Deep learning wave impedance inversion method and system based on virtual well

Abstract

The invention relates to the technical field of geophysical exploration, in particular to a deep learning wave impedance inversion method and system based on a virtual well, and the method achieves the high-precision inversion from seismic data to wave impedance parameters through building the statistical relation between lithology and wave impedance and combining the powerful nonlinear mapping capability of a neural network. And a reliable technical means is provided for underground reservoir prediction. A large amount of virtual well data is generated based on a small amount of real wave impedance real well data, so that deep learning network training is facilitated; the parameters do not need to be greatly adjusted; according to the method, a large amount of virtual well data is generated, the trained model has strong generalization ability and an accurate inversion effect, complex parameter adjustment is not needed in the training process, and the method has high applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of geophysical exploration, and particularly to a deep learning wave impedance inversion method and system based on virtual wells. Background Art

[0002] Seismic exploration is a technical means that uses artificial excitation of seismic waves and analyzes their propagation characteristics underground to infer geological structures. By studying parameters such as the propagation velocity, reflection, and transmission characteristics of seismic waves in different rock layers, important geological information can be obtained. Among them, seismic impedance, as a core parameter, is composed of the product of rock density and longitudinal wave velocity, and can effectively characterize key properties such as the porosity, fluid saturation, and lithology of rocks. These parameters are of great significance for oil and gas reservoir identification.

[0003] Against the backdrop of the continuously rising global energy demand, traditional oil and gas resources are becoming increasingly depleted, and exploration and development face greater challenges. This requires exploration technologies to develop towards higher precision and efficiency. Seismic impedance inversion technology can generate high-resolution underground imaging, providing geologists with accurate information such as the boundaries of oil and gas reservoirs, reservoir physical properties, and oil and gas-bearing properties, thereby significantly improving the exploration success rate and reducing drilling risks and costs. The essence of this technology is to convert seismic data into parameters that can directly reflect rock physical properties, and this conversion process is a key link in reservoir evaluation and reservoir characterization. As a bridge connecting the propagation characteristics of seismic waves and the geological properties of rocks, wave impedance parameters can be used to estimate key reservoir parameters such as formation porosity and permeability, which has important guiding value for production capacity assessment and development plan formulation.

[0004] Currently, various technical routes have been developed for wave impedance inversion methods, including trace integration inversion, generalized linear inversion, iterative inversion, and non-linear inversion, etc. Each method has its own characteristics: trace integration inversion is simple to operate but has limited accuracy; generalized linear inversion has higher accuracy but is sensitive to high-frequency noise. It is worth noting that model-based inversion technology can break through the limitations of traditional seismic resolution and theoretically obtain a resolution comparable to well logging data. However, this method has an obvious problem of non-uniqueness. Its inversion results are greatly affected by well logging information and the high and low-frequency components provided by the model, and are limited by the number and distribution range of well positions. A large amount of data sets are required to train the model to achieve a better inversion effect, while the real well data obtained from actual exploration cannot support the training process of the model.

[0005] Generally speaking, traditional wave impedance inversion methods have the following defects: 1. Inversion methods based on deep learning networks require a large amount of well data, which is difficult to meet the needs of actual exploration; 2. Conventional statistical inversion methods rely on accurate initial models and complex parameter adjustments; 3. Existing machine learning methods are limited by the quality of training data and have insufficient generalization ability. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a deep learning wave impedance inversion method and system based on virtual wells to solve the problems in the background technology.

[0007] To achieve the above purpose, the present invention adopts the following technical solutions:

[0008] A deep learning wave impedance inversion method based on virtual wells of the present invention includes the steps of:

[0009] Construct a formation lithology model for simulating the physical forms of multiple formations, and add initial wave impedance values and Gaussian noise to the formation lithology model based on lithology classification to obtain an initial wave impedance model;

[0010] Extract multiple wells from the initial wave impedance model to obtain multiple well data; classify the lithology of the multiple well data to obtain wave impedance data of multiple lithology categories; and calculate the distribution characteristics of the wave impedance data of the multiple lithology categories, wherein the well data includes the wave impedance of multiple formations in the well;

[0011] Construct multiple virtual well wave impedance models based on the distribution characteristics of the wave impedance data of multiple lithology categories, wherein the virtual well wave impedance model includes the wave impedance conforming to the distribution characteristics of multiple formations;

[0012] Extract the low-frequency smooth background of the multiple virtual well wave impedance models;

[0013] Calculate the reflection coefficient sequence of the virtual well wave impedance model, construct a seismic record based on the reflection coefficient sequence; and fuse the seismic record and the low-frequency smooth background to obtain a seismic data pair sample;

[0014] Label the seismic data sample based on the true wave impedance data to obtain training data; train an artificial neural network based on the training data to obtain an inversion model; and perform wave impedance inversion based on the inversion model.

[0015] The present application also provides a deep learning wave impedance inversion system based on virtual wells, and the system includes:

[0016] A model construction module for constructing a formation lithology model for simulating the physical forms of multiple formations, and adding initial wave impedance values and Gaussian noise to the formation lithology model based on lithology classification to obtain an initial wave impedance model;

[0017] A feature extraction module, configured to extract multiple well traces from the initial wave impedance model to obtain multiple well data; classify the lithology of the multiple well data to obtain wave impedance data of multiple lithology categories; and calculate the distribution characteristics of the wave impedance data of the multiple lithology categories, where the well data includes the wave impedance of multiple formations in the well.

[0018] A data generation module, configured to construct multiple virtual well wave impedance models based on the distribution characteristics of the wave impedance data of multiple lithology categories, where the virtual well wave impedance model includes the wave impedance conforming to the distribution characteristics of multiple formations.

[0019] A background extraction module, configured to extract the low-frequency smooth background of the multiple virtual well wave impedance models.

[0020] A sample generation module, configured to calculate the reflection coefficient sequence of the virtual well wave impedance model, construct a seismic record based on the reflection coefficient sequence; and fuse the seismic record and the low-frequency smooth background to obtain a seismic data pair sample.

[0021] An inversion module, configured to label the seismic data sample based on the true wave impedance data to obtain training data; train an artificial neural network based on the training data to obtain an inversion model; and perform wave impedance inversion based on the inversion model.

[0022] The beneficial effects of the present invention are as follows: A deep learning wave impedance inversion method and system based on virtual wells of the present invention. By establishing the statistical relationship between lithology and wave impedance and combining the powerful non-linear mapping ability of the neural network, the high-precision inversion from seismic data to wave impedance parameters is realized, providing a reliable technical means for underground reservoir prediction. The present application generates a large number of virtual well data based on a small amount of true wave impedance true well data, which is convenient for the training of the deep learning network; and there is no need to adjust a large number of parameters; by generating a large number of virtual well data, the trained model has a powerful generalization ability and an accurate inversion effect, and the training process does not require complex parameter tuning, having strong applicability. Description of the Drawings

[0023] The present invention will be further described below with reference to the drawings and embodiments:

[0024] Figure 1 is a flowchart of a deep learning wave impedance inversion method based on virtual wells shown in an embodiment of the present application;

[0025] Figure 2 is a schematic diagram of obtaining well data in an embodiment of the present application;

[0026] Figure 3 is a schematic diagram of the wave impedance distribution characteristics in an embodiment of the present application;

[0027] Figure 4 It is a schematic diagram of the wave impedance parameter model of the multi-channel virtual well in the embodiment of the present invention;

[0028] Figure 5 It is a comparison chart of the initial wave impedance model and the predicted wave impedance model in the embodiment of the present invention;

[0029] Figure 6 It is a structural diagram of a deep learning wave impedance inversion system based on virtual wells shown in an embodiment of the present application. Detailed implementation manners

[0030] Aiming at the instabilities, sparse well data, low precision and other defects and problems existing in the traditional wave impedance inversion method, the purpose of the present invention is to provide a deep learning wave impedance inversion method based on virtual wells to achieve high-precision inversion from seismic data to wave impedance parameters.

[0031] The present invention is proposed based on the study of the following problems:

[0032] 1. The inversion method based on the deep learning network requires a large amount of well data, which is difficult to meet the requirements in actual exploration;

[0033] 2. The conventional statistical inversion method depends on an accurate initial model and complex parameter adjustment;

[0034] 3. The existing machine learning methods are limited by the quality of training data and have insufficient generalization ability;

[0035] 4. It is difficult for a single network structure to capture the spatial features and time series features of seismic signals simultaneously.

[0036] Therefore, the present invention proposes a seismic wave impedance inversion method based on virtual wells and neural networks. First, by initializing geological parameters, a formation fold structure is simulated using a sine function to establish a framework for a formation lithology model. In this model, corresponding noisy wave impedance data is filled according to different lithology categories to form an initial wave impedance model. Lithology data in the model and its corresponding initial wave impedance model are randomly selected and classified and statistically analyzed according to lithology categories. For each type of lithology, the mean and variance of its wave impedance data are calculated, thereby establishing a Gaussian probability distribution model for the wave impedance corresponding to each type of lithology. Based on the above Gaussian distribution characteristics, a multi-trace statistically independent virtual well wave impedance model is simulated. Subsequently, the initial wave impedance model and the virtual well wave impedance model are normalized, and a corresponding low-frequency smoothed background field of wave impedance is generated. A reflection coefficient sequence is calculated from the wave impedance data, and by performing convolution operations with a standard Ricker wavelet, a synthetic seismic record is simulated. The synthetic seismic record and the low-frequency wave impedance background field are data-fused to generate data pairs convenient for neural network training. The virtual well data and its corresponding synthetic seismic data are used as training samples and input into the neural network for model training. The trained network model is used to perform inversion calculations on actual seismic data, and finally, a prediction result with a controllable error from the true wave impedance data is obtained.

[0037] Figure 1 is a flowchart of a deep learning wave impedance inversion method based on virtual wells shown in an embodiment of the present application, as Figure 1 shown: A deep learning wave impedance inversion method based on virtual wells in this embodiment includes steps S110 to S160:

[0038] S110, construct a formation lithology model for simulating the physical forms of multiple formations, and add initial wave impedance values and Gaussian noise to the formation lithology model based on lithology classification to obtain an initial wave impedance model;

[0039] In the technical process of seismic wave impedance inversion, first, the geometric and physical properties of the model are determined through parameter initialization, including the number of horizontal traces, vertical sampling interval, and lithology classification.

[0040] In the present application, based on a sine function model to simulate the formation fold structure, the construction process of the formation lithology model includes:

[0041] S1101, obtain a formation index L_index, a randomly generated amplitude A, and a randomly generated wavelength λ;

[0042] Among them, the formation index L_index is a set of standardized naming and numbering systems used in geology for systematic classification and identification of different rock layers or formation units, and is used to distinguish different formations.

[0043] S1102. Construct a sine function model based on the formation index, the randomly generated amplitude, and the randomly generated wavelength to obtain a formation lithology model that uses a sine wave to simulate the physical morphology of the formation.

[0044] The mathematical expression of the formation lithology model is:

[0045]

[0046]

[0047] is the mean value of the impedance corresponding to different lithologies obtained from actual data.

[0048] Simulate the formation fold structure based on the sine function model. Its core idea is to use a periodic undulation function to depict the bending deformation of the formation. The amplitude and wavelength parameters are randomly generated through a normal distribution, endowing the model with irregular geological features. The randomness of the amplitude reflects the intensity difference of tectonic movements, while the change in wavelength simulates the extension characteristics of folds in space.

[0049] Subsequently, assign initial wave impedance values according to the preset lithology classification, and simulate the heterogeneity of the actual geology by adding Gaussian noise. The noise intensity is adjusted according to the lithology category, so that the wave impedance data shows reasonable statistical fluctuations within the same lithology.

[0050] The process of adding Gaussian noise to form the initial wave impedance model includes:

[0051] S1111. Obtain the initial wave impedance values and wave impedance standard deviations corresponding to different lithology classifications;

[0052] S1112. Add the corresponding initial wave impedance values and Gaussian noise to different lithology formations in the formation lithology model to obtain an initial wave impedance model.

[0053] Different Gaussian noises are added corresponding to the wave impedance values of each lithology, and the noise is customized within a reasonable range. Among them, the mathematical expression of the initial wave impedance model is:

[0054]

[0055] In the formula, I(s) represents the initial wave impedance model data, I base (s) is the initial wave impedance value, represents Gaussian noise, δ s represents the relevant parameter standard deviation.

[0056] S120. Extract multiple well data from the initial wave impedance model, obtaining multiple sets of well data; classify the lithology of the multiple sets of well data to obtain wave impedance data for multiple lithology categories; and calculate the distribution characteristics of the wave impedance data for the multiple lithology categories. Herein, the well data includes the wave impedance of multiple strata within the well.

[0057] In the statistical modeling stage, randomly select four sets of data from the initial model, and extract the lithology labels and the corresponding wave impedance values. Figure 2 This is a schematic diagram of obtaining well data in an embodiment of the present application. For example, in Figure 2 , the left figure is the initial lithology model, and the right figure is the initial wave impedance model. The vertical line segments represent the well-taking positions.

[0058] After grouping by lithology category, use the mean-variance statistical model to calculate the wave impedance distribution characteristics of each lithology category. Figure 3 This is a schematic diagram of the wave impedance distribution characteristics in an embodiment of the present application. The distribution characteristics of each lithology are obtained as shown in Figure 3 .

[0059] The mean reflects the typical wave impedance value of the lithology, and the variance describes the degree of dispersion of the data. This process is achieved through arithmetic mean and sample variance calculations, and the results define the Gaussian probability distribution parameters for each lithology category, providing strict statistical constraints for subsequent data generation.

[0060] Among them, the calculation formulas for the mean and variance are as follows:

[0061]

[0062] In the formula, υ represents the mean wave impedance corresponding to the lithology, x i represents the wave impedance data, and N represents the number of wave impedance data.

[0063]

[0064] In the formula, δ 2 represents the variance of the wave impedance corresponding to the lithology.

[0065] S130. Construct multiple virtual well wave impedance models based on the distribution characteristics of the wave impedance data for multiple lithology categories. Herein, the virtual well wave impedance model includes the wave impedance conforming to the distribution characteristics of multiple strata.

[0066] Simulate and generate multiple uncorrelated virtual well wave impedance models according to the distribution of the wave impedance data corresponding to each lithology obtained previously.

[0067] In the present application, based on the distribution characteristics extracted above, use the multivariate normal distribution function to randomly generate virtual well wave impedance models. The process includes:

[0068] S131. Generate multiple virtual wells, where each virtual well includes multiple formation units with random thicknesses, and the lithology of each formation unit is determined by random sampling.

[0069] S132. Generate a multivariate normal distribution function based on the mean and variance of the wave impedance data of multiple lithology categories, and generate random wave impedance values for each formation unit based on the multivariate normal distribution function to obtain multiple virtual well wave impedance models.

[0070] Each virtual well is divided into multiple formation units with random thicknesses, and the lithology type is determined by Monte Carlo random sampling to ensure that the spatial combination of different lithology layers conforms to the geological stratification logic. The wave impedance values strictly follow the statistical distribution law of the previous step. Finally, multiple virtual well wave impedance models are generated. Figure 4 It is a schematic diagram of the wave impedance parameter model of multiple virtual wells in an embodiment of the present invention. The generated multiple virtual well wave impedance models are as Figure 4 shown.

[0071] In this embodiment, the wave impedance values in the virtual well wave impedance model are implemented through a multivariate normal distribution function (such as the mvnrnd function). The principle of the multivariate normal distribution function is to generate independent Gaussian distribution samples by using the Cholesky decomposition of the covariance matrix, ensuring that the wave impedance data of different formation units are statistically independent of each other and consistent with the distribution characteristics of the real data.

[0072] The mathematical expression of the multivariate normal distribution function in this application is:

[0073] Z~Ν(υ,Σ)(5)

[0074] In the formula, υ represents the wave impedance mean corresponding to the lithology, Σ represents the covariance matrix of the wave impedance corresponding to the lithology, and Z represents the generated random wave impedance value.

[0075]

[0076] In the formula, f(x) represents the probability density function, and k represents the dimension of the variable.

[0077] S140. Extract the low-frequency smooth background of the multiple virtual well wave impedance models, including:

[0078] S141. Normalize the initial wave impedance model and the virtual well wave impedance models to obtain normalized data.

[0079] S142. Extract the low-frequency smooth background in the normalized data based on the locally weighted regression smoothing algorithm.

[0080] Normalize the initial wave impedance model and the random virtual well wave impedance model, and generate the corresponding low-frequency smoothed background of wave impedance: First, perform linear normalization on the initial wave impedance and virtual well data. The mathematical expression for normalization is:

[0081]

[0082] In the formula, Z represents the normalized data, Z min represents the minimum wave impedance, and Z max represents the maximum wave impedance.

[0083] Through the normalization operation, the original numerical values are scaled to a reasonable range in a fixed ratio, eliminating the influence of dimensional differences on subsequent calculations. Subsequently, the locally weighted regression (LOWESS) smoothing algorithm is used to extract the low-frequency background field of wave impedance.

[0084] In this implementation, based on the locally weighted regression (LOWESS) smoothing algorithm, the low-frequency background field of wave impedance is extracted, and the mathematical expression for the smoothed value is:

[0085]

[0086] In the formula, w i represents the i-th weight, h represents the bandwidth, which is used to control the size of the smoothing window, β0 and β1 represent the intercept and slope of the regression coefficient, represents the smoothed value, x i represents the input variable, y i represents the smoothed data, and x c represents the current smoothing point.

[0087] The above algorithm fits the local trend through weighted least squares within the sliding window. The weight function uses a cubic kernel function, and the data closer to the current point has a higher weight, thereby generating the low-frequency smoothed background of wave impedance. The window width is set so that the local formation trend can be highlighted, while thin layer interference or random noise is effectively suppressed.

[0088] S150. Calculate the reflection coefficient sequence of the virtual well wave impedance model, construct a seismic record based on the reflection coefficient sequence; and fuse the seismic record and the low-frequency smoothed background to obtain a seismic data pair sample;

[0089] The reflection coefficient between formations refers to the ratio of the amplitude of the reflected wave to the amplitude of the incident wave when the seismic wave passes through the formation interface. In this embodiment, the reflection coefficient is calculated from the wave impedance data, convolved with the Ricker wavelet to obtain the corresponding synthetic seismic record, and the seismic record and the low-frequency smoothed background of wave impedance are integrated to construct a training data pair (sample data). The specific process includes:

[0090] S151. Calculate the reflection coefficient of adjacent formation units based on the wave impedance of adjacent formation units in the virtual well wave impedance model;

[0091] S152. Construct a reflection coefficient sequence based on the reflection coefficients of multiple adjacent formation units.

[0092] In the seismic data synthesis stage, first calculate the reflection coefficient sequence based on the impedance difference model. Its physical principle is that seismic waves are reflected at the interface where the wave impedance changes abruptly, and the value of the reflection coefficient is proportional to the wave impedance difference between adjacent layers. Among them, the mathematical expression of the reflection coefficient sequence is:

[0093]

[0094] In the formula, R i represents the reflection coefficient, Z i+1 represents the (i + 1)-th wave impedance data, and Z i represents the i-th wave impedance data.

[0095] S153. Multiply the reflection coefficient sequence by the Toeplitz matrix formed by the Ricker wavelet to obtain the seismic record.

[0096] Subsequently, simulate the propagation characteristics of the seismic wavelet through the Ricker wavelet model. The Ricker wavelet is a zero-phase wavelet, and its waveform is controlled by the dominant frequency parameter. The waveform of the wavelet has a symmetric main lobe and gradually decaying side lobes, which can accurately depict the energy distribution of seismic waves. Convolve the reflection coefficient sequence with the Ricker wavelet to generate the synthetic seismic record. Convert the convolution into matrix multiplication by constructing a Toeplitz matrix to achieve efficient calculation. Finally, fuse the synthetic seismic record with the low-frequency background field, and enhance the mutation response of the formation interface through the weighted superposition strategy to construct a data pair convenient for neural network training. The mathematical expression of the seismic data is:

[0097] S_record = M T *R(12)

[0098] In the formula, S_record represents the seismic record, M T is the Toeplitz matrix constructed by the Ricker wavelet, and R is the reflection coefficient.

[0099] S160. Annotate the seismic data samples based on the true wave impedance data to obtain training data; train the artificial neural network based on the training data to obtain an inversion model; and perform wave impedance inversion based on the inversion model.

[0100] Substitute the virtual well data and the corresponding seismic data into the neural network training model, and use the trained model to invert the seismic data corresponding to the initial wave impedance model to obtain a predicted wave impedance model with a sufficiently small error from the true initial wave impedance model: In the inversion stage, a convolutional neural network is used to establish a non-linear mapping relationship between the seismic data and the wave impedance. Using the virtual well seismic data as the input and the corresponding wave impedance data as the label, the mean square error loss function is used to quantify the deviation between the predicted value and the true value. During the network training process, the backpropagation algorithm is used to calculate the gradient, and the Adam optimizer is combined to dynamically adjust the learning rate to gradually optimize the network weights. The trained model performs inversion on the seismic data corresponding to the initial wave impedance, and corrects the prediction results through multiple rounds of iteration until the error between the predicted wave impedance and the true data is sufficiently small (such as Figure 5 shown), ensuring that the inversion results meet the geological interpretation requirements in terms of spatial distribution and numerical accuracy. Input the training data into the artificial neural network model to obtain wave impedance prediction data;

[0101] During training, calculate the deviation between the wave impedance prediction data and the label based on the mean square error loss function, and adjust the network weights of the artificial neural network based on the deviation; go back to inputting the training data into the artificial neural network model to obtain wave impedance prediction data until the deviation between the wave impedance prediction data and the label is less than a preset deviation threshold to obtain an inversion model.

[0102] A deep learning wave impedance inversion method based on virtual wells according to the present invention. Through the establishment of a statistical relationship between lithology and wave impedance and the combination of the powerful non-linear mapping ability of the neural network, the present application realizes high-precision inversion from seismic data to wave impedance parameters, providing a reliable technical means for underground reservoir prediction. The present application generates a large number of virtual well data based on a small amount of real wave impedance real well data, facilitating the training of the deep learning network; and there is no need to make a large number of parameter adjustments; through the generation of a large number of virtual well data, the trained model has strong generalization ability and accurate inversion effect, and the training process does not require complex parameter tuning, having strong applicability.

[0103] As Figure 6 shown, the present application also provides a deep learning wave impedance inversion system based on virtual wells, including:

[0104] A model construction module, configured to construct a formation lithology model for simulating the physical forms of multiple formations, and add an initial wave impedance value and Gaussian noise to the formation lithology model based on lithology classification to obtain an initial wave impedance model;

[0105] A feature extraction module, configured to extract multiple well traces from the initial wave impedance model to obtain a plurality of well data; classify the lithology of the plurality of well data to obtain wave impedance data of multiple lithology categories; and calculate the distribution characteristics of the wave impedance data of multiple lithology categories, where the well data includes the wave impedance of multiple formations in the well.

[0106] A data generation module, configured to construct a plurality of virtual well wave impedance models based on the distribution characteristics of the wave impedance data of multiple lithology categories, where the virtual well wave impedance model includes the wave impedance conforming to the distribution characteristics of multiple formations.

[0107] A background extraction module, configured to extract the low-frequency smooth background of the plurality of virtual well wave impedance models.

[0108] A sample generation module, configured to calculate the reflection coefficient sequence of the virtual well wave impedance model, construct a seismic record based on the reflection coefficient sequence; and fuse the seismic record and the low-frequency smooth background to obtain a seismic data pair sample.

[0109] An inversion module, configured to label the seismic data samples based on the real wave impedance data to obtain training data; train an artificial neural network based on the training data to obtain an inversion model; and perform wave impedance inversion based on the inversion model.

[0110] A deep learning wave impedance inversion system based on virtual wells according to the present invention. By establishing a statistical relationship between lithology and wave impedance and combining the powerful non-linear mapping ability of the neural network, the present application realizes high-precision inversion from seismic data to wave impedance parameters, providing a reliable technical means for underground reservoir prediction. The present application generates a large amount of virtual well data based on a small amount of real wave impedance real well data, facilitating the training of the deep learning network; and there is no need to adjust a large number of parameters; by generating a large amount of virtual well data, the trained model has a powerful generalization ability and an accurate inversion effect, and the training process does not require complex parameter tuning, having strong applicability.

Claims

1. A deep learning wave impedance inversion method based on virtual wells, characterized in that, Including the steps: Construct a formation lithology model for simulating the physical morphology of multiple formations, and add initial impedance values and Gaussian noise to the formation lithology model based on lithology classification to obtain an initial impedance model; Extract multiple well traces from the initial impedance model to obtain multiple well data; classify the lithology of the multiple well data to obtain impedance data of multiple lithology categories; and calculate the distribution characteristics of the impedance data of multiple lithology categories, where the well data includes the impedance of multiple formations in the well; Construct multiple virtual well impedance models based on the distribution characteristics of the impedance data of multiple lithology categories, where the virtual well impedance model includes the impedance conforming to the distribution characteristics of multiple formations; Extract the low-frequency smooth background of the multiple virtual well impedance models; Calculate the reflection coefficient sequence of the virtual well impedance model, construct a seismic record based on the reflection coefficient sequence; and fuse the seismic record and the low-frequency smooth background to obtain a seismic data pair sample; Label the seismic data sample based on the true impedance data to obtain training data; train an artificial neural network based on the training data to obtain an inversion model; and perform impedance inversion based on the inversion model.

2. The deep learning wave impedance inversion method based on virtual wells according to claim 1, wherein Construct a formation lithology model for simulating the physical morphology of multiple formations, including: Obtain a formation index, a randomly generated amplitude, and a randomly generated wavelength; Construct a sine function model based on the formation index, the randomly generated amplitude, and the randomly generated wavelength to obtain a formation lithology model that simulates the physical morphology of the formation using a sine wave.

3. A deep learning wave impedance inversion method based on virtual wells according to claim 1, characterized in that Add initial impedance values and Gaussian noise to the formation lithology model based on lithology classification to obtain an initial impedance model, including: Obtain the initial impedance values and impedance standard deviations corresponding to different lithology classifications; Add the corresponding initial impedance values and Gaussian noise to different lithology formations in the formation lithology model to obtain an initial impedance model.

4. A deep learning wave impedance inversion method based on virtual wells according to claim 1, characterized in that, Calculate the distribution characteristics of the impedance data of multiple lithology categories, including: Calculate the mean and variance of the impedance data of multiple lithology categories to obtain the distribution characteristics of the impedance data of multiple lithology categories.

5. A deep learning wave impedance inversion method based on virtual wells according to claim 4, characterized in that, Construct multiple virtual well impedance models based on the distribution characteristics of the impedance data of multiple lithology categories, including: Generate multiple virtual wells, where the virtual wells include multiple formation units with randomly varying thicknesses, and the lithology of each formation unit is determined by random sampling; Generate a multivariate normal distribution function based on the mean and variance of the impedance data of multiple lithology categories, and generate random impedance values for each formation unit based on the multivariate normal distribution function to obtain multiple virtual well impedance models.

6. The deep learning wave impedance inversion method based on virtual wells according to claim 1, wherein Extract the low-frequency smooth background of the multiple virtual well impedance models, including: Normalize the initial impedance model and the virtual well impedance model to obtain normalized data; Extract the low-frequency smooth background in the normalized data based on the locally weighted regression smoothing algorithm.

7. A deep learning wave impedance inversion method based on virtual wells according to claim 1, characterized in that Calculate the reflection coefficient sequence of the virtual well impedance model, including: Calculate the reflection coefficient of adjacent formation units based on the impedance of adjacent formation units in the virtual well impedance model; Construct a reflection coefficient sequence based on the reflection coefficients of multiple adjacent formation units.

8. A deep learning wave impedance inversion method based on virtual wells according to claim 1, characterized in that, Construct a seismic record based on the reflection coefficient sequence, including: Multiply the reflection coefficient sequence by the Toeplitz matrix formed by the Ricker wavelet to obtain the seismic record.

9. The deep learning wave impedance inversion method based on virtual wells according to claim 1, wherein Train an artificial neural network based on training data to obtain an inversion model, including: Input the training data into the artificial neural network model to obtain a wave impedance prediction model; Calculate the deviation between the wave impedance prediction data and the label based on the mean square error loss function, and adjust the network weights of the artificial neural network based on the deviation; return to input the training data into the artificial neural network model to obtain wave impedance prediction data until the deviation between the wave impedance prediction data and the label is less than a preset deviation threshold to obtain the inversion model.

10. A deep learning wave impedance inversion system based on virtual wells, which is used to implement a deep learning wave impedance inversion method based on virtual wells as described in claim 1, characterized in that, The system includes: A model construction module for constructing a formation lithology model for simulating the physical morphology of multiple formations, and adding an initial wave impedance value and Gaussian noise to the formation lithology model based on lithology classification to obtain an initial wave impedance model; A feature extraction module for extracting multiple well traces from the initial wave impedance model to obtain multiple well data; classifying the multiple well data by lithology to obtain wave impedance data of multiple lithology categories; and calculating the distribution characteristics of the wave impedance data of multiple lithology categories, where the well data includes the wave impedance of multiple formations in the well; A data generation module for constructing multiple virtual well wave impedance models based on the distribution characteristics of the wave impedance data of multiple lithology categories, where the virtual well wave impedance model includes the wave impedance of multiple formations that conforms to the distribution characteristics; A background extraction module for extracting the low-frequency smoothed background of the multiple virtual well wave impedance models; A sample generation module for calculating the reflection coefficient sequence of the virtual well wave impedance model, constructing a seismic record based on the reflection coefficient sequence; and fusing the seismic record and the low-frequency smoothed background to obtain a seismic data pair sample; An inversion module for annotating the seismic data samples based on the true wave impedance data to obtain training data; and Training an artificial neural network based on the training data to obtain an inversion model; and performing wave impedance inversion based on the inversion model.

Images