Unsupervised anti-q filtering method based on decay kernel matrix driving

Abstract

The application discloses an unsupervised anti-Q filtering method based on an attenuation kernel matrix drive, first, the data to be compensated is taken as the input of a DL network, and the compensated data is taken as the output. Then, the output is sent into an attenuation kernel matrix constructed by a near-surface Q model to perform forward attenuation. Then, by using the error between the attenuation seismic data obtained by forward and the original input data, the error is minimized by optimizing the network parameters, and the final compensation result is output. In the whole network prediction process, data labels are not needed, and the effect of unsupervised learning is achieved. The method introduces an attenuation kernel matrix physical model, thereby getting rid of the dependence on data labels, and solves the problems that the imaging precision of seismic data is seriously affected by the factors such as uneven high-frequency noise and complex surface environment, and the surface structure system cannot be effectively analyzed.

Description

Technical Field

[0001] This invention belongs to the field of near-surface exploration seismic data processing, and relates to an unsupervised inverse Q-filtering method driven by attenuation kernel matrix, particularly to an efficient near-surface seismic wave amplitude compensation method using deep learning during seismic data imaging. Background Technology

[0002] Inverse Q-filtering is widely used for seismic wave attenuation compensation to improve the resolution of seismic data. With the development of seismic exploration technology, the demand for higher imaging accuracy of seismic data is constantly increasing. To provide more accurate information, methods such as inverse Q-filtering, Gabor deconvolution, and Q-migrating have developed rapidly, especially in near-surface surveys. Among these, inverse Q-filtering is the most commonly used. Inverse Q-filtering can be viewed as the reverse propagation process of seismic waves. It is a method that can effectively compensate for the amplitude loss and phase distortion of seismic waves caused by formation attenuation effects. By calculating the formation quality factor Q and substituting it into the compensation function, it can effectively compensate for the energy loss and phase distortion caused by seismic waves propagating in viscoelastic media, thereby improving the continuity of the seismic data's phase axis, enhancing the energy of deep, weakly reflected waves, and playing a role in amplitude preservation and resolution improvement. However, this method faces many challenges in practical applications, such as non-uniform high-frequency noise and complex surface environments. It also suffers from low computational efficiency and insufficient amplitude compensation, which can seriously affect the imaging accuracy of seismic data, making it impossible to effectively analyze surface structural systems. Therefore, with the large-scale application of compensation technology in near-surface exploration, some scholars have studied inverse Q compensation technology for near-surface applications. Currently, there are phase, stability factor, variable stability factor, and adaptive gain limit based on Fourier transform of wavefield extension theory, which can improve the numerical stability, compensation intensity, frequency bandwidth, and phase correction to a certain extent. However, further research is needed to expand its efficiency and practical applications.

[0003] Machine learning and deep learning have been successfully applied to various aspects of geophysical exploration with good results. Examples include automatic first arrival picking, high-resolution reconstruction of seismic data, seismic data denoising, seismic data reconstruction, seismic fault identification, seismic facies classification, reservoir parameter prediction, and seismic data inversion. However, due to the differences between synthetic and real-world data, networks trained on synthetic data are difficult to use for practical compensation. Accurate inverse Q-filter labels cannot be obtained from field seismic data; if labels are obtained from real-world data using traditional methods, the compensation effect is at best comparable to, or even worse than, traditional methods. Therefore, due to the lack of a training set of real-world data, and the fact that the update of the network's hidden parameters in supervised learning depends on the difference between the network output and the loss in the labeled data, supervised learning is not suitable for inverse Q-filtering tasks. This has become a significant challenge for the practical application of deep learning methods in inverse Q-filtering. Summary of the Invention

[0004] The purpose of this invention is to provide an unsupervised inverse Q-filtering method driven by an attenuation kernel matrix. This method introduces a physical model of the attenuation kernel matrix, thereby eliminating the dependence on data labels. Based on constructing a near-surface attenuation kernel matrix, this method uses seismic data to perform inverse Q-filtering compensation prediction on a deep learning network. The prediction results of the network are then forward-modeled, and the errors of the forward-modeling results and the data to be compensated are backpropagated to the network for continuous optimization and iteration until convergence, thus achieving the effect of unsupervised learning.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: an unsupervised inverse Q-filtering method driven by an attenuation kernel matrix, which includes the following steps:

[0006] Step 1: Rearrange the known pre-stack or post-stack seismic data, input the original data to be compensated into the compensation module containing the U-Net network, and obtain the prediction model, which is the compensated seismic data, through the calculation of the network model.

[0007] Step 2: Input the known near-surface Q value and velocity model to establish a near-surface pre-stack or post-stack attenuation kernel matrix into the compensated data to obtain attenuated seismic data;

[0008] Step 3: Calculate the error between all input data to be compensated and the output attenuated seismic data, and backpropagate the error to the network to iteratively update the network parameters, thereby reducing the error between the data to be compensated and the attenuated seismic data and obtaining the optimal network parameters.

[0009] Step 4: Finally, the compensated seismic record is obtained directly using the network module with the optimal parameters.

[0010] Furthermore, in step one, the data selection and sorting before or after stacking mainly includes: CSP / CRP dual-domain gather and time-frequency domain normalization processing; treating the actual inverse Q filtering as an inversion process, using the Res-UNet network module and adding a residual network to establish compensation for the inverse Q filtering, and processing them in sequence.

[0011] Furthermore, the inversion objective function is expressed as:

[0012]

[0013] In the formula: B C The output is the compensated data; A is the wavefield of the original data to be compensated, and D is the attenuation kernel matrix.

[0014] Furthermore, in step two, since the near-surface contains different lithological layers, and the seismic wave travels a propagation path from the source to the reflection interface and back to the surface, the interlayer travel time with non-zero offset can be calculated by using the interlayer propagation path and velocity model. The interlayer travel time with zero offset can be extrapolated to obtain the corrected time-domain Q-value model and the corresponding attenuation kernel matrix. The attenuation kernel matrix Q-model is then used for positive stable absorption and attenuation of seismic waves.

[0015]

[0016] In the formula: t is time; P(t,w) is the seismic wave field at time t; Q is the quality factor; w is the angular frequency, w=2πf; γ=1 / (πQ);

[0017] Converting the above into matrix form, we have:

[0018] A = C.*D(Δt,Q)

[0019] In the formula: C and A are the wave fields before and after attenuation (time-frequency domain data), respectively; D is the attenuation kernel matrix, which is related to the near-surface Q value and Δt; * represents matrix dot product.

[0020] Furthermore, in step three, the error between the input data to be compensated and the output attenuated seismic data is calculated. The near-surface Q-model obtained by forward modeling is converted into different attenuation kernel matrices. Then, the compensated seismic data and the attenuation kernel matrix are multiplied by matrix to obtain the wavefield of the attenuated seismic data in forward modeling. The error between the attenuated seismic data and the data to be compensated is calculated. The error is backpropagated to the network to update the network parameters ω' through iteration. The Epoch is trained to perform iteration, so that the error between the data to be compensated and the attenuated seismic data is minimized, and the optimal network parameters are obtained.

[0021] Furthermore, the MSE loss function of the network is:

[0022]

[0023] In the formula: the wavefield A of the original data to be compensated is input into the compensation module containing the U-Net network, D is the forward transformation operator of the attenuation kernel matrix, and B... C To output the compensated data.

[0024] Furthermore, in step four, the seismic data is input into the network module with optimal parameters to directly obtain the compensated seismic record.

[0025] In summary, the basic principle of this invention is as follows:

[0026] Unlike traditional inverse Q-filtering methods, we treat formation absorption attenuation as a forward modeling problem and inverse Q-filtering as an inverse modeling problem. By constructing a near-surface attenuation kernel matrix, we perform forward modeling on the prediction results of the deep learning network (DL network). Then, the error between the forward modeling result and the data to be compensated is backpropagated to the network for continuous optimization and iteration until convergence, thus transforming data-driven supervised learning into physics-driven unsupervised learning. First, the data to be compensated is used as input to the DL network, and the compensated data is used as output. Then, the output is fed into the attenuation kernel matrix constructed from the near-surface Q-model for forward attenuation. Next, using the error between the attenuated seismic data obtained from the forward modeling and the original input data, the network parameters are optimized to minimize the error, outputting the final compensation result. Throughout the entire network prediction process, no data labels are required, achieving the effect of unsupervised learning. Application results using theoretical model data and actual seismic data show that, compared with the conventional inverse Q-filtering method, the unsupervised method can effectively compensate seismic signals and has high numerical stability, improving the resolution and signal-to-noise ratio of seismic records.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] First, the reliability of the method's principles. It primarily utilizes a well-established deep learning network framework. By combining deep learning networks with physical equations, the uninterpretability of deep learning can be made physically meaningful, thus transforming data-driven supervised learning into physics-driven unsupervised learning. Furthermore, positive decay avoids the numerical instability of reverse compensation, demonstrating the method's reliability.

[0029] Secondly, the method is highly efficient. Other traditional inverse Q-filter compensation methods suffer from low computational efficiency, poor stability, and poor generalization. In contrast, this method features a stable network structure model, requires no dataset creation, does not need to consider generalization, and requires minimal manual intervention, making it a simple and efficient inverse Q-filter compensation method.

[0030] Third, the stability of the actual results. In the actual processing, high-frequency noise, low-frequency loss, and numerical stability are considered first. It is also demonstrated that the unsupervised inverse Q-filter method has better compensation effect and noise resistance. The application results of actual data show that the compensation effectively improves the resolution and matches well data in the work area well, which is beneficial for further structural interpretation and reservoir prediction, proving the practicality and stability of the method.

[0031] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0032] Figure 1 It is a flowchart of an unsupervised inverse Q-filtering method driven by the attenuation kernel matrix.

[0033] Figure 2 This is an overview of region A; where (a) is a distribution map of Q-value survey points in region A, (b) is the near-surface Q-model, and (c) is the original two-dimensional post-stack data before compensation.

[0034] Figure 3 The results show the compensation effect in region A; (a) is the compensation result of the adaptive gain-limited inverse Q-filtering method, (b) is the compensation result of the unsupervised inverse Q-filtering method, and (c) is the spectrum comparison diagram of channel 200.

[0035] Figure 4 This is a comparison of the effects of data compensation before and after in region A; where (a) is the original two-dimensional logging post-stack data, (b) is the compensation result of the unsupervised inverse Q filtering method, and (c) is the spectrum comparison of channel 180.

[0036] Figure 5 The results show a comparison of the wave impedance inversion results of the two methods before and after compensation in area A; (a) is the wave impedance inversion result of the adaptive gain-limited inverse Q-filter method, (b) is the wave impedance inversion result of the unsupervised inverse Q-filter method, and (c) is a comparison of the inversion results of the well bypass section with the well data. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0038] like Figure 1 An unsupervised inverse Q-filtering method driven by a decay kernel matrix includes the following steps:

[0039] Step 1: Rearrange the known pre-stack or post-stack seismic data, input the original data to be compensated into the compensation module containing the U-Net network, and obtain the prediction model, which is the compensated seismic data, through the calculation of the network model.

[0040] Step 2: Input the known near-surface Q value and velocity model to establish a near-surface pre-stack or post-stack attenuation kernel matrix into the compensated data to obtain attenuated seismic data.

[0041] Step 3: Calculate the error between all input data to be compensated and the output attenuated seismic data. Propagate the error back to the network and iteratively update the network parameters to minimize the error between the data to be compensated and the attenuated seismic data, thus obtaining the optimal network parameters.

[0042] Step 4: Finally, the compensated seismic record is obtained directly using the network module with the optimal parameters.

[0043] In step one, the pre- or post-stack data selection and sorting mainly includes: CSP / CRP dual-domain gathers and time-frequency domain normalization processing; the actual inverse Q filtering is regarded as an inversion process, using the Res-UNet network module and adding a residual network to establish compensation for the inverse Q filtering, and processing is carried out in the above order. Its inversion objective function can be expressed as:

[0044]

[0045] In the formula: B C The output is the compensated data; A is the wavefield of the original data to be compensated, and D is the attenuation kernel matrix.

[0046] In step two, since the near-surface contains different lithological layers, and the seismic wave travels a propagation path from the source to the reflection interface and back to the surface, the interlayer travel time with non-zero offset is calculated by using the interlayer propagation path and velocity model. The interlayer travel time with zero offset can be extrapolated to obtain the corrected time domain Q-value model and the corresponding attenuation kernel matrix. The attenuation kernel matrix Q-model is then used for positive stable absorption and attenuation of seismic waves.

[0047]

[0048] In the formula: t is time; P(t,w) is the seismic wave field at time t; Q is the quality factor; w is the angular frequency, w=2πf; γ=1 / (πQ);

[0049] Converting the above into matrix form, we have:

[0050] A = C.*D(Δt,Q)

[0051] In the formula: C and A are the wave fields before and after attenuation (time-frequency domain data), respectively; D is the attenuation kernel matrix, which is related to the near-surface Q value and Δt; * represents matrix dot product.

[0052] In step three, the error between the input data to be compensated and the output attenuated seismic data is calculated. The near-surface Q-model obtained through forward modeling is converted into different attenuation kernel matrices. Then, the compensated seismic data and the attenuation kernel matrix are multiplied by a matrix to obtain the wavefield of the attenuated seismic data. The error between the attenuated seismic data and the data to be compensated is calculated, and this error is backpropagated to the network to iteratively update the network parameters ω'. The network is trained using Epochs to iterate until the error between the data to be compensated and the attenuated seismic data is minimized, thus obtaining the optimal network parameters. At this point, the network's MSE loss function is:

[0053]

[0054] In the formula: the wavefield A of the original data to be compensated is input into the compensation module containing the U-Net network, D is the forward transformation operator of the attenuation kernel matrix, and B... C To output the compensated data.

[0055] In step four, the seismic data is input into the network module with optimal parameters to directly obtain the compensated seismic record.

[0056] To verify the effectiveness of the unsupervised inverse Q-filtering method driven by the attenuation kernel matrix, it was applied to absorption attenuation compensation of real data. The following analysis uses a loess plateau region in western China as an example.

[0057] Micrologging data from area A was selected as a typical example for analysis. Four Q-value survey points were established on three large plateaus, using ultra-deep micrologging for Q-value investigation; five Q-value survey points were established on three river gullies and slopes, using small-refraction + dual-well micrologging for Q-value investigation; conventional micrologging was used for Q-value investigation on other plateaus, slopes, and gully bottoms in the work area. The distribution map of Q-value survey points in area A is shown below. Figure 2 As shown in (a), its near-surface Q-model is as follows: Figure 2 As shown in (b), in this example, we selected one line in the work area for the experiment and the results are presented. Figure 2 (b) The two-dimensional Q model corresponding to this line is extracted from the near-surface Q model obtained by establishing the near-surface QV fitting relationship through micro-logging data and lithology of the work area. Figure 2 (c) shows a portion of a selected post-stack profile with 400 data points and a sampling interval of 4 ms. The black vertical lines represent well locations. It can be seen that the resolution of the seismic data gradually decreases with increasing propagation time.

[0058] pass Figure 3A comparison clearly shows that both the adaptive gain limiting method and the unsupervised method significantly improve data resolution. For easier observation, the data within the red box is magnified. Compared to the traditional adaptive gain limiting method, the unsupervised method demonstrates a significantly improved overall resolution and a higher signal-to-noise ratio. However, the attenuation compensation results and amplitude spectrum results indicate that the unsupervised method performs better.

[0059] To further verify the reliability of the compensation results, we performed wave impedance inversion on the data from 0.5 to 2.5 s after obtaining the compensation results. Figure 5 The comparison of acoustic impedance inversion results before and after compensation is shown. It can be seen that the acoustic impedance results after unsupervised inverse Q filtering clearly show reservoir details, with a significant improvement in resolution. To further verify the reliability of the inversion, the inversion results from the well bypass were compared with well logging data. Figure 5 (c) shows a comparison of the inversion results for the two methods. Figure 5 As can be seen in (c), the trends of both the unsupervised learning method and the adaptive gain limit method are quite consistent with the logging data. However, upon closer observation, it is not difficult to see that the inversion results of the unsupervised learning method are closer to the logging data, and the resolution is significantly improved.

[0060] Application results using actual seismic data demonstrate that the proposed unsupervised inverse Q-filtering method improves resolution and enhances the quality and numerical stability of seismic imaging. Comparison with the adaptive gain-limiting method reveals that the proposed method exhibits good practicality and reliability, providing dependable data for subsequent compensation.

[0061] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention. Parts not covered in this invention are the same as or can be implemented using existing technology.

Claims

1. An unsupervised inverse Q-filtering method driven by a decay kernel matrix, characterized in that... Includes the following steps: Step 1: Rearrange the known pre-stack or post-stack seismic data, input the original data to be compensated into the compensation module containing the U-Net network, and obtain the prediction model, which is the compensated seismic data, through the calculation of the network model. Step 2: Input the known near-surface Q value and velocity model to establish a near-surface pre-stack or post-stack attenuation kernel matrix into the compensated data to obtain attenuated seismic data; Step 3: Calculate the error between all input data to be compensated and the output attenuated seismic data, and backpropagate the error to the network to iteratively update the network parameters, thereby reducing the error between the data to be compensated and the attenuated seismic data and obtaining the optimal network parameters. Step 4: Finally, the compensated seismic record is obtained directly using the network module with the optimal parameters.

2. The unsupervised inverse Q-filtering method based on attenuation kernel matrix driving according to claim 1, characterized in that: In step one, data rearrangement before or after stacking includes: CSP / CRP dual-domain gather and time-frequency domain normalization; the actual inverse Q filtering is regarded as an inversion process, and the compensation for inverse Q filtering is realized by using the Res-UNet network module and adding a residual network, and processed in sequence.

3. The unsupervised inverse Q-filtering method based on attenuation kernel matrix driving according to claim 1, characterized in that: The inversion objective function is expressed as: In the formula: To output the compensated data; The wave field of the original data to be compensated. For the decay kernel matrix, This is a matrix dot product.

4. The unsupervised inverse Q-filtering method based on attenuation kernel matrix driving according to claim 1, characterized in that: In step two, when calculating the inter-layer travel time with non-zero offset, the inter-layer travel time with zero offset is extrapolated to obtain the corrected time-domain Q-value model and the corresponding attenuation kernel matrix. The attenuation kernel matrix Q-model is then used for positive stable absorption and attenuation of seismic waves. In the formula: For time; for The seismic wave field at that moment; For quality factor; Angular frequency, ; ; ; Converting to matrix form, we have: In the formula: and These are the wave fields before and after attenuation, respectively; The attenuation kernel matrix is ​​related to the near-surface Q value and related; This is a matrix dot product.

5. The unsupervised inverse Q-filtering method based on attenuation kernel matrix driving according to claim 1, characterized in that: In step three, the error between the input data to be compensated and the output attenuated seismic data is calculated. The near-surface Q-model obtained through forward modeling is then converted into different attenuation kernel matrices. Next, the compensated seismic data and the attenuation kernel matrix are multiplied by a matrix to obtain the wavefield of the attenuated seismic data obtained through forward modeling. The error between the attenuated seismic data and the data to be compensated is calculated, and this error is backpropagated to the network to iteratively update the network parameters. The network is trained by iterative Epochs to minimize the error between the data to be compensated and the attenuated seismic data, thus obtaining the optimal network parameters.

6. The unsupervised inverse Q-filtering method based on attenuation kernel matrix driving according to claim 5, characterized in that: The MSE loss function of the network is: In the formula: the wave field of the original data to be compensated The input is fed into the compensation module containing the U-Net network. For the decay kernel matrix forward modeling operator , To output the compensated data, This is a matrix dot product.

7. The unsupervised inverse Q-filtering method based on attenuation kernel matrix driving according to claim 1, characterized in that: In step four, the seismic data is input into the network module with optimal parameters to directly obtain the compensated seismic record.

Images