Passive source seismic data joint intelligent waveform inversion method based on physical guidance
By constructing a joint intelligent waveform inversion method for active and passive source seismic data under sparse acquisition conditions, and utilizing a two-stage neural network driven by data and guided by physics, the shortcomings of traditional full waveform inversion and pure data-driven methods are solved, and high-precision and robust underground velocity structure inversion is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-05-13
- Publication Date
- 2026-06-09
AI Technical Summary
Under sparse acquisition conditions, traditional full waveform inversion relies heavily on high-precision initial models and is prone to falling into periodic jump traps. Pure data-driven deep learning inversion lacks physical mechanism constraints, resulting in poor interpretability of solutions, weak generalization ability, and the existence of non-physical geological artifacts.
A physical-guided joint intelligent waveform inversion method based on active and passive source seismic data is constructed. By building a two-stage neural network driven by data and guided by physics, the nonlinear mapping driven by data is performed using the joint training set of active and passive sources, and the velocity model is optimized by combining the forward modeling residual of the wave equation as a physical constraint.
It achieves high-precision and robust underground velocity structure inversion, avoids periodic jump obstacles, improves the model's generalization ability and physical interpretability, and enhances the accuracy and reliability of the inversion results.
Smart Images
Figure CN122172290A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent seismic exploration technology, specifically involving a physical-guided intelligent waveform inversion method for joint active and passive source seismic data. Under sparse acquisition conditions, the method uses a joint training set of active and passive sources to drive a neural network to extract wavefield features, and combines the forward modeling residuals of the wave equation as physical constraints for joint guidance to achieve intelligent inversion of high-precision underground velocity models. Background Technology
[0002] Seismic exploration is a core method for acquiring subsurface geological structural features and conducting resource exploration, and the establishment of high-precision subsurface velocity models is the foundation for high-quality seismic imaging. While traditional active-source seismic exploration offers high signal quality and clear wavefield characteristics, it is limited by high acquisition costs and stringent environmental requirements. In contrast, passive-source seismic exploration utilizes widely available environmental background noise, offering advantages such as low cost, environmental friendliness, and a wide detection range. However, noisy passive source data often faces challenges such as unknown source locations and low signal-to-noise ratios. Therefore, neither a single active-source nor a single passive-source mechanism can meet the high-precision detection requirements of deep and complex geological structures under sparse acquisition conditions.
[0003] Among existing seismic velocity modeling methods, full waveform inversion (FWI), as a high-precision, physically driven inversion method, can theoretically obtain extremely high-resolution velocity structures. However, traditional FWI faces extremely severe ill-conditioned inversion problems. Its objective function is highly nonlinear. Although the wavefield characteristics of multi-source seismic data have certain complementarity, the inherent limitations of the original observation data still lead to problems such as period jumps that are difficult to overcome. In addition, this method is highly dependent on the accuracy of the initial velocity model and is accompanied by huge computational overhead for forward modeling iterations. In recent years, data-driven deep learning technology has been introduced into the field of velocity inversion, effectively avoiding dependence on the initial model through end-to-end nonlinear mapping. However, pure data-driven deep learning inversion is essentially a statistical fitting, lacking explicit constraints on the physical laws of seismic wave propagation. This leads to severe black-box effects in the network, poor physical interpretability of the inversion results, and weak generalization ability. When faced with non-ideal observation data such as sparse acquisition, it is prone to producing non-physical geological artifacts, and the dynamic characteristics deviate significantly from the real medium.
[0004] Therefore, to address the challenge of subsurface velocity inversion under sparse acquisition conditions, and to overcome the computational bottleneck of traditional full-waveform inversion (which is prone to periodic jumps) and the shortcomings of pure data-driven deep learning models (poor generalization ability and lack of physical mechanisms), an improved intelligent inversion method is urgently needed. This method should fully utilize the combined training set of active and passive sources to achieve complementary advantages. Based on a data-driven deep convolutional network, it should introduce forward modeling residuals such as wave equations as physical guiding constraints to intelligently optimize the high wavenumber details of the velocity model. This ensures that the predicted velocity structure is both highly efficient and strictly follows the physical dynamics of seismic wave propagation. Thus, it achieves a more robust and accurate intelligent inversion of subsurface velocity structure than traditional full-waveform inversion and pure data-driven deep learning methods. Summary of the Invention
[0005] The purpose of this invention is to provide a physical-guided joint intelligent waveform inversion method for active and passive source seismic data, in order to solve the problems of traditional full waveform inversion (FWI) under sparse acquisition conditions, which heavily relies on high-precision initial models and is prone to falling into periodic jump traps, and the lack of physical mechanism constraints in pure data-driven deep learning inversion, which leads to poor interpretability of solutions, weak generalization ability and the existence of non-physical geological artifacts.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] To achieve the above objectives, this invention provides a physically guided joint intelligent waveform inversion method for active and passive source seismic data, which specifically includes the following steps:
[0008] A. Preprocess the sparsely acquired active source seismic data and the raw passive source seismic data respectively;
[0009] B. Perform seismic interferometry reconstruction on the passive source seismic data preprocessed in step A, and then normalize it to obtain normalized passive source seismic data. Then, interpolate and encrypt the passive source seismic data with the active source seismic data according to the shot point position to form a joint shot gather record, which is used as training data for deep learning.
[0010] C. Construct a two-stage neural network structure driven by data and guided by physics: The first stage constructs a mapping network and trains the network through supervised learning to quickly establish a nonlinear mapping from the joint seismic dataset to the velocity model. The loss function adopted is the MSE loss function.
[0011] D. After obtaining the velocity model predicted in the first stage, the second stage uses forward modeling operators to perform real-time forward modeling simulation and uses finite difference to numerically solve the wave equation to synthesize the seismic record.
[0012] E. Subtract the synthetic seismic record from the input training seismic record, and calculate the data residual between the synthetic record and the observation data by constructing a normalized cross-correlation loss function;
[0013] F. Calculate the model domain residuals while calculating the data domain residuals, and then weight the two residual losses together to obtain the total loss function, which drives the model update.
[0014] G. Perform forward simulation on the current velocity model and obtain the joint loss function value: determine whether the joint residual meets the accuracy requirements. If it does, output the predicted velocity inversion result; if it does not, proceed to step D.
[0015] Further, in step A, after obtaining the active and passive source seismic data acquired in the field in a collinear manner, the original active source seismic data is preprocessed by regularization, denoising, and filtering, and the original passive source seismic data is preprocessed by filtering and spectral whitening.
[0016] Further, step B specifically involves the following formula for normalizing the passive source:
[0017] (1)
[0018] in, and These represent unnormalized active and passive source seismic data, respectively. For calculating data Norm, This represents the normalized passive source seismic data;
[0019] Further, in step E, the data residuals of the second stage specifically involve using normalized zero-latency cross-correlation as the physics-driven loss function:
[0020] (2)
[0021] in, For batch size, Indicates the first in the batch One sample. This is the loss value; the closer it is to 0, the more consistent the predicted waveform is with the observed waveform. and These represent simulation records and observation records, respectively. The formula for normalizing the cross-correlation coefficient is as follows:
[0022] (3)
[0023] in, The total number of sampling points for a single input data. Indicates the first One sampling point, and Computed data and observed data respectively Norm, and The calculated wavefield and the observed wavefield respectively represent the wavefield at the t-th... The specific amplitude value of each sampling point The inner product of two wave field vectors is given, and the theoretical maximum value of the cross-correlation coefficient for perfectly matched waveforms is 1.
[0024] Furthermore, in step F, to balance the consistency of statistical regularity of the solution with the self-consistency of physical wave theory, the overall objective function is designed as a weighted combination of supervised learning loss and physical constraint loss, expressed mathematically as follows:
[0025] (4)
[0026] The first term on the right-hand side of the equation represents the supervision constraint of the model domain, utilizing... Norm Minimization Network Prediction Speed Model Compared with real geological models The residuals between labels, and the second term characterizes the physical consistency constraints of the data domain. Represents the finite difference forward modeling operator based on the wave equation. These are actual earthquake observation records. The weighting factors represent the loss function. This is the total loss function.
[0027] Compared with the prior art, the beneficial effects of the present invention are:
[0028] This invention addresses the shortcomings of traditional full-waveform inversion and pure data-driven approaches by innovatively constructing a two-stage intelligent velocity inversion architecture jointly guided by a data-driven network and wave equations. This effectively overcomes the ill-posedness of subsurface medium inversion. The combination of active and passive source data not only enriches the diversity of training samples but also compensates for the deficiencies of passive source data through the high signal-to-noise ratio of active source data. At the same time, it enhances the detection capability of structures by leveraging the wide coverage advantage of passive source data. Deep learning can learn the nonlinear mapping relationships in complex seismic data, effectively combining the two data types of active and passive sources, thereby improving the accuracy of the inversion model in representing geological structures.
[0029] Specifically, it includes the following advantages:
[0030] 1. Traditional FWI is prone to getting trapped in local minima when faced with sparse acquisition or missing low-frequency components from passive sources. This invention, in its first stage, leverages a mapping deep neural network to fully utilize the high-frequency constraints of sparse, high-quality active source data and the rich low-frequency content of supplementary passive source data. This multi-source seismic data compensates for the inherent limitations of single-source data, providing more comprehensive wavefield information for inversion. Utilizing the network's powerful global nonlinear fitting and feature extraction capabilities, it can directly and rapidly map a low-wavenumber macroscopic background velocity model from the seismic record end-to-end. This strategy avoids the dependence on high-precision initial models in traditional physical inversion and avoids period jump obstacles.
[0031] 2. In the second stage, this invention innovatively cascades a fully differentiable forward modeling operator of the acoustic wave equation as a physical constraint layer, transforming the inversion problem into an optimization process that strictly follows the laws of seismic wave propagation dynamics. By utilizing physical residuals to drive the network to refine high wavenumber microstructural details, the output velocity model not only statistically conforms to the data regularity but also strictly satisfies wave theory in physics, significantly improving the model's generalization ability and reliability when facing real-world complex and heterogeneous geological structures.
[0032] 3. In the physical guidance stage, a physical objective function based on normalized cross-correlation, which is applicable to passive source data, is innovatively adopted. This mechanism avoids the interference of dynamic amplitude distortion, so that the update of network parameters is dominated by reliable kinematic phase information, thereby enhancing the inversion robustness. Attached Figure Description
[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0034] Figure 1 Flowchart of the steps of a sparsely acquired intelligent seismic velocity inversion method driven by a combination of active and passive source physical guidance data.
[0035] Figure 2 The nine velocity models in the training model of this invention are experimental models of the CurvedFault subset of the wrinkled non-uniform medium model randomly selected from the OpenFWI open source data.
[0036] Figure 3 Ideal distribution of earthquake source locations;
[0037] Figure 4Active source earthquake records and ideally distributed source virtual source records, where (a) is the active source earthquake reference record and (b) is the passive source virtual source record corresponding to the active source shot point location after earthquake interferometric reconstruction;
[0038] Figure 5 Deployment of active and passive sources under sparse acquisition conditions;
[0039] Figure 6 This invention is based on a physically guided data-driven joint inversion neural network architecture diagram;
[0040] Figure 7 The comparison diagram between the predicted results of this invention and the actual velocity models is shown in the figure. (a) is the actual velocity model one, (b) is the inversion result obtained by pure data-driven mapping in the first stage of model one, (c) is the inversion result obtained by adding the wave equation forward modeling operator to model one, (d) is the actual velocity model two, (e) is the inversion result obtained by pure data-driven mapping in the first stage of model two, (f) is the inversion result obtained by adding the wave equation forward modeling operator to model two, (g) is the actual velocity model three, (h) is the inversion result obtained by pure data-driven mapping in the first stage of model three, and (i) is the inversion result obtained by adding the wave equation forward modeling operator to model three.
[0041] Figure 8 A comparison diagram of the predicted results of this invention with the actual velocity model profile;
[0042] Figure 9 The comparison diagram of the prediction results of this invention with the forward modeling seismic records of the actual velocity model is shown in the figure. (a) is the seismic record input by the network, (b) is the forward modeling record of the pure data-driven inversion result in the first stage, and (c) is the forward modeling record of the combined data-driven and physical-guided inversion result in the first and second stages. Detailed Implementation
[0043] The present invention will be further described below with reference to embodiments:
[0044] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.
[0045] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0046] This invention addresses the problems of traditional full-waveform inversion, which easily falls into periodic jump traps and heavily relies on the initial model under actual sparse acquisition conditions, while pure data-driven deep learning inversion lacks physical mechanism constraints, resulting in weak model generalization ability and non-physical geological artifacts in the prediction results. It proposes a physically guided joint intelligent waveform inversion method based on active and passive source seismic data. First, a training model dataset is constructed by acquiring preprocessed sparse active and passive source seismic data. Seismic interferometry reconstruction is performed on the passive source data. The reconstructed passive and active source records are jointly used as input samples, and the corresponding subsurface velocity models are used as labels. The samples and labels are normalized separately to create a joint training dataset. Then, an intelligent inversion neural network model is established: an end-to-end mapping deep learning network is built to learn the nonlinear mapping relationship between complex seismic wavefields and subsurface velocity structures. Based on this, a fully differentiable wave equation forward modeling operator is cascaded to construct a physical constraint layer. Finally, the model is trained using a two-stage cascaded training strategy. The first stage is the data-driven stage, where the created training set is input into the neural network. By adaptively focusing on effective wavefield features, the network parameters are iteratively updated, and a low wavenumber background velocity model is directly output. The second stage is the physics-guided stage: the predicted velocity model is forward-modeled using a physical constraint layer. The normalized cross-correlation loss function is innovatively used to calculate the wavefield residuals. The residual gradient is backpropagated to drive the network to refine the microstructural details of high wavenumbers and eliminate geological artifacts that are detached from physical reality. Finally, the model is tested and applied. Using the trained physics-guided joint network model, sparse active and passive source seismic records to be tested (or actually acquired) are input for end-to-end prediction, completing intelligent velocity modeling along the seismic line and obtaining a high-resolution, highly robust, and high-precision underground velocity structure model that strictly follows the laws of seismic wave propagation dynamics.
[0047] The physical-guided joint intelligent waveform inversion method for active and passive source seismic data described in this invention is implemented using the MATLAB and PyCharm platforms, and specifically includes the following steps:
[0048] A. After obtaining the active and passive source seismic data acquired in the field in a collinear manner, the original active source seismic data is preprocessed by regularization, denoising, filtering, etc., and the original passive source seismic data is preprocessed by filtering, spectral whitening, etc.
[0049] B. Perform seismic interferometry reconstruction on the preprocessed passive source seismic data, and then normalize it to obtain normalized passive source seismic data. The data is then interpolated and encrypted with active source seismic data according to shot point location to form a joint shot gather record, which serves as training data for deep learning.
[0050] C. Construct a two-stage neural network structure driven by data and guided by physics. The first stage constructs a mapping network, which is trained through supervised learning to quickly establish a nonlinear mapping from the joint seismic dataset to the velocity model. The loss function adopted is the mean squared error (MSE) loss function.
[0051] D. After obtaining the velocity model predicted in the first stage, the second stage no longer relies on the pre-labeled velocity model data, but performs real-time forward modeling simulation through forward modeling operators and uses finite difference to numerically solve the wave equation to synthesize the seismic record.
[0052] E. Subtract the synthetic seismic record from the input training seismic record, and calculate the data residual between the synthetic record and the observed data by constructing a normalized cross-correlation loss function. ;
[0053] F. While calculating the data domain residuals, calculate the model domain residuals, and then weight the two residual losses together to obtain the total loss function. This drives model updates. The goal is to simultaneously consider both model features and physical characteristics during network weight optimization to obtain the currently updated model.
[0054] G. Perform forward simulation on the current velocity model and obtain the joint loss function value. Determine whether the joint residuals meet the accuracy requirements. If they do, output the predicted velocity inversion results; otherwise, proceed to step D.
[0055] Further, step B specifically involves the following formula for normalizing the passive source:
[0056] (1)
[0057] in, and These represent unnormalized active and passive source seismic data, respectively. For calculating data Norm, This represents the normalized passive source seismic data;
[0058] Further, in step E, the data residuals of the second stage specifically involve using normalized zero-latency cross-correlation as the physics-driven loss function:
[0059] (2)
[0060] in, For batch size, Indicates the first in the batch One sample. This is the loss value, i.e., the data residual. The closer it is to 0, the more consistent the predicted waveform is with the observed waveform. and These represent simulated records and observational records, respectively. The formula for normalizing the cross-correlation coefficient is as follows:
[0061] (3)
[0062] in, The total number of sampling points for a single input data. Indicates the first One sampling point, and Computed data and observed data respectively Norm, and The calculated wavefield and the observed wavefield respectively represent the wavefield at the t-th... The specific amplitude value of each sampling point The inner product of two wave field vectors is given, and the theoretical maximum value of the cross-correlation coefficient for perfectly matched waveforms is 1.
[0063] Furthermore, in step F, to balance the consistency of statistical regularity of the solution with the self-consistency of physical wave theory, the overall objective function is designed as a weighted combination of supervised learning loss and physical constraint loss, expressed mathematically as follows:
[0064] (4)
[0065] The first term on the right-hand side of the equation represents the supervision constraint of the model domain, utilizing... Norm Minimization Network Prediction Speed Model Compared with real geological models The residuals between labels aim to guide the network to converge quickly to the attraction domain of the global optimum and construct an accurate low-wavenumber velocity background; the second term characterizes the physical consistency constraints of the data domain. Represents the finite difference forward modeling operator based on the wave equation. These are actual earthquake records observed. The weighting factors represent the loss function. This is the total loss function.
[0066] Example 1
[0067] A physically guided joint intelligent waveform inversion method for active and passive source seismic data includes the following steps, such as... Figure 1 As shown.
[0068] The actual velocity model used is as follows: Figure 2As shown, each graph represents a different velocity structure, with significant horizontal variability reflecting the lateral heterogeneity of the subsurface medium. This is an experimental model based on the OpenFWI open-source dataset, using a subset of the CurvedFault model (a non-uniform medium with folds) for training. The models are all 70×70 grids with a grid spacing of 10m, and velocities ranging from 1500m / s to 4500m / s. Seismic detectors are placed on the model surface, with a detector at each grid point (10m interval). The seismic source locations are set as follows... Figure 3 As shown. The passive source wavelet is set with a random noise sequence with a maximum frequency of 30Hz. The sampling interval is 0.001s. The active source wavelet uses a real source function with a dominant frequency of 15Hz, the Ricker wavelet.
[0069] Active source earthquake records of surface blasting, such as Figure 4 As shown in Figure a, not only is the waveform continuous and free from coherent noise interference, but it also presents reflected wave information with correct shape and position. After multidimensional deconvolution, the obtained noise record yields a passive virtual source record corresponding to the active source location, as shown in Figure a. Figure 4 As shown in b, it can be clearly observed that the reconstructed virtual source seismic record exhibits coherent noise and false phase axes compared to the active source record.
[0070] In this embodiment, the passive virtual source data is normalized and then combined with the active source data to form a training set input on a sideline. The distribution pattern of the active and passive sources is as follows: Figure 5 As shown, a joint inversion was performed using 5-shot active source seismic records and 5-shot passive source records.
[0071] Figure 6 This is the inversion network framework of the present invention.
[0072] The physical-guided data-driven intelligent seismic velocity inversion network framework adopts a cascaded two-stage architecture: In the first stage, the data-driven layer, the network takes joint active and passive source seismic records as input and, relying on the mapping backbone network, rapidly maps a low wavenumber background velocity model that avoids the periodic jump trap by adaptively focusing on the spatial and channel characteristics of the effective wavefield. In the second stage, the physical constraint layer, a fully differentiable acoustic equation forward modeling operator is seamlessly cascaded at the mapping output to perform forward modeling simulation on the predicted initial velocity. It innovatively uses a normalized cross-correlation objective function to quantify the kinematic phase residual between the simulated wavefield and the actual observed wavefield. Then, the backpropagation of the residual gradient drives the network to refine the high wavenumber microstructural details of the velocity model. Finally, a high-precision intelligent imaging closed-loop system for underground velocity is constructed, which has the advantages of deep learning nonlinear rapid fitting and strictly follows the propagation law of seismic wave dynamics.
[0073] Figure 7 To invert and predict the effect, Figure 7 'a' represents the true velocity model in the first prediction model. The inversion results show that, even with significant differences in the horizontal velocity structure, the approximate stratigraphic information of the first-stage inversion results can be matched. Figure 7 As shown in b, the velocity model in the second stage has a higher resolution, but the velocity representation of the stratigraphic information is incorrect, such as... Figure 7 As shown in c. Figure 7 d represents the true velocity model in the second prediction model. In the inversion results, because the difference in the horizontal direction of the velocity structure is smaller than that of the first model, and the low-frequency information is abundant, the inversion results of the first stage can also recover the velocity accuracy well. Figure 7 As shown in e, the second step provides improved layer information compared to the first model, but the deep high-velocity volume cannot be recovered well, such as... Figure 7 As shown in f. Figure 7 g represents the true velocity model in the third prediction model. Given that the difference in the horizontal direction of the velocity structure is significantly smaller than in the first two models, it can accurately characterize stratigraphic information in the first stage, but its resolution is lower. Figure 7 As shown in h, the second stage can clearly express the shallow and mid-to-deep stratigraphic information of the velocity model, such as... Figure 7 As shown in i.
[0074] Figure 8 Yes Figure 7 The predicted velocity model is plotted as a single-track velocity profile, with the vertical velocity profile positioned at a horizontal location of x=0.3km. In the figure, the blue line represents the actual velocity values from the true velocity profile, the red line represents the predicted velocity values from the inversion results obtained through pure data-driven mapping, and the yellow line represents the predicted velocity values from the two-stage inversion results combining data-driven and wave equation physics-guided approaches. Comparing the results, it can be found that the two-stage inversion results based on physics guidance, i.e., the yellow line, have a higher degree of overlap with the actual velocity model, i.e., the blue line, especially in deeper parts of the structure, where the yellow and blue lines fit together more closely. Since the physics-guided inversion results are based on and modified from the pure mapping inversion results, a wave equation physics-guided network is added to areas where the pure mapping velocity predictions (i.e., the red line) are not well-fitted, in order to fit the actual velocity values as closely as possible.
[0075] Figure 9 This is a comparison chart of the prediction results of the first model and the forward modeled seismic records of the actual velocity model. Figure 9 'a' represents the input observation data and the passive source virtual shot record. Figure 9 b is Figure 9 a is the forward modeling record of the pure data-driven inversion results corresponding to the shot point location. Figure 9 c is Figure 9Figure 'a' shows the forward modeling record of the data-driven and physics-guided joint inversion results corresponding to the shot point location. As can be seen from the figure, the seismic record obtained through the wave equation-guided inversion results has richer waveform information, and the waveform residuals are reduced to some extent compared to the pure data-driven algorithm, indicating that the physics-guided inversion network can extract more data features.
[0076] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.
Claims
1. A physically guided joint intelligent waveform inversion method for active and passive source seismic data, characterized in that, Includes the following steps: A. Preprocess the sparsely acquired active source seismic data and the raw passive source seismic data respectively; B. After performing seismic interferometry reconstruction on the passive source seismic data preprocessed in step A, normalize the data to obtain normalized passive source seismic data. Then, interpolate and encrypt the data with the active source seismic data according to the shot point location to form a joint shot gather record, which is used as training data for deep learning. C. Construct a two-stage neural network structure driven by data and guided by physics: The first stage constructs a mapping network and trains the network through supervised learning to quickly establish a nonlinear mapping from the joint seismic dataset to the velocity model. The loss function adopted is the MSE loss function. D. After obtaining the velocity model predicted in the first stage, the second stage uses forward modeling operators to perform real-time forward modeling simulation and uses finite difference to numerically solve the wave equation to synthesize the seismic record. E. Subtract the synthetic seismic record from the input training seismic record, and calculate the data residual between the synthetic record and the observation data by constructing a normalized cross-correlation loss function; F. Calculate the model domain residuals while calculating the data domain residuals, and then weight the two residual losses together to obtain the total loss function, which drives the model update. G. Perform forward simulation on the current velocity model and obtain the joint loss function value: determine whether the joint residual meets the accuracy requirements. If it does, output the predicted velocity inversion result; if it does not, proceed to step D.
2. The method for joint intelligent waveform inversion of physically guided active and passive source seismic data according to claim 1, characterized in that, Step A: After obtaining the active and passive source seismic data acquired in the field using collinear acquisition, the original active source seismic data is preprocessed by regularization, denoising, and filtering, while the original passive source seismic data is preprocessed by filtering and spectral whitening.
3. The method for joint intelligent waveform inversion of physically guided active and passive source seismic data according to claim 1, characterized in that, Step B, the formula for normalizing the passive source is as follows: （1） in, and These represent unnormalized active and passive source seismic data, respectively. For calculating data Norm, This represents the normalized passive source seismic data.
4. The method for joint intelligent waveform inversion of physically guided active and passive source seismic data according to claim 1, characterized in that, Step E, the second stage of data residuals, specifically utilizes normalized zero-latency cross-correlation as the physics-driven loss function: （2） in, For batch size, Indicates the first in the batch One sample, This is the loss value; the closer it is to 0, the more consistent the predicted waveform is with the observed waveform. and These represent simulation records and observation records, respectively. The formula for normalizing the cross-correlation coefficient is as follows: （3） in, The total number of sampling points for a single input data. Indicates the first One sampling point, and Computed data and observed data respectively Norm, and The calculated wavefield and the observed wavefield respectively represent the wavefield at the t-th... The specific amplitude value of each sampling point The inner product of two wave field vectors is given, and the theoretical maximum value of the cross-correlation coefficient for perfectly matched waveforms is 1.
5. The method for joint intelligent waveform inversion of physically guided active and passive source seismic data according to claim 1, characterized in that, Step F: The overall objective function is designed as a weighted combination of supervised learning loss and physical constraint loss, and its mathematical expression is as follows: （4） The first term on the right-hand side of the equation represents the supervision constraint of the model domain, utilizing... Norm Minimization Network Prediction Speed Model Compared with real geological models The residuals between labels, and the second term characterizes the physical consistency constraints of the data domain. Represents the finite difference forward modeling operator based on the wave equation. These are actual earthquake observation records. The weighting factors represent the loss function. This is the total loss function.