 3 convolutional layers and a ReLU activation operation. The connection can combine the high resolution features from the contraction path with the up-sampled output from the expansion path. The network body in this disclosure is similar to the original U-Net architecture, and contains a total of 23 convolutional layers.
In step 3 and step 4 of the method, the built convolutional neural network is trained to obtain a prediction result: firstly, dividing a generated data set into a training set, a verification set and a test set, wherein the training set, the verification set and the test set respectively comprise 2300, 500 and 200 pieces of 2D seismic data; then training the convolutional neural network by using a training set and a verification set; and finally, predicting 200 seismic data in the test set by the trained network model to obtain a prediction result. The training process also only needs to be performed once, and once the training is completed, the trained network model can be used for predicting new 2D data.
Referring to fig. 5, 6 and 7, the results of comparing the predicted results of the horizontal laminar velocity model, the upward-inclined laminar velocity model and the downward-inclined laminar velocity model with the actual velocity model show that the method provided by the present disclosure can reconstruct the underground velocity model by using the 2DVSP data, and compare the predicted results with the actual velocity model, and the convolutional neural network can successfully invert all parallel velocity interfaces regardless of whether the number of layers of the velocity model is large or small, and regardless of whether the velocity model is the inclined velocity model or the horizontal laminar velocity model, and has accurate depth, velocity and inclination angle, and furthermore, most of the velocity interfaces are clearly and continuously depicted and are consistent with the actual velocity model. The preliminary prediction results show that the method provided by the disclosure can quickly and accurately predict the 2D velocity model and has higher accuracy.
Example 2: three-dimensional VSP data prediction speed model
As shown in fig. 14, a method for reconstructing a subsurface velocity model from 3DVSP data by using a convolutional neural network according to an embodiment of the present disclosure mainly includes the following steps:
and 4, inputting new seismic data into the trained network model, and predicting the velocity model.
In step 1 of the method, the method for generating 3D common detection point data includes: firstly, 3000 3D speed models are established, wherein 500 horizontal layered speed models, 800 inclined layered speed models and 1700 other irregular-shape speed models are established, as shown in FIG. 9, all the speed models are limited to 3-8 layers, the speed range is 1500-3500 m/s, and the speed is increased along with the increase of the depth; then, as shown in fig. 10, a 3D observation system is established, 10 receivers (an inverted triangle in the middle of fig. 10 a) are placed in a vertical well, every 30 meters of the receivers are from 800 meters to 1070 meters, shot points (small dots uniformly distributed in fig. 10 b) are placed on the ground, the well (a vertical line in fig. 10 a) is used as a circle center, the distance from the outermost shot point (the small dot in fig. 10 b) to the well head (a large dot in the middle of fig. 10 b) is used as a radius, the radius is 400 meters at most, and the intervals between the shot points in the x direction and the y direction are 20 meters; performing waveform forward modeling by using a finite difference method to obtain 3D seismic data, wherein the main frequency of a Rake wavelet is 20 Hz, the sampling interval is 1 millisecond, and the duration is two seconds; finally, as shown in fig. 11, the common checkpoint data is extracted as input for network training. Although it takes time to generate training data in this process, the generation of training data and the subsequent network training need only be performed once, and once the training data is prepared and the network training is completed, the operation does not need to be repeated.
Referring to fig. 12, the method for extracting the common-detector-point data in step 1 mainly includes: and selecting a detector for extracting the common detection point data, and screening the seismic signals of all shot points received by the detector to obtain the common detection point data of the detector.
Fig. 10a and 10b are a front view and a top view, respectively, of the 3D observation system established in step 1 of the method, wherein the middle vertical line represents a vertical well, the vertical well is located at a position where x is 500 meters and y is 500 meters, 10 inverted triangles represent 10 detectors, the detectors are distributed at intervals of 30 meters from a depth of 800 meters, shot points are located on the ground surface, the shot points are distributed in a circle with a radius of 400 meters and a well (500m ) as a center, and the intervals of the shot points in the x direction and the y direction are both 20 meters.
In step 2 of the method, according to the generated 3D seismic data, a training network for building the velocity modeling comprises the following steps:
the initial convolutional neural network is modified according to the matrix size of the 3D seismic data and the corresponding training labels (velocity model) to be suitable for prediction of the 3DVSP velocity model of this embodiment.
In step 3 and step 4 of the method, training with a convolutional neural network to obtain a prediction result of a 3D velocity model mainly comprises: firstly, dividing a generated data set into a training set, a verification set and a test set, wherein the training set, the verification set and the test set respectively comprise 2300, 500 and 200 pieces of 3D seismic data; then training the convolutional neural network by using a training set and a verification set; and finally, predicting 200 seismic data in the test set by the trained network model to obtain a prediction result. The training process needs to be performed only once, and once training is completed, the trained network model can be used for prediction of new 3D data.
The present disclosure utilizes 2D or 3DVSP prestack seismic data to invert a subsurface velocity model through a convolutional neural network. The method solves the problems that the traditional method is large in calculation amount and time-consuming due to an iterative optimization process, an initial velocity model needs to be given for inversion, and the like, can be used for inversion of 2D and 3DVSP velocity models, and the prediction result shows that the method is high in calculation speed and accurate in prediction result, is an efficient, reliable and accurate velocity model inversion technology, is expected to form an international advanced velocity model inversion technology in the VSP velocity modeling field, promotes the development of the velocity model inversion technology, and plays an important role in improving the onshore VSP velocity model inversion technology level.
The above-mentioned embodiments are intended to illustrate the objects, aspects and advantages of the present disclosure in further detail, and it should be understood that the above-mentioned embodiments are only illustrative of the present disclosure and are not intended to limit the present disclosure, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present disclosure should be included in the scope of the present disclosure.
Claims (10)
1. A method for predicting a velocity model based on two-dimensional VSP data, comprising:
acquiring a mass two-dimensional velocity model, and acquiring two-dimensional VSP seismic data generated by forward modeling of seismic waveforms through a two-dimensional observation system;
inputting each two-dimensional velocity model into a convolutional neural network in combination with the two-dimensional VSP seismic data for training;
and acquiring two-dimensional VSP seismic data to be predicted, and inputting the two-dimensional VSP seismic data into the convolutional neural network to obtain a corresponding two-dimensional velocity model.
2. The method of predicting a velocity model based on two-dimensional VSP data of claim 1, wherein the acquiring two-dimensional VSP seismic data generated by forward evolution of seismic waveforms through a two-dimensional observation system further comprises establishing a two-dimensional observation system comprising:
placing at least one detector in a vertical well, and sequentially setting the detectors according to a fixed distance;
and placing the shot points on the ground surface, wherein the shot points comprise one shot and a plurality of shots.
3. The method of predicting velocity models based on two-dimensional VSP data of claim 2, wherein prior to training each said two-dimensional velocity model in conjunction with said two-dimensional VSP seismic data input convolutional neural network, further comprises:
and setting parameters of a convolutional neural network according to the two-dimensional seismic data and the matrix size of the corresponding two-dimensional velocity model.
4. The method of predicting velocity models based on two-dimensional VSP data of claim 3, wherein said training each said two-dimensional velocity model in conjunction with said two-dimensional VSP seismic data input convolutional neural network further comprises:
dividing the two-dimensional seismic data and the corresponding two-dimensional velocity model into a training set, a verification set and a test set, wherein the three data sets are different;
training the convolutional neural network by using the training set and the verification set;
predicting the convolutional neural network using the test set.
5. The method of predicting a velocity model based on two-dimensional VSP data of claim 1, wherein the shape of the mass of two-dimensional velocity models comprises: the device comprises a horizontal layer, an inclined layer and a bent layer, wherein the number of layers is 5-13, the speed is 2000-5000 m/s, and the speed is increased along with the increase of the depth.
6. A method for predicting a velocity model based on three-dimensional VSP data, comprising:
acquiring a mass three-dimensional velocity model, acquiring three-dimensional VSP seismic data generated by forward modeling of seismic waveforms through a three-dimensional observation system, and extracting common detection point data according to the three-dimensional VSP seismic data;
inputting each three-dimensional speed model and the common detection point data into a convolutional neural network for training;
acquiring three-dimensional VSP seismic data to be predicted, extracting common detection point data and inputting the common detection point data into the convolutional neural network to obtain a corresponding three-dimensional velocity model.
7. The method of claim 6, wherein the obtaining three-dimensional VSP seismic data generated by forward evolution of seismic waveforms through a three-dimensional observation system comprises establishing a three-dimensional observation system comprising:
placing at least one detector in a vertical well, and sequentially setting the detectors according to a fixed distance;
and placing the shot points on the ground surface, wherein the shot points are uniformly distributed in a circle by taking the wellhead of the vertical well as the center of the circle.
8. The method of predicting velocity models based on three-dimensional VSP data of claim 7, wherein said training each said three-dimensional velocity model in conjunction with said co-detector point data input to a convolutional neural network further comprises:
and setting parameters of the convolutional neural network according to the co-detection point data and the matrix size of the corresponding three-dimensional velocity model.
9. The method of predicting velocity models based on three-dimensional VSP data of claim 8, wherein said training each said three-dimensional velocity model in conjunction with said inputting of co-detector point data into a convolutional neural network further comprises:
dividing the three-dimensional seismic data and the corresponding three-dimensional velocity model into a training set, a verification set and a test set, wherein the three data sets are different;
training the convolutional neural network by using the training set and the verification set;
predicting the convolutional neural network using the test set.
10. The method of predicting a velocity model based on three dimensional VSP data of claim 6, wherein the shape of the mass of three dimensional velocity models to the X-Z plane comprises: the device comprises a horizontal layer, an inclined layer and a bent layer, wherein the number of layers is 3-8, the speed is 1500-3500 m/s, and the speed is increased along with the increase of the depth.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110611434.5A CN113514881A (en) | 2021-06-01 | 2021-06-01 | Method for predicting speed model based on two-dimensional or three-dimensional VSP data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110611434.5A CN113514881A (en) | 2021-06-01 | 2021-06-01 | Method for predicting speed model based on two-dimensional or three-dimensional VSP data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113514881A true CN113514881A (en) | 2021-10-19 |
Family
ID=78065226
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110611434.5A Pending CN113514881A (en) | 2021-06-01 | 2021-06-01 | Method for predicting speed model based on two-dimensional or three-dimensional VSP data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113514881A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2024030565A1 (en) * | 2022-08-05 | 2024-02-08 | Saudi Arabian Oil Company | Method to correct vsp data |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN111239802A (en) * | 2020-01-19 | 2020-06-05 | 中国海洋大学 | Deep learning speed modeling method based on seismic reflection waveform and velocity spectrum |
| US20200183035A1 (en) * | 2018-12-11 | 2020-06-11 | Exxonmobil Upstream Research Company | Data Augmentation for Seismic Interpretation Systems and Methods |
| US20200184374A1 (en) * | 2018-12-11 | 2020-06-11 | Exxonmobil Upstream Research Company | Automated Seismic Interpretation Systems and Methods for Continual Learning and Inference of Geological Features |
| US20200183031A1 (en) * | 2018-12-11 | 2020-06-11 | Exxonmobil Upstream Research Company | Automated seismic interpretation-guided inversion |
| US20200183032A1 (en) * | 2018-12-11 | 2020-06-11 | Exxonmobil Upstream Research Company | Training machine learning systems for seismic interpretation |
| CN111626355A (en) * | 2020-05-27 | 2020-09-04 | 中油奥博(成都)科技有限公司 | Unet + + convolutional neural network-based seismic data first arrival pickup method |
| CN111723329A (en) * | 2020-06-19 | 2020-09-29 | 南京大学 | Seismic phase feature recognition waveform inversion method based on full convolution neural network |
-
2021
- 2021-06-01 CN CN202110611434.5A patent/CN113514881A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200183035A1 (en) * | 2018-12-11 | 2020-06-11 | Exxonmobil Upstream Research Company | Data Augmentation for Seismic Interpretation Systems and Methods |
| US20200184374A1 (en) * | 2018-12-11 | 2020-06-11 | Exxonmobil Upstream Research Company | Automated Seismic Interpretation Systems and Methods for Continual Learning and Inference of Geological Features |
| US20200183031A1 (en) * | 2018-12-11 | 2020-06-11 | Exxonmobil Upstream Research Company | Automated seismic interpretation-guided inversion |
| US20200183032A1 (en) * | 2018-12-11 | 2020-06-11 | Exxonmobil Upstream Research Company | Training machine learning systems for seismic interpretation |
| 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 |
| CN111239802A (en) * | 2020-01-19 | 2020-06-05 | 中国海洋大学 | Deep learning speed modeling method based on seismic reflection waveform and velocity spectrum |
| CN111626355A (en) * | 2020-05-27 | 2020-09-04 | 中油奥博(成都)科技有限公司 | Unet + + convolutional neural network-based seismic data first arrival pickup method |
| CN111723329A (en) * | 2020-06-19 | 2020-09-29 | 南京大学 | Seismic phase feature recognition waveform inversion method based on full convolution neural network |
Non-Patent Citations (1)
| Title |
|---|
| 毛博: "基于卷积神经网络的地震数据重构与模型构建研究", 《中国优秀博硕士学位论文全文数据库（博士）基础科学辑》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2024030565A1 (en) * | 2022-08-05 | 2024-02-08 | Saudi Arabian Oil Company | Method to correct vsp data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP3254142B1 (en) | A method for determining sedimentary facies using 3d seismic data | |
| US5999488A (en) | Method and apparatus for migration by finite differences | |
| US20100054082A1 (en) | Reverse-time depth migration with reduced memory requirements | |
| US10234582B2 (en) | Joint inversion of seismic data | |
| CN106526674A (en) | Three-dimensional full waveform inversion energy weighted gradient preprocessing method | |
| EA011519B1 (en) | Methods foe earth modeling and seismic imaging using interactive and selective updating | |
| WO2012139082A1 (en) | Event selection in the image domain | |
| US10816685B2 (en) | System and method for reconstructed wavefield imaging | |
| AU2015280679B2 (en) | Instantaneous isochron attribute-based geobody identification for reservoir modeling | |
| CN102395902A (en) | Seismic imaging systems and methods employing a fast target-oriented illumination calculation | |
| AU2012260680A1 (en) | A method to aid in the exploration, mine design, evaluation and/or extraction of metalliferous mineral and/or diamond deposits | |
| US6018499A (en) | Three-dimensional seismic imaging of complex velocity structures | |
| CN107765337B (en) | Electric method and earthquake synchronous joint inversion method and system | |
| EP3436849B1 (en) | Determining displacement between seismic images using optical flow | |
| CN109459787B (en) | coal mine underground structure imaging method and system based on seismic channel wave full-waveform inversion | |
| EP3436850B1 (en) | Determining displacement between seismic images using optical flow | |
| Le Meur et al. | Seismic tomography of the Gulf of Corinth: a comparison of methods | |
| CN108680968B (en) | Evaluation method and device for seismic exploration data acquisition observation system in complex structural area | |
| US5764514A (en) | Method for modelling kinematic seismic data processed with at least one motion operator | |
| CN113514881A (en) | Method for predicting speed model based on two-dimensional or three-dimensional VSP data | |
| Ma et al. | Velocity model building from one-shot VSP data via a convolutional neural network | |
| Brunini et al. | Differential evolution for microseismic event location | |
| Noble et al. | Large-sized nonlinear inversion of a marine data set: retrieving the source, background velocity, and impedance contrasts | |
| Tang et al. | Subsalt velocity analysis by target-oriented wavefield tomography: A 3-D field-data example | |
| Feighner | Three-dimensional seismic imaging of the Ryepatch geothermal reservoir |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20211019 |