Seismic wave impedance inversion method and system based on combination of CNN and transformer

Abstract

The application discloses a seismic wave impedance inversion method and system based on a CNN and a Transformer, and the method comprises the following steps: constructing an inversion network, fusing global features and local features through a residual connection, and outputting a wave impedance prediction result; constructing a forward operator, and mapping the wave impedance into synthetic seismic data by solving a first-order velocity-stress acoustic wave equation in a differentiable finite difference format; constructing a time-varying wavelet model of multi-base time-varying weighted combination, and the base wavelet main frequency and the time-varying mixed weight are all learnable parameters; calculating a wave impedance loss, a seismic loss, a physical constraint loss and a wavelet regularization loss, jointly updating all parameters through back propagation, jointly optimizing the wavelet parameters and the inversion network parameters by constructing the time-varying wavelet model of multi-base time-varying weighted combination, and directly outputting the time-varying wavelet estimation of a target work area after training, so that the wavelet does not need to be manually extracted.

Description

Technical Field

[0001] This invention relates to the field of seismic exploration data processing technology, specifically to a seismic impedance inversion method and system based on a combination of CNN and Transformer. Background Technology

[0002] With the development of deep learning technology, neural network-based methods have been introduced into the field of seismic impedance inversion. Compared with traditional inversion methods based on model optimization or convolution theory, neural network methods can automatically learn the mapping relationship between seismic data and wave impedance from the data, reducing the dependence on prior geological models.

[0003] In seismic impedance inversion, the forward modeling operator is used to convert the predicted wave impedance into synthetic seismic data so that it can be compared with the actual seismic data, thus forming a semi-supervised training closed loop.

[0004] In existing technologies, forward modeling operators employ explicit wavelet convolution models, which generate synthetic seismic records through the convolution of known seismic wavelets and reflection coefficients. However, this method has the following problems: seismic wavelets are often unknown in the actual work area and need to be estimated by additional means, and the estimation error will be directly transmitted to the inversion results; furthermore, the convolution model is based on the one-dimensional vertical incidence assumption and does not consider wave field propagation effects, making it difficult to characterize complex wave field phenomena such as multiples and absorption attenuation.

[0005] In existing technologies, forward modeling operators use pure convolutional neural networks to learn the mapping from wave impedance to seismic data. This method does not require explicit extraction of wavelets, but it has the following problems: pure data-driven forward modeling operators lack physical constraints, and the mapping relationship they learn may violate the physical laws of the wave equation. They are prone to overfitting and have insufficient generalization ability when training data is limited.

[0006] Furthermore, existing methods typically assume that seismic wavelets are time-invariant, meaning that the same wavelet is used throughout the entire seismic trace depth. However, in actual seismic data, due to the absorption and attenuation effects of the strata, seismic wavelets are time-varying: shallow wavelets have higher dominant frequencies and wider bandwidths, while deep wavelets have lower dominant frequencies and narrower bandwidths. The time-invariant assumption leads to a decrease in the accuracy of deep inversion and fails to characterize the physical characteristics of wavelet variations with depth. Summary of the Invention

[0007] This invention proposes a seismic wave impedance inversion method and system based on the combination of CNN and Transformer to solve the problems mentioned in the background.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: The seismic wave impedance inversion method based on the joint CNN and Transformer of the present invention includes the following steps: S1, acquire seismic data of the target work area and logging impedance data of at least one well; S2, construct an inversion network, fuse the global and local features of the seismic data through residual connections, and output the wave impedance prediction results; S3, Construct a forward modeling operator, which maps wave impedance to synthetic seismic data by numerically solving the discrete wave equation; S4, Construct a time-varying wavelet model; S5. Calculate wave impedance loss based on the well logging wave impedance data and the wave impedance prediction results; calculate seismic loss based on the synthetic seismic data output by the forward model and the seismic data; calculate physical constraint loss based on the discrete residual of the wave equation; and calculate wavelet regularization loss based on the parameters of the time-varying wavelet model. S6, the wave impedance loss, the seismic loss, the physical constraint loss and the wavelet regularization loss are weighted and combined into a total loss, and the inversion network parameters, the forward modeling operator parameters and the time-varying wavelet model parameters are jointly updated through backpropagation; S7, under the condition that the total loss converges, output the final wave impedance prediction result of the inversion network and the wavelet estimation result of the time-varying wavelet model.

[0009] Preferably, the neural network correction component is used to learn the residual correction term of the physical forward modeling component, wherein the numerical solution process of the physical forward modeling component is implemented in a differentiable finite difference scheme, so that the physical constraints of the wave equation are satisfied at the architecture level. The physical forward modeling component maps wave impedance to synthetic seismic data by numerically solving discrete wave equations, including one or more of the following: The acoustic wave equation adopts the first-order velocity-stress form, and the acoustic wave equation is as follows:

[0010]

[0011] in, Let the velocity of the particle be... For sound pressure, For density, Bulk modulus For spatial coordinates, For time, the speed of sound wave impedance ; The acoustic wave equations are discretized using an interleaved grid finite difference scheme to obtain the discrete equations:

[0012]

[0013] in, For spatial grid indexing, For time step index, For spatial step size, For time step; The output of the physical forward modeling component For the sound pressure recording at the receiving point, the final output of the forward operator is:

[0014] in, The residual correction term learned by the neural network correction component.

[0015] Preferably, the physical constraint loss is calculated based on the discrete residuals of the wave equation, including: The final output of the forward operator Substituting into the residual expression of the discrete wave equation:

[0016] in, Let be the discrete differential operator matrix of the wave equation. The source term matrix is ​​related to wave impedance. Wave impedance; The physical constraint loss is:

[0017] in, These are the weighting coefficients of the residuals in the wave equation. The L2 regularization coefficient of the neural network correction term; During the training process, Dynamically adjust based on the current training error: when the training error exceeds a preset threshold, reduce... This allows the neural network correction component to fit flexibly; when the training error is less than the preset threshold, the value is increased. This causes the forward operator to converge to the vicinity of the physical forward component.

[0018] Preferably, the time-varying wavelet model is a time-varying weighted combination of multiple fundamental wavelets with different dominant frequencies, including:

[0019] in, This is the time offset relative to the center of the wavelet. The depth during the two-way travel of the earthquake path. For the first A fundamental wavelet, For the first The dominant frequency parameters of each fundamental wavelet are learnable parameters. For the first The fundamental wavelet at depth The time-varying hybrid weights at the point are learnable parameters, and K is the number of fundamental wavelets.

[0020] Preferably, the dominant frequency parameter and the time-varying mixing weight of the fundamental wavelet are both learnable parameters, and the time-varying mixing weight is a function of the two-way travel time depth, used to characterize the time-varying characteristics of the wavelet caused by the formation absorption attenuation effect. The fundamental wave Using Ricker wavelet form:

[0021] The time-varying hybrid weights are parameterized using the softmax function:

[0022] in, and As learnable parameters, this parameterization method guarantees ; The wavelet regularization loss includes the smoothness constraint and the time-varying weight continuity constraint of the time-varying wavelet model:

[0023] in, For the total variation of the wavelet, and These are the weighting coefficients. The distance between the two sides in the depth direction is the distance between the two sides.

[0024] Preferably, the inversion network includes a Transformer encoder and a multi-scale convolutional neural network, wherein the Transformer encoder is used to extract global features of the seismic data, and the multi-scale convolutional neural network is used to extract local features of the seismic data.

[0025] Preferably, the inversion network includes: The Transformer encoder, comprising a multi-head self-attention layer and a feedforward network layer, is used to extract global dependency features after location encoding of the input seismic data. The multi-head self-attention mechanism is calculated as follows:

[0026] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively. The dimension of the key vector; Multi-scale convolutional neural networks, which include multiple branches, each using dilated convolution with different dilation coefficients, are used to extract local features of seismic data at different scales. The residual connection module is used to fuse the output of the Transformer encoder and the output of the multi-scale convolutional neural network by element-wise addition, and output the wave impedance prediction result.

[0027] Where X represents the input seismic data. For the mapping function of the Transformer encoder, This is the mapping function for a multi-scale convolutional neural network.

[0028] Preferably, the total loss is:

[0029] in, The mean square error is used to represent the wave impedance loss. The earthquake loss is calculated using the mean square error between the synthetic seismic data and the original seismic data. This refers to the physical constraint loss; The wavelet regularization loss; , , , These are the weighting coefficients for each loss term.

[0030] Preferably, when the total loss converges, the final wave impedance prediction result of the inversion network and the wavelet estimation result of the time-varying wavelet model are output, including one or more of the following: The seismic data of the target work area is input into the trained inversion network, and the wave impedance profile of the target work area is output. Output the time-varying hybrid weights at each depth τ in the time-varying wavelet model after training. and the dominant frequency parameters of each fundamental wavelet , as the wavelet estimation result of the target work area; The wavelet estimation results include time-varying characteristics where the dominant frequency of the shallow wavelet is higher than that of the deep wavelet.

[0031] A seismic impedance inversion system based on a combination of CNN and Transformer includes: The data acquisition unit is used to acquire seismic data of the target work area and logging impedance data of at least one well. The inversion network unit is used to extract features and output wave impedance prediction results through a Transformer encoder and a multi-scale convolutional neural network. Forward modeling operator unit, used to map wave impedance to synthetic seismic data through physical forward modeling components and neural network correction components; Time-varying wavelet units are used to construct time-varying wavelet models and output wavelet estimation results; The training unit is used to calculate various losses and jointly update parameters.

[0032] As can be seen from the above technical solution, this invention provides a seismic wave impedance inversion method based on the joint use of CNN and Transformer. Compared with the prior art, this invention has the following advantages: This invention embeds the wave equation into the physical forward modeling component of the forward modeling operator in the form of a differentiable finite difference solver, so that the output of the forward modeling operator naturally satisfies the physical constraints of the wave equation, rather than only using soft constraints through the loss function; the hard constraints at the architecture level enable the forward modeling operator to maintain physical rationality even with limited training data; furthermore, by constructing a time-varying wavelet model with a time-varying weighted combination of multi-basic wavelets, the wavelet parameters and the inversion network parameters are jointly optimized, and the time-varying wavelet estimate of the target work area is directly output after training, without the need for manual extraction of wavelets; in addition, the parallel fusion architecture of multi-scale convolutional neural network and Transformer encoder takes into account the modeling capabilities of local detailed features and global dependencies. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the seismic impedance inversion method based on the combined CNN and Transformer of this invention.

[0034] Figure 2 This is a diagram of the joint inversion network structure of the present invention.

[0035] Figure 3 This is a diagram of the physical constraint forward modeling operator structure of the present invention.

[0036] Figure 4 This is a structural diagram of the time-varying wavelet model of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0038] like Figure 1As shown, the seismic impedance inversion method based on the joint CNN and Transformer in this embodiment includes the following steps: S1: Obtain seismic data of the target work area and well logging impedance data of at least one well; S2: Construct an inversion network; S3: Construct a forward modeling operator; S4: Construct a time-varying wavelet model; S5: Calculate various losses; S6: Jointly update parameters; S7: Output the final result.

[0039] S1: Acquire seismic data for the target work area and acoustic impedance logging data from at least one well. In one possible implementation, the seismic data is post-stack seismic profile data for the target work area, and the acoustic impedance logging data is the acoustic impedance logging curve of at least one drilled well within the target work area. Preprocess the seismic data, including bandpass filtering, amplitude equalization, and phase correction. Resample the acoustic impedance logging data to match the sampling rate of the seismic data.

[0040] In addition, seismic traces near the well were extracted and paired with well logging impedance data to form a labeled training dataset; the remaining seismic traces were used as an unlabeled dataset.

[0041] S2: Construct the inversion network; such as... Figure 2 As shown, the inversion network includes a Transformer encoder and a multi-scale convolutional neural network.

[0042] In one possible implementation, the Transformer encoder includes six multi-head self-attention layers with eight attention heads, a hidden layer dimension of 512, and a feedforward network dimension of 2048; for the input seismic trace data... Where L is the number of time sampling points, first a linear projection is performed to The dimensional features are then applied with sine-cosine position encoding, and then passed sequentially through 6 Transformer encoder layers to output a global feature representation. The calculation of the multi-head self-attention mechanism is as follows:

[0043] in, These are the query matrix, key matrix, and value matrix, respectively. is the dimension of the key vector.

[0044] MSCNN comprises multiple branches, each employing dilated convolutions with different dilation coefficients to extract local features at different scales. In one possible implementation, MSCNN includes three branches: Branch 1 uses a convolution with a dilation coefficient of 1, a kernel size of 3, and a receptive field of 3 sampling points to extract fine local features; Branch 2 uses a convolution with a dilation coefficient of 2, a kernel size of 3, and a receptive field of 5 sampling points to extract medium-scale local features; and Branch 3 uses a convolution with a dilation coefficient of 4, a kernel size of 3, and a receptive field of 9 sampling points to extract large-scale local features. The outputs of the three branches are concatenated along the channel dimension and then reduced to 512 dimensions using a 1×1 convolution to obtain the local feature representation. The residual connection module fuses the output of the Transformer encoder with the output of MSCNN by element-wise addition.

[0045] Where X represents the input seismic data. For the mapping function of the Transformer encoder, This is the mapping function for a multi-scale convolutional neural network.

[0046] Understandably, the multi-head self-attention mechanism of the Transformer encoder can capture the global dependency between any two sampling points on the seismic trace, overcoming the limitations of the convolutional network's limited receptive field and the long-distance decay of the recurrent network; the multi-branch dilated convolutional structure of MSCNN can acquire multi-scale local features with fewer parameters; the parallel residual fusion of the two enables the inversion network to have both global modeling capabilities and local detail capture capabilities.

[0047] S3: Constructing forward operators; such as... Figure 3 As shown, the forward operator includes a physical forward modeling component and a neural network correction component. In this embodiment, the first-order velocity-stress form of the acoustic wave equation is embedded in the forward operator architecture in the form of a differentiable finite difference solver, so that the physical constraints of the forward operator are satisfied at the architecture level, rather than just through soft constraints of the loss function.

[0048] The physical forward modeling component uses a first-order velocity-stress form of the acoustic wave equation:

[0049]

[0050] in, Let the velocity of the particle be... For sound pressure, For density, Bulk modulus For spatial coordinates, For time, the speed of sound wave impedance ; based on wave impedance prediction results The density ρ and bulk modulus K can be derived given the wave impedance Z and the sound velocity c. .

[0051] The acoustic wave equations are discretized using an interleaved grid finite difference scheme:

[0052]

[0053] in, For spatial grid indexing, For time step index, For spatial step size, For time step.

[0054] It should be noted that this finite difference scheme is implemented in a differentiable form, meaning that all operations can be backpropagated using an automatic differentiation framework; the output of the physical forward modeling component... For the sound pressure recording at the receiving point, the final output of the forward operator is:

[0055] in, The residual correction term learned by the neural network correction component.

[0056] The neural network correction component is a lightweight convolutional network used to learn complex wave field effects that the physical forward modeling component cannot model, including multiples, absorption attenuation, anisotropy, etc.

[0057] In one possible implementation, the neural network correction component comprises three 1D convolutional layers, each with a kernel size of 3 and 64 channels, followed by a ReLU activation function; its input is the wave impedance prediction result. The output is the same as Residual correction terms of the same dimension .

[0058] It should be noted that the standard PINN uses the residuals of the physical equations as soft constraints in the loss function, indirectly guiding the network output to satisfy physical laws through gradient descent. However, the network architecture itself is still a purely data-driven black box, and may still output results that violate physical laws when training data is limited. In this embodiment, the finite difference solver of the wave equation is used as the physical forward modeling component of the forward modeling operator. The output of this solver is directly determined by the discrete format of the wave equation, naturally satisfying physical constraints, which is a hard constraint at the architecture level. The neural network correction component only learns the residual correction term of the physical forward modeling component and is subject to additional constraints from the physical constraint loss. Therefore, even if the correction component has biases, the overall behavior of the forward modeling operator will not deviate significantly from physical laws.

[0059] Furthermore, the differentiable finite difference solver enables the gradient of the wave impedance parameter to be directly propagated to the inversion network through the wave equation solution process, resulting in a shorter and more stable gradient path.

[0060] S4: Construct a time-varying wavelet model; such as Figure 4 As shown, the time-varying wavelet model is a time-varying weighted combination of multiple fundamental wavelets with different dominant frequencies:

[0061] in, This is the time offset relative to the center of the wavelet. The depth during the two-way travel of the earthquake path. For the first A fundamental wavelet, For the first The dominant frequency parameters of each fundamental wavelet are learnable parameters. For the first The fundamental wavelet at depth The time-varying hybrid weights at the point are learnable parameters, and K is the number of fundamental wavelets.

[0062] In one possible implementation, K=5, and the initial dominant frequencies of the five fundamental wavelets are 10Hz, 20Hz, 30Hz, 40Hz, and 50Hz, respectively.

[0063] Fundamental wave Using Ricker wavelet form:

[0064] Ricker wavelet is a commonly used zero-phase wavelet model in seismic exploration. Its shape is determined by the dominant frequency. The only certainty.

[0065] The time-varying hybrid weights are parameterized using the softmax function:

[0066] in, and These are learnable parameters; this parameterization method guarantees... That is, the sum of the mixing weights at any depth is always 1. During training, the time-varying mixing weights automatically learn the contribution ratio of different dominant frequency fundamental wavelets at each depth. It can be understood that due to the formation absorption and attenuation effect, the dominant frequency of wavelets in shallow seismic data is higher, and the dominant frequency of wavelets in deep seismic data is lower; the time-varying wavelet model in this embodiment can automatically learn this physical trend: driven by training data in shallow layers (where τ is smaller), the mixing weights corresponding to high dominant frequency fundamental wavelets... Automatically increases; driven by deep (larger τ) training data, the hybrid weights corresponding to the low-frequency fundamental wavelet... Automatically increases in size.

[0067] It should be noted that the learnable filter bank in speech processing is usually a fixed-frequency bandpass filter whose parameters are independent of time and is used to perform frequency domain decomposition on the entire speech signal. The time-varying wavelet model in this embodiment uses the Ricker wavelet in seismic exploration as the basis function, and its time-varying mixing weight is a function of the two-way travel time depth, which directly describes the time-varying physical characteristics of the wavelet caused by the formation absorption attenuation effect and has clear physical meaning. Furthermore, the time-varying wavelet model is not an independent signal processing module, but a component embedded in the forward model and jointly trained with the inversion network. Its parameters are automatically calibrated by seismic loss during the end-to-end training process.

[0068] S5: Calculate the various losses; the loss function in this embodiment includes four items: wave impedance loss. Earthquake losses Physical constraint loss Sum wavelet regularization loss .

[0069] Wave impedance loss The mean squared error is used to constrain the prediction accuracy of the inversion network on labeled data.

[0070] in, The number of labeled data. The wave impedance prediction results output by the inversion network, This is logging impedance data.

[0071] earthquake losses To constrain the consistency between the output of the forward modeling operator and the actual seismic data, the mean square error is used.

[0072] in, For the total number of data points, The synthetic seismic data output by the forward modeling operator. The data used is actual earthquake data; earthquake loss is the core of semi-supervised training, enabling unlabeled earthquake data to participate in training through a forward-inversion closed loop.

[0073] Physical constraint loss This includes wave equation residual constraints and neural network correction term regularization:

[0074] in, The discrete residuals of the wave equation are... Let S be the discrete differential operator matrix of the wave equation, and let S be the source term matrix related to wave impedance. These are the weighting coefficients of the residuals in the wave equation. is the L2 regularization coefficient of the neural network correction term.

[0075] During the training process, Dynamically adjust based on the current training error: when the training error exceeds a preset threshold, reduce... This allows the neural network correction components to fit flexibly; when the training error is less than a preset threshold, the [function / mechanism] is increased. This is to make the forward operator converge to the vicinity of the physical forward component.

[0076] Wavelet regularization loss This includes wavelet smoothness constraints and time-varying weighted continuity constraints:

[0077] in, For the total variation of the wavelet, and These are the weighting coefficients. The distance between the two sides in the depth direction is the distance between the two sides.

[0078] Wavelet smoothness constraints prevent non-physical high-frequency oscillations of the wavelet during training; time-varying weight continuity constraints prevent abrupt changes in the mixed weights in the depth direction, ensuring the continuity of the wavelet's time-varying process.

[0079] S6: Weight the four losses to form the total loss, and update all parameters jointly through backpropagation:

[0080] in, , , , These are the weighting coefficients for each loss term. In one possible implementation, ; By backpropagation, the parameters of the inversion network (weights of the Transformer encoder and MSCNN) and the parameters of the forward operators (weights of the neural network correction components) are updated simultaneously. and Dynamic adjustment strategy parameters) and time-varying wavelet model parameters (fundamental wavelet dominant frequency) Time-varying hybrid weight parameters and It should be noted that, since the differentiable finite difference solver of the physical forward modeling component satisfies the wave equation at the architecture level, the gradient can be directly propagated to the wave impedance parameter through the solution process of the wave equation, resulting in a shorter and more stable gradient path.

[0081] S7: Upon convergence of the total loss, output the final wave impedance prediction results of the inversion network and the wavelet estimation results of the time-varying wavelet model; input all seismic data of the target work area into the trained inversion network, and output the wave impedance profile of the target work area; simultaneously, output the wave impedance profiles of each depth in the trained time-varying wavelet model. Time-varying hybrid weights at the point and the dominant frequency parameters of each fundamental wavelet The wavelet estimation results for the target work area include the time-varying characteristics of the shallow wavelet dominant frequency being higher than that of the deep wavelet dominant frequency.

[0082] In one possible implementation, the target work area is 50 square kilometers, the seismic data sampling rate is 2ms, the recording length is 3 seconds, and there are a total of 1500 sampling points; there are 3 drilled wells in the work area, and the depth range of well logging impedance data is 1000-3000ms; the seismic traces next to the wells of the 3 wells are paired with the well logging impedance data to form a labeled training set, and the remaining seismic traces are used as an unlabeled dataset; the inversion network has a Transformer encoder with 6 layers, 8 attention heads, and a hidden layer dimension of 512; the MSCNN includes three dilated convolution branches with dilation coefficients of 1, 2, and 4; the time-varying wavelet model contains 5 Ricker basis wavelets with initial dominant frequencies of 10Hz, 20Hz, 30Hz, 40Hz, and 50Hz.

[0083] During training, the Adam optimizer was used with an initial learning rate of 0.001, and a cosine annealing strategy was employed to reduce the learning rate; the total number of training rounds was 300. The dynamic adjustment strategy is as follows: when the training error is large in the first 100 rounds, This allows the neural network correction component to fit flexibly; after 100 rounds, when the training error drops below a preset threshold, Gradually increase it to 0.1 to make the forward operator converge to the vicinity of the physical forward component.

[0084] After training, the average correlation coefficient between the wave impedance profile output by the inversion network and the well logging wave impedance data of the three wells reached 0.92. The wavelet estimation results of the time-varying wavelet model showed that the dominant fundamental wavelet frequency in the shallow layer (1000-1500 ms) converged to approximately 45 Hz, in the middle layer (1500-2200 ms) to approximately 35 Hz, and in the deep layer (2200-3000 ms) to approximately 28 Hz. This time-varying trend is consistent with the wavelet frequency variation trend extracted by commercial seismic processing software through spectral analysis (approximately 46 Hz in the shallow layer, approximately 34 Hz ​​in the middle layer, and approximately 29 Hz in the deep layer), verifying the adaptive estimation capability of the time-varying wavelet model.

[0085] Furthermore, to verify the effectiveness of the hard constraint on the wave equation in the forward modeling operator of this embodiment, a comparative experiment was conducted. The comparative scheme adopted the standard PINN method, where the residual of the wave equation is used only as a soft constraint term in the loss function, and the forward modeling operator is a pure CNN structure. With only one well as training data (a scenario with limited data), the average correlation coefficient between the wave impedance profile and the well logging data of the comparative scheme was 0.78, while the average correlation coefficient of the scheme in this embodiment was 0.88. With three wells as training data, the average correlation coefficient of the comparative scheme was 0.86, while the average correlation coefficient of the scheme in this embodiment was 0.92. This result demonstrates that the hard constraint on the wave equation at the architecture level has a significant advantage in generalization ability when training data is limited.

[0086] It is understood that the technical effects of this embodiment include the following four aspects. First, the forward operator embeds the wave equation into the architecture level through a differentiable finite difference solver, achieving physical hard constraints. Even with limited training data, it maintains physical rationality and generalization ability, overcoming the problems of pure CNN forward operators lacking physical constraints and standard PINN relying solely on soft constraints. Second, the time-varying wavelet model automatically characterizes the time-varying features of wavelets caused by formation absorption attenuation through a multi-basic wavelet time-varying weighted combination, eliminating the need for manual wavelet extraction and directly outputting physically meaningful wavelet estimation results after training. Third, the inversion network, through parallel residual fusion of a Transformer encoder and a multi-scale convolutional neural network, simultaneously captures global dependencies and local detailed features, overcoming the limitations of a single network architecture. Fourth, the inversion network, forward operator, and time-varying wavelet model are jointly trained end-to-end, with the parameters of each component mutually calibrated, forming a physically consistent inversion system.

[0087] On the other hand, this embodiment also provides a seismic impedance inversion system based on a combination of CNN and Transformer, characterized by including units for implementing the above-described method. The system may include: a data acquisition unit for acquiring seismic data of the target work area and well logging impedance data of at least one well; an inversion network unit for extracting features and outputting impedance prediction results through a Transformer encoder and a multi-scale convolutional neural network; a forward modeling operator unit for mapping impedance to synthetic seismic data through a physical forward modeling component and a neural network correction component; a time-varying wavelet unit for constructing a time-varying wavelet model and outputting wavelet estimation results; and a training unit for calculating various losses and jointly updating parameters.

[0088] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the flow or function according to this embodiment is generated, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk).

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0090] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0091] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A seismic wave impedance inversion method based on the joint use of CNN and Transformer, characterized in that, Includes the following steps: S1, acquire seismic data of the target work area and logging impedance data of at least one well; S2, construct an inversion network, fuse the global and local features of the seismic data through residual connections, and output the wave impedance prediction results; S3, Construct a forward modeling operator, which maps wave impedance to synthetic seismic data by numerically solving the discrete wave equation; S4, Construct a time-varying wavelet model; S5. Calculate wave impedance loss based on the well logging wave impedance data and the wave impedance prediction results; calculate seismic loss based on the synthetic seismic data output by the forward model and the seismic data; calculate physical constraint loss based on the discrete residual of the wave equation; and calculate wavelet regularization loss based on the parameters of the time-varying wavelet model. S6, the wave impedance loss, the seismic loss, the physical constraint loss and the wavelet regularization loss are weighted and combined into a total loss, and the inversion network parameters, the forward modeling operator parameters and the time-varying wavelet model parameters are jointly updated through backpropagation; S7, under the condition that the total loss converges, output the final wave impedance prediction result of the inversion network and the wavelet estimation result of the time-varying wavelet model.

2. The seismic impedance inversion method based on the joint CNN and Transformer as described in claim 1, characterized in that: The neural network correction component is used to learn the residual correction term of the physical forward modeling component, wherein the numerical solution process of the physical forward modeling component is implemented in a differentiable finite difference scheme, so that the physical constraints of the wave equation are satisfied at the architecture level. The physical forward modeling component maps wave impedance to synthetic seismic data by numerically solving discrete wave equations, including one or more of the following: The acoustic wave equation adopts the first-order velocity-stress form, and the acoustic wave equation is as follows: in, Let the velocity of the particle be... For sound pressure, For density, Bulk modulus For spatial coordinates, For time, the speed of sound wave impedance ; The acoustic wave equations are discretized using an interleaved grid finite difference scheme to obtain the discrete equations: in, For spatial grid indexing, For time step index, For spatial step size, For time step; The output of the physical forward modeling component For the sound pressure recording at the receiving point, the final output of the forward operator is: in, The residual correction term learned by the neural network correction component.

3. The seismic impedance inversion method based on the joint CNN and Transformer as described in claim 2, characterized in that: The physical constraint loss is calculated based on the discrete residuals of the wave equation, including: The final output of the forward operator Substituting into the residual expression of the discrete wave equation: in, Let be the discrete differential operator matrix of the wave equation. The source term matrix is ​​related to wave impedance. Wave impedance; The physical constraint loss is: in, These are the weighting coefficients of the residuals in the wave equation. The L2 regularization coefficient of the neural network correction term; During the training process, Dynamically adjust based on the current training error: when the training error exceeds a preset threshold, reduce... This allows the neural network correction component to fit flexibly; when the training error is less than the preset threshold, the value is increased. This causes the forward operator to converge to the vicinity of the physical forward component.

4. The seismic impedance inversion method based on the joint CNN and Transformer as described in claim 3, characterized in that: The time-varying wavelet model is a time-varying weighted combination of multiple fundamental wavelets with different dominant frequencies, including: in, This is the time offset relative to the center of the wavelet. The depth during the two-way travel of the earthquake path. For the first A fundamental wavelet, For the first The dominant frequency parameters of each fundamental wavelet are learnable parameters. For the first The fundamental wavelet at depth The time-varying hybrid weights at the point are learnable parameters, and K is the number of fundamental wavelets.

5. The seismic impedance inversion method based on the joint CNN and Transformer as described in claim 4, characterized in that: The fundamental wavelet's dominant frequency parameter and time-varying hybrid weight are both learnable parameters. The time-varying hybrid weight is a function of the two-way travel time depth and is used to characterize the wavelet's time-varying features caused by the formation absorption attenuation effect. The fundamental wave Using Ricker wavelet form: The time-varying hybrid weights are parameterized using the softmax function: in, and As learnable parameters, this parameterization method guarantees ; The wavelet regularization loss includes the smoothness constraint and the time-varying weight continuity constraint of the time-varying wavelet model: in, For the total variation of the wavelet, and These are the weighting coefficients. The distance between the two sides in the depth direction is the distance between the two sides.

6. The seismic impedance inversion method based on the joint CNN and Transformer as described in claim 5, characterized in that: The inversion network includes a Transformer encoder and a multi-scale convolutional neural network. The Transformer encoder is used to extract global features of the seismic data, and the multi-scale convolutional neural network is used to extract local features of the seismic data.

7. The seismic impedance inversion method based on the joint CNN and Transformer as described in claim 6, characterized in that: The inversion network includes: The Transformer encoder, comprising a multi-head self-attention layer and a feedforward network layer, is used to extract global dependency features after location encoding of the input seismic data. The multi-head self-attention mechanism is calculated as follows: Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively. The dimension of the key vector; Multi-scale convolutional neural networks, which include multiple branches, each using dilated convolution with different dilation coefficients, are used to extract local features of seismic data at different scales. The residual connection module is used to fuse the output of the Transformer encoder and the output of the multi-scale convolutional neural network by element-wise addition, and output the wave impedance prediction result. Where X represents the input seismic data. For the mapping function of the Transformer encoder, This is the mapping function for a multi-scale convolutional neural network.

8. The seismic impedance inversion method based on the joint CNN and Transformer as described in claim 7, characterized in that: The total loss is: in, The mean square error is used to represent the wave impedance loss. The earthquake loss is calculated using the mean square error between the synthetic seismic data and the original seismic data. This refers to the physical constraint loss; The wavelet regularization loss; , , , These are the weighting coefficients for each loss term.

9. The seismic impedance inversion method based on the joint CNN and Transformer as described in claim 8, characterized in that: When the total loss converges, the final wave impedance prediction result of the inversion network and the wavelet estimation result of the time-varying wavelet model are output, including one or more of the following: The seismic data of the target work area is input into the trained inversion network, and the wave impedance profile of the target work area is output. Output the time-varying hybrid weights at each depth τ in the time-varying wavelet model after training. and the dominant frequency parameters of each fundamental wavelet , as the wavelet estimation result of the target work area; The wavelet estimation results include time-varying characteristics where the dominant frequency of the shallow wavelet is higher than that of the deep wavelet.

10. A seismic impedance inversion system based on a combination of CNN and Transformer, employing the seismic impedance inversion method based on a combination of CNN and Transformer as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to acquire seismic data of the target work area and logging impedance data of at least one well. The inversion network unit is used to extract features and output wave impedance prediction results through a Transformer encoder and a multi-scale convolutional neural network. Forward modeling operator unit, used to map wave impedance to synthetic seismic data through physical forward modeling components and neural network correction components; Time-varying wavelet units are used to construct time-varying wavelet models and output wavelet estimation results; The training unit is used to calculate various losses and jointly update parameters.

Images