阅读说明：本技术 一种构建高分辨率波阻抗反演标签的自洽深度学习方法 (Self-consistent deep learning method for constructing high-resolution wave impedance inversion label ) 是由 许辉群 于 2021-07-16 设计创作，主要内容包括：本发明公开了一种构建高分辨率波阻抗反演标签的自洽深度学习方法,属于深度学习地震波阻抗反演领域,解决了现有技术难以获取高精度大容量的样本问题,井插值具有高纵向分辨率,插值无法获取与地层一致性较好的横向特征的问题,本发明提供一种构建高分辨率波阻抗反演标签的自洽深度学习方法,其目的在于：通过对井插值和井震联合反演数据集进行迭代训练,从而构建高分辨率波阻抗反演标签的深度学习方法,其不需要准确的地质模型,训练难度低且经济可靠,效率高,具有重要现实意义。(The invention discloses a self-consistent depth learning method for constructing a high-resolution wave impedance inversion label, belongs to the field of deep learning seismic wave impedance inversion, and solves the problems that in the prior art, high-precision and large-capacity samples are difficult to obtain, well interpolation has high longitudinal resolution, and interpolation cannot obtain transverse features with good stratum consistency, and the invention provides the self-consistent depth learning method for constructing the high-resolution wave impedance inversion label, which aims to: the depth learning method of the high-resolution wave impedance inversion label is constructed by performing iterative training on the well interpolation and well-seismic joint inversion data set, does not need an accurate geological model, is low in training difficulty, economical and reliable, is high in efficiency, and has important practical significance.)

1. A self-consistent deep learning method for constructing a high-resolution wave impedance inversion label is characterized by comprising the following steps:

step A: constructing single well wave impedance under a variable frequency condition to form diversified forward data and logging wave impedance data;

and B: constructing wave impedance of well logging and seismic inversion under a variable frequency condition, and forming forward data with larger capacity compared with a single well and wave impedance data of well seismic joint inversion;

and C: combining the forward data and the logging wave impedance data in the step A, the forward data and the wave impedance data of the well-seismic joint inversion in the step B, and constructing an initial forward data and an initial wave impedance data set;

step D: c, performing deep learning training on the depth inversion model on the initial forward modeling data and the initial wave impedance data set in the step C to obtain a wave impedance label inversion model;

step E: d, applying the wave impedance label inversion model obtained in the step D to forward-calculated seismic data under a known geological model, completing wave impedance prediction of the seismic data under the known geological model, and obtaining a predicted wave impedance label corresponding to the seismic data under the known geological model;

step F: e, comparing the real wave impedance data set of the known geological model with the wave impedance label data set predicted in the step E, and if the real wave impedance data set of the known geological model is consistent with the wave impedance label data set predicted in the step E, ending the training; if the wave impedance tag data set is inconsistent with the real wave impedance tag data set, outputting inconsistent tags in the predicted wave impedance tag data set and the real wave impedance tag data set, and carrying out forward simulation correction on inconsistent tag samples to obtain corrected seismic data and corrected wave impedance data;

step G: and D, replacing the initial forward modeling data and the initial wave impedance data set which are constructed in the step C with the seismic data and the wave impedance data which are corrected in the step F and accord with the geological pattern, repeating the steps D to F, thereby completing the correspondence between the construction and the prediction of the wave impedance inversion label, and enabling the construction and the prediction to be self-consistent through a deep learning method.

2. The self-consistent deep learning method for constructing the high-resolution wave impedance inversion tag according to claim 1, wherein the step C further comprises comparing the wave impedance jointly inverted based on the well logging and the well seismic with the well logging and the seismic to ensure that the impedance in the longitudinal direction is consistent with that of a single well and the impedance in the transverse direction is consistent with that of a velocity model, and then enhancing the wave impedance tag sample by using a random sampling mode.

3. The self-consistent deep learning method for constructing the high-resolution wave impedance inversion label according to claim 1, wherein in the step D, the depth inversion model is a depth inversion network CNN _ inv _ D model improved according to a convolutional neural network CNN, the depth inversion model includes a convolutional layer, a pooling layer, a Dropout layer, and an activation layer, wherein the Dropout layer can be selected as an overfitting of a complex data set and improves calculation efficiency.

4. The self-consistent deep learning method for constructing the high-resolution wave impedance inversion label according to claim 1, wherein in the step D, the initial wave impedance data set is divided into a training set, a verification set and a test set, and specifically: the data of logging interpolation are divided into a training set, a verification set and a test set, and the data volume proportion of the training set, the verification set and the test set is 1: 1: 1; the data based on logging and seismic joint inversion are also divided into a training set, a verification set and a testing set, and the data volume proportion of the training set, the verification set and the testing set is 2: 1: 1; and dividing the final data of the combined logging interpolation and logging earthquake joint inversion into a training set, a verification set and a test set, wherein the data volume ratio of the training set, the verification set and the test set is 3: 2: 2.

5. the self-consistent deep learning method for constructing the high-resolution wave impedance inversion label according to claim 4, wherein in the step D, the deep learning training comprises the following steps:

step 1, training a depth inversion model on a training set to obtain an initial tag impedance inversion model;

step 2, evaluating the initial tag impedance inversion model in a verification set, and obtaining an evaluated initial tag impedance inversion model after evaluation by using the criterion that verification is performed under the condition that the loss function tends to be stable and the training error is small;

step 3, training the evaluation initial tag impedance inversion model on a training set and evaluating the evaluation initial tag impedance inversion model on a verification set in sequence, and obtaining a quasi-tag impedance inversion model after training and evaluating for a plurality of times;

and 4, evaluating the quasi-tag impedance inversion model in the test set to obtain the tag impedance inversion model after evaluation.

Technical Field

The invention belongs to the field of deep learning seismic wave impedance inversion, and particularly relates to a self-consistent deep learning method for constructing a high-resolution wave impedance inversion label.

Background

The high-resolution seismic inversion is an important technology for providing technical support for oil field exploration and reservoir development prediction by using technical means such as earthquake, well logging, geology and the like, and the production benefit can be improved by improving the reservoir prediction precision. Oil and gas production units always pursue quality improvement and efficiency enhancement, the efficiency is improved to achieve the highest yield and benefit, various data are complete and rich along with the promotion of oil field exploration and development, and the problem to be solved urgently under the background of low oil price is solved by fully utilizing various data and improving the reservoir prediction precision.

The determination of the geological model in the exploration and development of the oil field is established on the basis of well logging and earthquake, and is continuously updated according to experience and development practice, the modification is mainly to modify and perfect the geological model according to the introduction of well drilling data and development data, and further modify and perfect the inversion model, so as to reduce the multi-solution of the geological model and further reduce the multi-solution of the inversion result. However, geological knowledge in geological model construction highly depends on the experience of experts, the acquisition difficulty is high, the quality depends on known drilling logging data and seismic data, so that samples used for deep learning are limited, various unknown conditions must be faced during model training, and the acquired impedance tag is subjected to various factors such as the expert experience, geological models, logging data quality, seismic data quality and the like, so that the samples under the acquired stratum conditions are seriously unbalanced, namely the proved impedance has great inconsistency with the correlation of a predicted target, namely the tag in the predicted target does not exist in the samples, and the nonexistent part can seriously influence the effect of the training model. Due to the large sample volume requirement, for each tagged data set, it can be easily obtained by well-seismic combined wave impedance inversion.

Seismic wave impedance inversion based on deep learning often needs a large amount of correct marked data, mainly depends on drilling data of a research area, a geological model interpolated by the drilling data of the research area is insufficient to represent the real underground condition, and part of data can be polluted due to errors caused by factors such as instruments. In addition, the fitting problem is also caused by the density degree of well data, the existing deep learning model is difficult to express a data set, so that the model is poor in performance on a test set, and the problem cannot be well solved by a common label acquisition method.

Disclosure of Invention

The invention provides a self-consistent deep learning method for constructing a high-resolution wave impedance inversion label, aiming at solving the problems that in the prior art, samples with high precision and large capacity are difficult to obtain, logging interpolation has high longitudinal resolution, and interpolation cannot obtain transverse features with good consistency with a stratum, and the self-consistent deep learning method is characterized in that: iterative training is carried out on the logging interpolation and the logging and earthquake joint inversion wave impedance data set, so that a high-resolution wave impedance label is constructed, the reliability of the method is detected through deep learning, and finally a self-consistent method for label construction and deep learning detection is formed.

In order to achieve the purpose, the invention adopts the technical scheme that:

a self-consistent deep learning method for constructing a high-resolution wave impedance inversion label is characterized by comprising the following steps:

step A: constructing single well wave impedance under a variable frequency condition to form diversified forward data and logging wave impedance data;

and B: constructing wave impedance of well logging and seismic inversion under a variable frequency condition, and forming forward data with larger capacity compared with a single well and wave impedance data of well seismic joint inversion;

and C: combining the forward data and the logging wave impedance data in the step A, the forward data and the wave impedance data of the well-seismic joint inversion in the step B, and constructing an initial forward data and an initial wave impedance data set;

step D: c, performing deep learning training on the depth inversion model on the initial forward modeling data and the initial wave impedance data set in the step C to obtain a wave impedance label inversion model;

step E: d, applying the wave impedance label inversion model obtained in the step D to forward-calculated seismic data under a known geological model, completing wave impedance prediction of the seismic data under the known geological model, and obtaining a predicted wave impedance label corresponding to the seismic data under the known geological model;

step F: e, comparing the real wave impedance data set of the known geological model with the wave impedance label data set predicted in the step E, and if the real wave impedance data set of the known geological model is consistent with the wave impedance label data set predicted in the step E, ending the training; if the wave impedance tag data set is inconsistent with the real wave impedance tag data set, outputting inconsistent tags in the predicted wave impedance tag data set and the real wave impedance tag data set, and carrying out forward simulation correction on inconsistent tag samples to obtain corrected seismic data and corrected wave impedance data;

step G: and D, replacing the initial forward modeling data and the initial wave impedance data set which are constructed in the step C with the seismic data and the wave impedance data which are corrected in the step F and accord with the geological pattern, repeating the steps D to F, thereby completing the correspondence between the construction and the prediction of the wave impedance inversion label, and enabling the construction and the prediction to be self-consistent through a deep learning method.

Further, the step C also comprises the step of comparing the wave impedance jointly inverted by the well logging-based wave impedance and the well earthquake with the well logging and the earthquake to ensure that the impedance in the longitudinal direction is consistent with that of a single well and the transverse direction has better consistency with the velocity model, and then enhancing the wave impedance label sample by using a random sampling mode.

Further, in the step D, the depth inversion model is a depth inversion network CNN _ inv _ D model improved according to the convolutional neural network CNN, and the depth inversion model includes a convolutional layer, a pooling layer, a Dropout layer, and an activation layer, where the Dropout layer may be selected as an overfitting of the complex data set and improves the calculation efficiency.

Further, in the step D, the initial wave impedance data set is divided into a training set, a verification set, and a test set, specifically: the data of logging interpolation are divided into a training set, a verification set and a test set, and the data volume proportion of the training set, the verification set and the test set is 1: 1: 1; the data based on logging and seismic joint inversion are also divided into a training set, a verification set and a testing set, and the data volume proportion of the training set, the verification set and the testing set is 2: 1: 1; and dividing the final data of the combined logging interpolation and logging earthquake joint inversion into a training set, a verification set and a test set, wherein the data volume ratio of the training set, the verification set and the test set is 3: 2: 2.

further, in the step D, the deep learning training includes the following steps:

step 1, training a depth inversion model on a training set to obtain an initial tag impedance inversion model;

step 2, evaluating the initial tag impedance inversion model in a verification set, and obtaining an evaluated initial tag impedance inversion model after evaluation by using the criterion that verification is performed under the condition that the loss function tends to be stable and the training error is small;

step 3, training the evaluation initial tag impedance inversion model on a training set and evaluating the evaluation initial tag impedance inversion model on a verification set in sequence, and obtaining a quasi-tag impedance inversion model after training and evaluating for a plurality of times;

and 4, evaluating the quasi-tag impedance inversion model in the test set to obtain the tag impedance inversion model after evaluation.

Compared with the prior art, the invention has the beneficial effects that:

the deep learning method for the high-resolution wave impedance inversion label is constructed by performing iterative training on the well interpolation and well-seismic joint inversion data set, does not need an accurate geological model, is low in training difficulty, economical and reliable, and high in efficiency, and has important practical significance.

Logging interpolation can obtain wave impedance labels of the stratum, and a large amount of wave impedance label data are obtained through variable frequency wavelets on the basis, namely, new labels are artificially added and cover the wave impedance labels in the predicted target. The diversity of the wave impedance tags obtained by logging interpolation is limited by the influence of human factors, and the wave impedance tags which really exist as far as possible under the condition of conforming to the stratum can be constructed by combining the seismic data, so that a large number of wave impedance tags irrelevant to the stratum are avoided. Therefore, under the constraint of the frequency conversion synthetic record, the rationality of the newly added wave impedance label can be ensured by comparing the difference with the real seismic record according to the principle of small difference, and a large-capacity sample is provided for seismic wave impedance inversion based on deep learning.

Drawings

Fig. 1 is a schematic diagram of an inversion process based on CNN _ inv _ d wave impedance.

FIG. 2 is a continental facies sand shale wave impedance geological model in an embodiment of the invention.

FIG. 3 is a diagram of pseudo-well interpolation wave impedance in an embodiment of the present invention.

FIG. 4 is a graph of inversion wave impedance of seismic logging in an embodiment of the invention.

Fig. 5 is a wave impedance inversion based on CNN _ inv _ d deep learning in an embodiment of the present invention.

Fig. 6 is a graph showing a change in loss value in the embodiment of the present invention.

Detailed Description

The present invention will be described in detail with reference to the following examples, which are provided for the purpose of illustrating the present invention and are not to be construed as limiting the scope of the present invention.

A self-consistent deep learning method for constructing a high-resolution wave impedance inversion label is characterized by comprising the following steps:

step A: constructing single well wave impedance under a variable frequency condition to form diversified forward data and logging wave impedance data;

and B: constructing wave impedance of well logging and seismic inversion under a variable frequency condition, and forming forward data with larger capacity compared with a single well and wave impedance data of well seismic joint inversion;

and C: combining the forward data and the logging wave impedance data in the step A, the forward data and the wave impedance data of the well-seismic joint inversion in the step B, and constructing an initial forward data and an initial wave impedance data set;

step D: c, performing deep learning training on the depth inversion model on the initial forward modeling data and the initial wave impedance data set in the step C to obtain a wave impedance label inversion model;

step E: d, applying the wave impedance label inversion model obtained in the step D to forward-calculated seismic data under a known geological model, completing wave impedance prediction of the seismic data under the known geological model, and obtaining a predicted wave impedance label corresponding to the seismic data under the known geological model;

step F: e, comparing the real wave impedance data set of the known geological model with the wave impedance label data set predicted in the step E, and if the real wave impedance data set of the known geological model is consistent with the wave impedance label data set predicted in the step E, ending the training; if the wave impedance tag data set is inconsistent with the real wave impedance tag data set, outputting inconsistent tags in the predicted wave impedance tag data set and the real wave impedance tag data set, and carrying out forward simulation correction on inconsistent tag samples to obtain corrected seismic data and corrected wave impedance data;

step G: and D, replacing the initial forward modeling data and the initial wave impedance data set which are constructed in the step C with the seismic data and the wave impedance data which are corrected in the step F and accord with the geological pattern, repeating the steps D to F, thereby completing the correspondence between the construction and the prediction of the wave impedance inversion label, and enabling the construction and the prediction to be self-consistent through a deep learning method.

And C, comparing the wave impedance jointly inverted by the well logging-based wave impedance and the well earthquake with the well logging and the earthquake to ensure that the impedance in the longitudinal direction is consistent with that of a single well and has better consistency with the velocity model in the transverse direction, and then enhancing the wave impedance label sample by using a random sampling mode.

In the step D, the depth inversion model is a depth inversion network CNN _ inv _ D model formed by improving a convolutional neural network CNN, and the depth inversion model includes a convolution layer, a pooling layer, a Dropout layer, and an activation layer, where the Dropout layer may be selected as an overfitting of a complex data set and improves calculation efficiency.

In the step D, the initial wave impedance data set is divided into a training set, a verification set and a test set, and the specific method is as follows: the data of logging interpolation are divided into a training set, a verification set and a test set, and the data volume proportion of the training set, the verification set and the test set is 1: 1: 1; the data based on logging and seismic joint inversion are also divided into a training set, a verification set and a testing set, and the data volume proportion of the training set, the verification set and the testing set is 2: 1: 1; and dividing the final data of the combined logging interpolation and logging earthquake joint inversion into a training set, a verification set and a test set, wherein the data volume ratio of the training set, the verification set and the test set is 3: 2: 2.

in the step D, the deep learning training includes the following steps:

step 1, training a depth inversion model on a training set to obtain an initial tag impedance inversion model;

step 2, evaluating the initial tag impedance inversion model in a verification set, and obtaining an evaluated initial tag impedance inversion model after evaluation by using the criterion that verification is performed under the condition that the loss function tends to be stable and the training error is small;

step 3, training the evaluation initial tag impedance inversion model on a training set and evaluating the evaluation initial tag impedance inversion model on a verification set in sequence, and obtaining a quasi-tag impedance inversion model after training and evaluating for a plurality of times;

and 4, evaluating the quasi-tag impedance inversion model in the test set to obtain the tag impedance inversion model after evaluation.

The following are specific examples.

A self-consistent deep learning method for constructing a high-resolution wave impedance inversion label comprises the following stages:

1. precondition:

the data set is marked as sedimentary rock, metamorphic rock and volcanic rock, and mainly aims at the stratum with a better stratum in the rock stratum, the specific implementation process is shown in figure 1, and the geological model is shown in figure 2.

The data set is generated as a single well single pass computed wave impedance, and seismic logging inverted wave impedance.

2. A model training stage:

building a neural network on a data set of single-well single-channel calculated wave impedance and seismic logging joint inversion wave impedance, determining a loss function, training parameters by using a gradient descent technology to obtain a group of model parameters which enable the loss function to be relatively small, and finishing preliminary training of a model;

predicting a part of seismic logging joint inversion wave impedance data training set by using a model trained for the first time;

and comparing the wave impedance label predicted by the model with a real label, calibrating and screening the label which is not in accordance with the real label according to the earthquake synthetic record, and re-developing earthquake inversion to realize well earthquake matching.

3. Model iteration stage:

and training the model on the data set after the marking is carried out again, and continuously repeating the model training stage until the model prediction result is basically consistent with the actual data.

Constructing wave impedance of logging interpolation and wave impedance of seismic logging joint inversion as a data set, wherein the data set is characterized in that: the data sets are logging interpolation wave impedance (figure 3) and logging and earthquake joint inversion wave impedance (figure 4) under the condition of a specific geological model; the method comprises the following steps that wrong labels exist among wells in the logging interpolation wave impedance, namely certain wrong labels are mixed in the logging interpolation wave impedance labels, and labels of the data are used as training labels under the specific geological model condition; the precision of the wave impedance label of seismic logging inversion depends on an inversion geological model, and a certain error label can be generated due to the error of the constructed inversion geological model and the specific geological condition, namely a certain error label is mixed in the seismic wave impedance inversion label; the initial wave impedance dataset is randomly divided into a training set, a validation set, and a test set, wherein the training set is much larger in size than the validation set and the test set.

Before training, further processing data, and processing the problem of sample imbalance by using a sampling and data synthesis method, wherein the specific method comprises the following steps: performing data enhancement on the wave impedance dataset using random sampling; and further, more samples are generated by using the method for changing the phase and the frequency of the seismic wavelet in the step A and the step B, so that the application effect of a small data scene is improved.

Constructing a depth inversion model by using a CNN network architecture mainly comprising a convolutional layer, a Dropout, a pooling layer and an activation function structure; depth inversion networks CNN _ inv and CNN _ inv _ d formed by improving the CNN are two deformation forms in a Convolutional Neural Network (CNN), and the wave impedance effects of the CNN _ inv and the CNN _ inv _ d on the data sets generated in the steps A and B are compared, so that the accuracy of the CNN _ inv _ d is improved compared with that of other neural networks. CNN _ inv _ d is therefore selected as the final depth inversion model.

The general flow of the CNN _ inv _ d training model is as follows:

a. and (6) inputting a model. CNN _ inv _ d is a method for deep learning geophysical inversion, where the model input is a numerical matrix and the model input is M × N. Where M x N are the dataset rows and columns.

b model structure. The basic CNN mainly consists of 4 elements, which are convolutional layers, pooling layers, activation functions, and full-link layers, respectively. And CNN _ inv _ d is an inversion of deep learning geophysical inversion and mainly comprises 4 elements, namely a convolutional layer, a pooling layer, Dropout and an activation function, wherein Dropout can be selected as overfitting of a complex data set and improves the computational efficiency.

The purpose of the convolutional layer is to extract data features, which contain relative position information in addition to the feature values themselves. The purpose of the pooling layer is to compress convolutional layer information, reducing the number of feature-mapped neural elements to reduce the data dimension. Dropout mainly aims to improve the generalization capability of a model, force a neuron to work together with other randomly selected neurons, weaken and reduce joint adaptability among neuron nodes, and prevent the synergistic effect of certain characteristics so that the model does not depend on certain local characteristics too much, thereby enhancing the robustness of the model. The main purpose of the activation function is to construct a non-linear mapping relationship to establish the relationship between the input and the output. The CNN _ inv _ d inherits the CNN characteristic extraction advantages, nonlinear mapping of data and labels is constructed through an activation function, a trained CNN _ inv _ d inversion model is reapplied to an original training set and label prediction is carried out, and label inversion is carried out on samples on the training set. The specific method of prediction is to use the CNN _ inv _ d inverse model to obtain various parameters, including loss values (fig. 6), weights and biases.

And comparing the inverted label result with the original label, if the inverted label result is inconsistent with the original label result, outputting a sample inconsistent with the original label to carry out manual inspection, and if the label result is definitely wrong, correcting the sample label with the wrong label. If the two are consistent, the training is ended. And (5) obtaining a final model after finishing training, and inputting the model by using data to obtain an inversion result (figure 5).

The above-mentioned embodiments only express the specific embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, without departing from the technical idea of the present application, several changes and modifications can be made, which are all within the protection scope of the present application.

Details not described in this specification are within the skill of the art that are well known to those skilled in the art.