A seismic wavelet and reflection coefficient joint inversion method and system

Abstract

The present application is suitable for the field of seismic exploration technology, and provides a joint inversion method and system for seismic wavelet and reflection coefficient, which comprises the following steps: constructing a data set conforming to actual seismic data; based on the data set, constructing and training an end-to-end parallel network; the parallel network comprises two parallel independent sub-networks based on deep learning training, which are respectively used for estimating corresponding seismic wavelet and reflection coefficient from seismic signals; inputting the seismic signal to be processed into the parallel network, and synchronously outputting the estimated seismic wavelet and reflection coefficient. Based on the deep learning method, the corresponding seismic wavelet and reflection coefficient are estimated from the seismic signal, the two tasks are processed in parallel, so that the seismic wavelet and the reflection coefficient can be echoed with each other, while meeting the geophysical constraints, the method is suitable for the wavelet estimation and reflection coefficient inversion tasks in seismic data processing, and can be widely applied to the fields of oil and gas exploration, underground structure imaging and the like.

Description

Technical Field

[0001] This invention belongs to the field of seismic exploration technology, and particularly relates to a joint inversion method and system for seismic wavelet and reflection coefficient. Background Technology

[0002] Seismic wavelet estimation is a crucial step in seismic data processing. As the coupling carrier between the source signal and the subsurface reflection sequence, the accuracy of its estimation directly impacts the vertical resolution and amplitude fidelity, and also affects the reliability of subsequent quantitative interpretation. Errors in the seismic wavelet are amplified through processes such as forward and inversion modeling. Seismic signal processing is primarily based on the assumption that the seismic wavelet is of minimum phase and that the reflection coefficient sequence is a statistically independent stationary Gaussian process. However, in actual seismic data, the seismic wavelet is generally of mixed phase, and the reflection coefficients exhibit a symmetrical or near-symmetrical distribution, approximating a distribution between Gaussian and Cauchy. Therefore, accurate estimation of the seismic wavelet is of paramount importance. Methods for seismic wavelet estimation can be categorized into deterministic and statistical methods. Deterministic methods require obtaining the corresponding reflection coefficients from well logging data, then performing deconvolution on the seismic record and reflection coefficients to obtain the corresponding true wavelet. Therefore, this deterministic method is heavily reliant on well logging data, computationally complex, and prone to errors. While statistical methods can estimate seismic wavelets without well logging data, the accuracy of the estimated seismic wavelets may be compromised.

[0003] Seismic data can be viewed as the convolution of a seismic wavelet and a reflection coefficient sequence. Therefore, to obtain an accurate seismic wavelet, the existence of the reflection coefficient cannot be ignored, as the accuracy of the reflection coefficient estimation directly affects the vertical resolution of the seismic signal. If only the seismic wavelet is estimated from the seismic signal without considering the reflection coefficient, the accuracy of the estimated seismic wavelet cannot be determined. Estimating the corresponding reflection coefficient sequence from the seismic signal without a definite wavelet is a "blind" problem, and the accuracy of the estimated reflection coefficient sequence will also be questionable. Therefore, the estimation of the seismic wavelet and the reflection coefficient sequence from the seismic signal are mutually constraining.

[0004] In recent years, deep learning has been gradually applied to seismic data interpretation and attribute prediction, but a systematic end-to-end solution is still lacking in the joint inversion of seismic wavelets and reflection coefficients. Therefore, there is an urgent need for a method that can synchronously and accurately invert seismic wavelets and reflection coefficients to improve the automation of seismic data processing and the physical reliability of the inversion results. Summary of the Invention

[0005] The purpose of this invention is to provide a joint inversion method for seismic wavelet and reflection coefficient, aiming to solve the above-mentioned technical problems.

[0006] This invention is implemented as follows: a joint inversion method for seismic wavelet and reflection coefficient, comprising the following steps:

[0007] Construct a dataset that conforms to actual earthquake data; the dataset includes earthquake signals and their corresponding seismic wavelets and reflection coefficients;

[0008] Based on the dataset, an end-to-end parallel network is constructed and trained. The parallel network includes two parallel and independent sub-networks trained based on deep learning, which are used to estimate the corresponding seismic wavelet and reflection coefficient from the seismic signal, respectively. During the training process of the parallel network, the two sub-networks output the preliminary estimated seismic wavelet and reflection coefficient, respectively. Then, the preliminary estimated seismic wavelet and reflection coefficient are convolved to obtain the reconstructed signal. The reconstruction loss between the reconstructed signal and the input seismic signal is calculated, and the independent loss functions of each sub-network are combined to jointly guide the parameter update of the parallel network.

[0009] The seismic signal to be processed is input into the parallel network, and the estimated seismic wavelet and reflection coefficient are output synchronously.

[0010] Furthermore, the steps for constructing a dataset that conforms to actual earthquake data specifically include:

[0011] Based on the seismic wavelet and reflection coefficient, a seismic signal that conforms to the actual seismic data is constructed.

[0012] Each seismic signal is mapped one-to-one with its corresponding seismic wavelet and reflection coefficient to form a dataset; where the seismic signal is used as input, the seismic wavelet is used as the label for estimating the seismic wavelet, and the reflection coefficient is used as the label for estimating the reflection coefficient.

[0013] Furthermore, the steps for constructing seismic signals that conform to actual seismic data based on seismic wavelets and reflection coefficients specifically include:

[0014] First, a seismic wavelet is constructed, using an ideal Ricker wavelet as the source wavelet for synthesizing the seismic signal. The expression for the Ricker wavelet is as follows:

[0015] ;

[0016] In the formula, Represents the seismic wavelet. Indicates the dominant frequency of the seismic wavelet. Indicates time;

[0017] Secondly, a reflection coefficient that conforms to actual seismic data is constructed. The reflection coefficient in actual seismic data is due to the difference in wave impedance between the upper and lower rock layers at the subsurface interface. The expression for the reflection coefficient is as follows:

[0018] ;

[0019] In the formula, Represents the reflection coefficient. Indicates the density of the strata beneath the interface. This indicates the propagation speed of seismic waves in the strata below the interface. Indicates the density of the strata at the interface. This indicates the propagation speed of seismic waves in the strata at the interface;

[0020] When constructing the reflection coefficients in the dataset, to match the reflection coefficients in actual seismic data, the constructed reflection coefficients are a pulse sequence. The intensity of the pulse signal, i.e., the amplitude of the corresponding reflection coefficient, matches the amplitude of the reflection coefficient in actual seismic data, with the amplitude controlled between -1 and 1. This pulse signal is caused by the inconsistency in wave impedance between two consecutive strata. The expression for wave impedance is as follows:

[0021] ;

[0022] In the formula, Indicates wave impedance, Indicates the density of the strata. This indicates the propagation speed of seismic waves within the Earth's strata.

[0023] After the seismic wavelet and reflection coefficients are constructed, the seismic signal is constructed. The seismic signal is represented as a single-channel seismic signal, which is obtained by the convolution of the seismic wavelet and reflection coefficients. The expression for the single-channel seismic signal is as follows:

[0024] ;

[0025] In the formula, Indicates wave impedance, Indicates the density of the strata. This indicates the propagation speed of seismic waves within the Earth's strata.

[0026] After the seismic wavelet and reflection coefficients are constructed, the seismic signal is constructed. The seismic signal is represented as a single-channel seismic signal, which is obtained by the convolution of the seismic wavelet and reflection coefficients. The expression for the single-channel seismic signal is as follows:

[0027] ;

[0028] In the formula, Indicates earthquake signal, Represents the seismic wavelet. A pulse sequence representing the reflection coefficient. This indicates noise interference.

[0029] Furthermore, the two sub-networks are the wavelet estimation sub-network and the reflection coefficient estimation sub-network, both of which adopt the U-Net network architecture combined with LSTM.

[0030] Furthermore, in the wavelet estimation subnetwork, the independent loss functions used include mean square error loss and Pearson correlation coefficient loss; in the reflection coefficient estimation subnetwork, the independent loss functions used include mean square error loss, Pearson correlation coefficient loss and average error loss; the reconstruction loss is composed of the main losses in the two subnetworks, including mean square error loss and Pearson correlation coefficient loss.

[0031] The expression for the mean squared error loss (MSE) is as follows:

[0032] ;

[0033] In the formula, The total number of samples, Indicates the first The true value of each sample Indicates the first Predicted values ​​for each sample;

[0034] Pearson correlation coefficient loss C orr The expression is as follows:

[0035] ;

[0036] In the formula, Indicates the first The predicted value for each sample, This represents the mean of the predicted values. Indicates the first The true value of each sample The mean of the true values;

[0037] The expression for the average error loss L1 is as follows:

[0038] ;

[0039] In the formula, The total number of samples, Indicates the first The true value of each sample Indicates the first The predicted value for each sample.

[0040] Furthermore, after the step of constructing and training an end-to-end parallel network based on the dataset, the method further includes:

[0041] After the parallel network training is completed, the test set data in the dataset is used for testing. By using the correlation coefficient index and intuitive observation, it is determined whether the estimated seismic wavelet is consistent with the real seismic wavelet, whether the estimated reflection coefficient is consistent with the real reflection coefficient, and whether the reconstructed signal is consistent with the input seismic signal.

[0042] Another object of the present invention is to provide a joint inversion system for seismic wavelet and reflection coefficient, for implementing the above-mentioned joint inversion method, comprising:

[0043] A dataset construction module is used to construct a dataset that conforms to actual earthquake data; the dataset includes earthquake signals and their corresponding seismic wavelets and reflection coefficients.

[0044] A parallel network training module is used to construct and train an end-to-end parallel network based on the dataset. The parallel network includes two parallel independent sub-networks, which are used to estimate the corresponding seismic wavelet and reflection coefficient from the seismic signal, respectively. During the training process of the parallel network, the two sub-networks output the preliminary estimated seismic wavelet and reflection coefficient, respectively. Then, the preliminary estimated seismic wavelet and reflection coefficient are convolved to obtain the reconstructed signal. The reconstruction loss between the reconstructed signal and the input seismic signal is calculated, and the independent loss functions of each sub-network are combined to jointly guide the parameter update of the parallel network.

[0045] The joint inversion module is used to input the seismic signal to be processed into the parallel network and synchronously output the estimated seismic wavelet and reflection coefficient.

[0046] This invention employs a joint inversion method using an end-to-end parallel network, solving the two-step problem of traditional methods. The two-step process involves first estimating the seismic wavelet and then inverting the reflection coefficient. This method can simultaneously output the seismic wavelet and reflection coefficient, and the two can be convolved to reconstruct the seismic signal. This feature, introduced as a geophysical constraint into the parallel network, increases the interpretability of the parallel network and effectively solves the problem of multiple solutions in seismic data processing. This method can estimate the corresponding seismic wavelet and reflection coefficient even without well logging data, solving the "blind" problem. Furthermore, the introduction of neural networks improves computational speed and efficiency. Since the network is labeled, as a supervised learning network, the quality of the constructed dataset also affects the experimental results. A dataset that matches real seismic data will result in accurate experimental results. The method uses a parallel and independent network architecture with two branches, employing a U-Net network architecture incorporating LSTM. The introduced LSTM is responsible for memorizing the global depth direction sequence, which improves the accuracy of the prediction results. In the test experiment to verify the accuracy of the model prediction, the predicted seismic wavelet and reflection coefficient had small errors compared with the actual seismic wavelet and reflection coefficient, indicating that the prediction results were accurate.

[0047] The joint inversion method of seismic wavelet and reflection coefficient provided by this invention is based on deep learning. It estimates the corresponding seismic wavelet and reflection coefficient from the seismic signal and processes the two tasks in parallel, so that the seismic wavelet and reflection coefficient can complement each other and meet the constraints of geophysics. This method is applicable to wavelet estimation and reflection coefficient inversion tasks in seismic data processing and can be widely used in fields such as oil and gas exploration and underground structure imaging. Attached Figure Description

[0048] Figure 1 This is a diagram of the parallel network structure used in the embodiments of the present invention;

[0049] Figure 2 The test set uses seismic wavelets estimated using the method provided in the embodiments of the present invention;

[0050] Figure 3 This represents the original true reflectance coefficient;

[0051] Figure 4 The reflection coefficients estimated using the method provided in the embodiments of the present invention are used in the test set;

[0052] Figure 5 This is real earthquake data;

[0053] Figure 6 The seismic wavelet estimated by traditional methods;

[0054] Figure 7 The seismic wavelet estimated by the method provided in the embodiments of the present invention;

[0055] Figure 8 The reflection coefficient estimated by the method provided in the embodiments of the present invention;

[0056] Figure 9 The reconstructed signal is obtained by convolving the estimated seismic wavelet and the estimated reflection coefficient. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] like Figure 1 As shown, in one embodiment of the present invention, a joint inversion method for seismic wavelet and reflection coefficient is provided, which includes the following steps:

[0059] S1. Construct a dataset that conforms to actual earthquake data; the dataset includes earthquake signals and their corresponding seismic wavelets and reflection coefficients;

[0060] S2. Based on the dataset, construct and train an end-to-end parallel network. The parallel network includes two parallel and independent sub-networks trained based on deep learning, which are used to estimate the corresponding seismic wavelet and reflection coefficient from the seismic signal, respectively. During the training process of the parallel network, the two sub-networks output the preliminary estimated seismic wavelet and reflection coefficient, respectively. Then, the preliminary estimated seismic wavelet and reflection coefficient are convolved to obtain the reconstructed signal. The reconstruction loss between the reconstructed signal and the input seismic signal is calculated, and the independent loss functions of each sub-network are combined to jointly guide the parameter update of the parallel network.

[0061] S3. Input the seismic signal to be processed into the parallel network and output the estimated seismic wavelet and reflection coefficient simultaneously.

[0062] In a preferred embodiment of the present invention, the step of constructing a dataset that conforms to actual earthquake data specifically includes:

[0063] S11. Construct a seismic signal that conforms to the actual seismic data based on the seismic wavelet and reflection coefficient;

[0064] S12. Each seismic signal is mapped one-to-one with its corresponding seismic wavelet and reflection coefficient to form a dataset; where the seismic signal is used as input, the seismic wavelet is used as the label for estimating the seismic wavelet, and the reflection coefficient is used as the label for estimating the reflection coefficient.

[0065] The steps for constructing a seismic signal that conforms to actual seismic data based on the seismic wavelet and reflection coefficient specifically include:

[0066] First, seismic wavelets are constructed, with different dominant frequencies. The Ricker wavelet, commonly used in seismic exploration, is a zero-phase, symmetrical pulse signal. This embodiment of the invention uses an ideal Ricker wavelet as the source wavelet for synthesizing the seismic signal. The expression for the Ricker wavelet is as follows:

[0067] ;

[0068] In the formula, Represents the seismic wavelet. Indicates the dominant frequency of the seismic wavelet. Indicates time;

[0069] Secondly, a reflection coefficient consistent with actual seismic data is constructed. In actual seismic data, the reflection coefficient is due to the difference in wave impedance between the upper and lower rock layers at the subsurface interface. When seismic waves generated by an artificial seismic source propagate downwards, upon encountering a stratum interface with different lithologies, some energy is reflected back to the surface, forming a reflected wave. The strength of the reflection is determined by the reflection coefficient of that interface. The expression for the reflection coefficient is as follows:

[0070] ;

[0071] In the formula, Represents the reflection coefficient. Indicates the density of the strata beneath the interface. This indicates the propagation speed of seismic waves in the strata below the interface. Indicates the density of the strata at the interface. This indicates the propagation speed of seismic waves in the strata at the interface;

[0072] The greater the difference in wave impedance between the rocks on both sides of the interface, the greater the corresponding reflection coefficient and the stronger the reflected signal. If the wave impedances are similar, the reflection coefficient is smaller. When the reflection coefficient is close to 0, there is almost no reflection signal. When constructing the reflection coefficients in the dataset, to match the reflection coefficients in actual seismic data, the constructed reflection coefficients are a pulse sequence. A specific pulse signal exists at a specific location, and the intensity of the pulse signal, i.e., the amplitude of the corresponding reflection coefficient, matches the amplitude of the reflection coefficient in actual seismic data, with the amplitude controlled between -1 and 1. This pulse signal is caused by the inconsistency in wave impedance between the two strata. The expression for wave impedance is as follows:

[0073] ;

[0074] In the formula, Indicates wave impedance, Indicates the density of the strata. This indicates the propagation speed of seismic waves within the Earth's strata.

[0075] After the seismic wavelet and reflection coefficients are constructed, the seismic signal is constructed. The seismic signal is represented as a single-channel seismic signal, which is obtained by the convolution of the seismic wavelet and reflection coefficients. The expression for the single-channel seismic signal is as follows:

[0076] ;

[0077] In the formula, Indicates earthquake signal, Represents the seismic wavelet. A pulse sequence representing the reflection coefficient. This indicates noise interference.

[0078] After constructing the seismic wavelet, reflection coefficient, and seismic signal, the seismic wavelet and reflection coefficient corresponding to each seismic signal are matched one-to-one and saved as an npy file. In other words, the seismic signal, reflection coefficient, and corresponding seismic wavelet are paired up, with the seismic signal as the input, the seismic wavelet as the label for the seismic wavelet estimation task, and the reflection coefficient as the label for the reflection coefficient estimation.

[0079] In this embodiment of the invention, a dual-branch parallel network is used. One branch is used to estimate the corresponding seismic wavelet from the seismic signal, and the other branch is used to estimate the reflection coefficient sequence from the seismic signal. Since the two independent networks have different tasks and require different data, to make the constructed dataset convenient and responsive to network calls, the seismic wavelet, reflection coefficient sequence, and seismic signal are placed in the same file. In the seismic wavelet estimation sub-network, the seismic signal from the file is used as input, and the seismic wavelet is used as the label. In the reflection coefficient estimation sub-network, the seismic signal from the file is used as input, and the reflection coefficient sequence is used as the label. There are a total of 2100 .npy format files, where the first column of each file is the actual seismic wavelet, the second column is the actual reflection coefficient sequence, and the third column is the seismic signal.

[0080] like Figure 1 As shown, in a preferred embodiment of the present invention, the step of constructing and training an end-to-end parallel network based on the dataset specifically includes:

[0081] S21. Constructing a Parallel Network: The end-to-end parallel network consists of two independent sub-networks, each performing a corresponding task. One sub-network estimates the corresponding wavelet from the seismic signal, and the other estimates the reflection coefficient. Specifically, the two sub-networks are for wavelet estimation and reflection coefficient estimation, respectively. Both sub-networks use a U-Net architecture combined with LSTM. During training, the estimated seismic wavelet and the estimated reflection coefficient, initially estimated from the seismic signal, are convolved to construct a reconstructed signal. The reconstructed signal is compared with the input seismic signal to construct a reconstruction loss, which guides the training of both networks. The reconstruction loss serves as the primary loss return parameter to guide the training of both sub-networks. Each of the two individual sub-networks also has its own independent loss function to guide their training.

[0082] In the wavelet estimation subnetwork, the independent loss functions used include mean squared error loss (MSE) and Pearson correlation coefficient loss. In the reflection coefficient estimation subnetwork, the independent loss functions used include mean squared error loss, Pearson correlation coefficient loss, and average error loss. The reconstruction loss is a combination of the main losses from the two subnetworks, including mean squared error loss and Pearson correlation coefficient loss. Mean squared error loss is a commonly used loss function, calculated by subtracting the predicted and true values ​​point by point, squaring the results, and then averaging them. Each sample is considered a point in a high-dimensional space; the mean squared error loss is the mean of the squared Euclidean distance. The gradient direction always follows the line connecting the predicted and true values, and its magnitude is proportional to the absolute value of the error. Therefore, convergence is faster when the gradient descent occurs. The Pearson correlation coefficient is an indicator of the degree of linear correlation between two variables, ranging from -1 to 1, where 1 represents perfect positive correlation, -1 represents perfect negative correlation, and 0 represents no linear correlation. In the subnetwork of seismic wavelet estimation, this is used to represent the correlation between the predicted and actual seismic wavelets, indicating whether their waveforms are identical. In seismic data, the reflection coefficient is inherently sparse; adding an average error loss can suppress most points to zero, leaving only the pulses corresponding to the main interfaces.

[0083] The expression for the mean squared error loss (MSE) is as follows:

[0084] ;

[0085] In the formula, The total number of samples, Indicates the first The true value of each sample Indicates the first Predicted values ​​for each sample;

[0086] Pearson correlation coefficient loss C orr The expression is as follows:

[0087] ;

[0088] In the formula, Indicates the first The predicted value for each sample, This represents the mean of the predicted values. Indicates the first The true value of each sample The mean of the true values;

[0089] The expression for the average error loss L1 is as follows:

[0090] ;

[0091] In the formula, The total number of samples, Indicates the first The true value of each sample Indicates the first The predicted value for each sample.

[0092] In this parallel network, two independent sub-networks are constructed to perform different tasks. One branch of the network is used to estimate the corresponding seismic wavelet from the seismic signal, and the other network is used to estimate the corresponding reflection coefficient sequence from the seismic signal. The network framework for both tasks is consistent, using a U-Net network incorporating LSTM. LSTM, or Long Short-Term Memory network, is a special type of recurrent neural network. Unlike general feedforward neural networks, LSTM can analyze the input using time series data. Since seismic signals can be viewed as signals along the time or depth direction, the reflection coefficients can also be viewed as signals along the time direction. The purpose of introducing LSTM into the U-Net network is that LSTM is responsible for remembering the global depth direction sequence, which can improve the prediction accuracy.

[0093] S22. Training the Parallel Network: In terms of dataset partitioning, 90% of the dataset is used as the training set, and the remaining 10% is used as the test set to verify the performance and accuracy of the trained network. When training the seismic wavelet estimation sub-network, seismic signals and seismic wavelets from the dataset are used, with the seismic signals as input and the seismic wavelets as labels. In training the reflection coefficient estimation sub-network, seismic signals from the dataset are used as input, and the reflection coefficient sequence is used as labels. After initially obtaining the estimated seismic wavelets and reflection coefficients, they are convolved to reconstruct the network, obtaining the reconstructed signal. The reconstructed signal and the input signal are used to construct a reconstruction loss to guide the updating of the parameters in both sub-networks during training.

[0094] S23. Verify the performance of the parallel network: After the parallel network training is completed, test it using the test set data in the dataset. Analyze the correlation coefficient and visually verify whether the estimated seismic wavelet matches the actual seismic wavelet, the estimated reflection coefficient matches the actual reflection coefficient, and the reconstructed signal matches the input seismic signal. Repeat the parallel network training steps until the errors between the estimated and actual seismic wavelet, the estimated and actual reflection coefficients, and the reconstructed signal are very small. This indicates that the parallel network's estimation of the seismic wavelet and reflection coefficient is accurate.

[0095] The following verification experiments illustrate the predictive effectiveness of the method provided in this embodiment of the invention:

[0096] I. The corresponding seismic wavelet and reflection coefficient sequence were estimated from the seismic signals in the test set:

[0097] The dataset used in this experiment consists of 2100 .npy files. The first column of each .npy file contains the actual seismic wavelet, the second column contains the actual reflection coefficient sequence, and the third column contains the seismic signal. 90% of the dataset is used as the training set, and the remaining 10% is used as the test set. After the parallel network training is complete, the seismic signals from the test set are input into the network to test the accuracy of the predicted seismic wavelet and the actual wavelet, as well as the accuracy of the predicted reflection coefficient sequence and the actual reflection coefficient sequence. The correlation coefficients between the estimated seismic wavelet and the actual seismic wavelet are shown in Table 1 and... Figure 2 As shown.

[0098] Table 1. Correlation coefficients between predicted seismic wavelets and actual seismic wavelets

[0099]

[0100] The correlation coefficients between the estimated reflectance and the true reflectance are shown in Table 2 and... Figures 3-4 As shown:

[0101] Table 2. Correlation coefficients between predicted and actual reflectance coefficients

[0102]

[0103] From Table 1-2 and Figures 2-4 As can be seen from the test data, the correlation coefficient between the estimated seismic wavelet and the actual seismic wavelet is high, and the correlation coefficient between the estimated reflection coefficient and the actual reflection coefficient is also high. This indicates that the parallel network provided in this embodiment of the invention can effectively estimate the corresponding seismic wavelet and reflection coefficient from the seismic signal.

[0104] II. Based on real earthquake data (such as...) Figure 5 As shown in the figure, the corresponding seismic wavelet and the corresponding reflection coefficient are estimated from the seismic signal. The seismic wavelet estimated from the seismic signal using traditional methods is as follows. Figure 6 As shown; the traditional method uses the cepstral method, whose formula is as follows:

[0105] ;

[0106] ;

[0107] ;

[0108] In the formula, For time-domain seismic signals, The time-domain signal of the seismic wavelet. The time-domain signal of the reflection coefficient, For frequency domain seismic signals, The frequency domain signal of the seismic wavelet. The frequency domain signal is the reflection coefficient; the cepstral method is the time domain signal of the seismic signal. Transform to the frequency domain, and then convert the signal in the frequency domain. Take the logarithm and transform to the cepstral domain. By applying a window, the seismic wavelet and reflection coefficient in the cepstral domain are separated.

[0109] The seismic wavelet and reflection coefficient estimated from the seismic signal using the method provided in this embodiment of the invention are as follows: Figures 7-8 As shown; furthermore, by convolving the estimated seismic wavelet and reflection coefficient, the reconstructed signal can be obtained, as shown in the figure. Figure 9 As shown.

[0110] In another embodiment of the present invention, a joint inversion system for seismic wavelet and reflection coefficient is also provided to implement the above-mentioned joint inversion method, specifically including:

[0111] A dataset construction module is used to construct a dataset that conforms to actual earthquake data; the dataset includes earthquake signals and their corresponding seismic wavelets and reflection coefficients.

[0112] A parallel network training module is used to construct and train an end-to-end parallel network based on the dataset. The parallel network includes two parallel independent sub-networks, which are used to estimate the corresponding seismic wavelet and reflection coefficient from the seismic signal, respectively. During the training process of the parallel network, the two sub-networks output the preliminary estimated seismic wavelet and reflection coefficient, respectively. Then, the preliminary estimated seismic wavelet and reflection coefficient are convolved to obtain the reconstructed signal. The reconstruction loss between the reconstructed signal and the input seismic signal is calculated, and the independent loss functions of each sub-network are combined to jointly guide the parameter update of the parallel network.

[0113] The joint inversion module is used to input the seismic signal to be processed into the parallel network and synchronously output the estimated seismic wavelet and reflection coefficient.

[0114] It should be noted that each of the above modules can be implemented as a computer program, which can run on a computer device. The computer device's memory can store the computer program that makes up each module, enabling the processor to execute each step of the above method.

[0115] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0116] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0117] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A joint inversion method for seismic wavelet and reflection coefficient, characterized in that, Includes the following steps: Construct a dataset that conforms to actual earthquake data; the dataset includes earthquake signals and their corresponding seismic wavelets and reflection coefficients; Based on the dataset, an end-to-end parallel network is constructed and trained. The parallel network includes two parallel and independent sub-networks trained based on deep learning, which are used to estimate the corresponding seismic wavelet and reflection coefficient from the seismic signal, respectively. During the training process of the parallel network, the two sub-networks output the preliminary estimated seismic wavelet and reflection coefficient, respectively. Then, the preliminary estimated seismic wavelet and reflection coefficient are convolved to obtain the reconstructed signal. The reconstruction loss between the reconstructed signal and the input seismic signal is calculated, and the independent loss functions of each sub-network are combined to jointly guide the parameter update of the parallel network. The seismic signal to be processed is input into the parallel network, and the estimated seismic wavelet and reflection coefficient are output synchronously. The two sub-networks are the wavelet estimation sub-network and the reflection coefficient estimation sub-network, both of which adopt the U-Net network architecture combined with LSTM; In the wavelet estimation subnetwork, the independent loss functions used include mean square error loss and Pearson correlation coefficient loss; in the reflection coefficient estimation subnetwork, the independent loss functions used include mean square error loss, Pearson correlation coefficient loss and average error loss; the reconstruction loss is composed of the main losses in the two subnetworks, including mean square error loss and Pearson correlation coefficient loss. The expression for the mean squared error loss (MSE) is as follows: ； In the formula, The total number of samples, Indicates the first The true value of each sample Indicates the first Predicted values ​​for each sample; Pearson correlation coefficient loss C orr The expression is as follows: ； In the formula, Indicates the first The predicted value for each sample, This represents the mean of the predicted values. Indicates the first The true value of each sample The mean of the true values; The expression for the average error loss L1 is as follows: ； In the formula, The total number of samples, Indicates the first The true value of each sample Indicates the first Predicted values ​​for each sample; The steps for constructing a dataset that matches actual earthquake data include: Based on the seismic wavelet and reflection coefficient, a seismic signal that conforms to the actual seismic data is constructed. Each seismic signal is mapped one-to-one with its corresponding seismic wavelet and reflection coefficient to form a dataset; where the seismic signal is used as input, the seismic wavelet is used as the label for estimating the seismic wavelet, and the reflection coefficient is used as the label for estimating the reflection coefficient. The steps for constructing a seismic signal that conforms to actual seismic data based on the seismic wavelet and reflection coefficient specifically include: First, a seismic wavelet is constructed, using an ideal Ricker wavelet as the source wavelet for synthesizing the seismic signal. The expression for the Ricker wavelet is as follows: ； In the formula, Represents the seismic wavelet. Indicates the dominant frequency of the seismic wavelet. Indicates time; Secondly, a reflection coefficient that conforms to actual seismic data is constructed. The reflection coefficient in actual seismic data is due to the difference in wave impedance between the upper and lower rock layers at the subsurface interface. The expression for the reflection coefficient is as follows: ； In the formula, Represents the reflection coefficient. Indicates the density of the strata beneath the interface. This indicates the propagation speed of seismic waves in the strata below the interface. Indicates the density of the strata at the interface. This indicates the propagation speed of seismic waves in the strata at the interface; When constructing the reflection coefficients in the dataset, to match the reflection coefficients in actual seismic data, the constructed reflection coefficients are a pulse sequence. The intensity of the pulse signal, i.e., the amplitude of the corresponding reflection coefficient, matches the amplitude of the reflection coefficient in actual seismic data, with the amplitude controlled between -1 and 1. This pulse signal is caused by the inconsistency in wave impedance between two consecutive strata. The expression for wave impedance is as follows: ； In the formula, Indicates wave impedance, Indicates the density of the strata. This indicates the propagation speed of seismic waves within the Earth's strata. After the seismic wavelet and reflection coefficients are constructed, the seismic signal is constructed. The seismic signal is represented as a single-channel seismic signal, which is obtained by the convolution of the seismic wavelet and reflection coefficients. The expression for the single-channel seismic signal is as follows: ； In the formula, Indicates earthquake signal, Represents the seismic wavelet. A pulse sequence representing the reflection coefficient. This indicates noise interference.

2. The joint inversion method of seismic wavelet and reflection coefficient according to claim 1, characterized in that, Following the step of constructing and training an end-to-end parallel network based on the dataset, the method further includes: After the parallel network training is completed, the test set data in the dataset is used for testing. By using the correlation coefficient index and intuitive observation, it is determined whether the estimated seismic wavelet is consistent with the real seismic wavelet, whether the estimated reflection coefficient is consistent with the real reflection coefficient, and whether the reconstructed signal is consistent with the input seismic signal.

3. A joint inversion system for seismic wavelet and reflection coefficient, used to implement the joint inversion method according to any one of claims 1-2, characterized in that, include: A dataset construction module is used to construct a dataset that conforms to actual earthquake data; the dataset includes earthquake signals and their corresponding seismic wavelets and reflection coefficients. A parallel network training module is used to construct and train an end-to-end parallel network based on the dataset. The parallel network includes two parallel independent sub-networks, which are used to estimate the corresponding seismic wavelet and reflection coefficient from the seismic signal, respectively. During the training process of the parallel network, the two sub-networks output the preliminary estimated seismic wavelet and reflection coefficient, respectively. Then, the preliminary estimated seismic wavelet and reflection coefficient are convolved to obtain the reconstructed signal. The reconstruction loss between the reconstructed signal and the input seismic signal is calculated, and the independent loss functions of each sub-network are combined to jointly guide the parameter update of the parallel network. The joint inversion module is used to input the seismic signal to be processed into the parallel network and synchronously output the estimated seismic wavelet and reflection coefficient.

Images