Full waveform inversion method for enhancing boundary perception

Abstract

The invention discloses a full-waveform inversion method for enhancing boundary perception. The full-waveform inversion method comprises the following steps: acquiring seismic data; constructing a seismic waveform inversion network model; the seismic waveform inversion network model comprises an encoder, a decoder and a jump connection; carrying out the size adjustment of the seismic data through an encoder, and then extracting the advanced features of the reserved low-frequency information, and obtaining a multi-scale coding feature map; inputting the multi-scale encoding feature maps extracted by the encoder into different levels of the decoder by utilizing jump connection; a decoder is used for obtaining distinguishing information of the spatial relation contained in the multi-scale coding feature map, and a prediction speed model is generated; multi-scale analysis is executed through a wavelet convolution layer in an encoder, and low-frequency features are enhanced; then, distinguishing information is extracted by utilizing the spatial relationship of the features through a spatial attention module in a decoder so as to improve the resolution; finally, a reflection coefficient is introduced into the loss function to emphasize punishment on the low-contrast boundary; therefore, the full-waveform inversion precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of full waveform inversion, and in particular to a full waveform inversion method for enhancing boundary perception. Background Art

[0002] Traditional seismic imaging methods usually require multiple steps and the collaboration of geophysical experts, but the resulting velocity model has a low resolution. In contrast, deep learning full waveform inversion (DL-FWI) can achieve an end-to-end workflow without human intervention, especially when dealing with complex geological conditions. Traditional full waveform inversion (FWI) relies on an initial velocity model and often faces challenges such as pathological conditions, cycle jumps, and high computational costs when applied to field data. At the same time, due to the slight changes in the physical properties of the medium at the geological boundary, the waveform changes generated by the seismic wave during propagation are weak, making it difficult to capture small-scale changes in the underground structure.

[0003] To solve these problems, DL-FWI uses the powerful nonlinear fitting ability of deep neural networks to significantly improve the accuracy of velocity model estimation.

[0004] The network architecture is particularly critical in DL-FWI. Typical network structures such as GeoDNN and enhanced FCN both use fully connected networks (FCN), while FCNVMB directly extracts velocity models from raw seismic data through supervised deep fully convolutional neural networks, automatically extracts multi-layer information features, and does not require initial velocity settings. InversionNet uses an encoding-decoding structure to predict geological structures through position interaction, effectively reducing the calculation time and improving the inversion accuracy. To adapt to different inversion tasks, various loss functions are designed accordingly. For example, AD-FWI uses an adversarial loss function that combines Wasserstein distance and L2 metric, which converges quickly and has the ability to capture details. The reweighted variational method generates high-resolution velocity models by introducing data mismatch terms with adjustable weights. In addition, the loss function with structural constraints can achieve more efficient super-resolution FWI. Incorporating domain knowledge can avoid unreasonable predictions because immutable physical laws play a guiding role in the training process. For example, UPFWI uses the wave equation to constrain the dimension of the solution space and extracts style and content features through VGG16 as the perceptual loss function, thereby improving the inversion quality through triple physical constraints. VelocityGAN learns data distribution representation to narrow the distribution gap between the source domain and the target domain. In addition, the performance of DL-FWI can be further improved by using parameter distribution and low-frequency components of the signal.

[0005] However, existing methods still face challenges with position drift and velocity errors when dealing with complex or low-contrast stratigraphic boundaries. Complex geological structures increase the complexity of the paths of seismic wave propagation in the medium, resulting in complex waveforms that increase the difficulty of accurate inversion. Under low-contrast boundary conditions, waveform changes are subtle and difficult to capture accurately. These challenges require the design of more advanced network architectures and robust constraints so that the algorithm can better adapt to complex geological conditions and achieve enhanced perception of boundaries.

[0006] DL-FWI aims to predict the accurate velocity value for each sampling point. Visually, the predicted velocity model requires accurate color mapping and clear boundaries. This process does not rely on the wave equation, but instead uses the powerful nonlinear fitting capabilities of deep learning networks, making it a very challenging task. Effectively integrating physical knowledge into the deep learning process can help generate more accurate velocity models. DL-FWI can be roughly divided into three categories according to the degree of physical knowledge integration.

[0007] First, various existing neural network structures directly learn the mapping from seismic data to geological models. The first strategy uses an encoder-decoder structure to generate structural predictions. SeisInvNet enhances each seismic waveform with auxiliary knowledge of neighboring traces to improve spatial consistency. InversionNet and InversionNet3D model the interaction between different locations to invert geological structures. InvLINT exploits the property that the integration result of velocity with Gaussian kernel is linearly related to the integration of seismic data with sine kernel. This makes it a lightweight inversion network that does not require encoder fine-tuning. In addition, SimFWI decouples the encoder and decoder and trains them separately through self-supervised learning. It helps understand the relationship between multiple FWI datasets and different subsurface structures by connecting paired seismic data and geological models using linear transformers. The second strategy adopts a U-Net structure with skip connections, which facilitates better information flow during encoding and decoding. FCNVMB first reconstructs subsurface parameters using a multi-level feature fusion U-Net structure. DD-Net further develops a U-Net with velocity and boundary decoders, and the two decoders interact to enhance the encoding of spatial information at the edges. The third strategy involves a generative adversarial network structure with data generation capabilities to alleviate the problem of insufficient training data. VelocityGAN learns regularization from geological models to improve inversion accuracy. FWIGAN combines the wave equation with the discriminant network by incorporating physical constraints into the discriminant network. It reduces the sensitivity to the initial model or data noise and alleviates the local minimum problem. AD-FWI uses an adversarial loss function that combines Wasserstein distance and L2 metric. It converges faster and has stronger detail capture capabilities. In addition, conditional generative adversarial networks can also be used for full waveform inversion. However, its performance is poor when there are significant differences between synthetic data and field data.

[0008] Secondly, appropriate physical constraints can prevent unreasonable outputs and alleviate the problem of data scarcity. SeisInvNet enhances seismic traces by exploiting the global context of neighboring information, observation settings, and corresponding seismic profiles. It learns spatially aligned feature maps and further connects them to reconstruct geological models. This approach alleviates problems such as weak spatial correspondence and time-varying characteristics of seismic data. UPFWI integrates partial differential equations and convolutional networks into a cyclic unsupervised full waveform inversion scheme. It uses partial differential equations for forward modeling and convolutional networks for reverse modeling, achieving comparable accuracy to supervised methods. Fourier-DeepONet uses Fourier neural operators as the decoder of DeepONet, which has good interpretability. In addition, some studies integrate FWI gradient operators into multi-depth networks or apply RTM constraints to iteratively optimize low-wavenumber geological models, improving the details of geological model reconstruction. However, due to the complex geological conditions at the exploration site and the low signal-to-noise ratio of seismic signals, the performance of field data inversion is still poor.

[0009] Third, the interpretable inversion network can clarify the physical meaning of each module. The hidden layers and weights in the network can be replaced by the calculation process and parameters of the physical model. Then, the RNN is constructed by the finite difference time domain method. This is a method to simulate the propagation of wave equations and estimate medium parameters. In addition, the moment method can also be used to construct an interpretable network. It helps to calculate the forward scattered field in the two-dimensional plane and predict the backscattering parameters. However, when the model cannot be decomposed due to its nonlinearity, the interpretable network cannot be used as a substitute. Summary of the invention

[0010] In view of the above-mentioned deficiencies in the prior art, the present invention provides a full waveform inversion method for enhancing boundary perception.

[0011] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0012] A full waveform inversion method for enhancing boundary perception comprises the following steps:

[0013] Acquisition of seismic data;

[0014] Constructing a seismic waveform inversion network model; the seismic waveform inversion network model includes an encoder, a decoder and a skip connection;

[0015] The encoder is used to resize the seismic data and extract high-level features that retain low-frequency information to obtain a multi-scale encoded feature map;

[0016] Use skip connections to input the multi-scale encoded feature maps extracted by the encoder into different levels of the decoder;

[0017] The decoder is used to extract discriminative information about the spatial relationship contained in the multi-scale encoded feature map to generate a prediction speed model.

[0018] Optionally, the encoder comprises:

[0019] A scale adjustment module, a first encoding module, a second encoding module, a third encoding module and a fourth encoding module are arranged in sequence.

[0020] Optionally, the first encoding module, the second encoding module, the third encoding module and the fourth encoding module are all provided with a wavelet convolution layer after the convolution layer.

[0021] Optionally, the wavelet convolution layer first uses a weight tensor to perform a convolution operation on the input feature map to obtain a convolution feature map; then multi-layer wavelet decomposition is used on the input feature map. During the decomposition process, wavelet transform is used on the low-frequency feature map separated in the previous stage to obtain a reduced low-frequency feature map and a high-frequency feature map, and then a convolution operation is performed on these feature maps of different frequencies to obtain a low-frequency convolution feature map and a high-frequency convolution feature map; then an inverse wavelet transform is used on the low-frequency convolution feature map and the high-frequency convolution feature map to obtain a transformed feature map; finally, the transformed feature map and the convolution feature map are cascaded to obtain an output feature map.

[0022] Optionally, the decoder comprises:

[0023] A first decoding module, a second decoding module, a third decoding module and a fourth decoding module are arranged in sequence.

[0024] Optionally, the first decoding module, the second decoding module, the third decoding module and the fourth decoding module all have a spatial attention module after the convolution layer.

[0025] Optionally, the spatial attention module first performs average pooling and maximum pooling operations on the input feature map, then concatenates the pooled feature maps and performs a convolution operation, and then performs weighted fusion of the convolution feature map with the input feature map after passing the activation function to generate a spatial attention feature map.

[0026] Alternatively, pixel-level loss and boundary loss are used as the total loss when training the seismic waveform inversion network model, specifically:

[0027]

[0028] in, is the total loss, is the pixel-level loss, is the boundary loss, α is the weight, v is the true value, The velocity model for prediction.

[0029] Optionally, the pixel-level loss is specifically:

[0030]

[0031] in, is the pixel-level loss, is the L1 norm, is the L2 norm, λ is the pixel loss weight, v is the true value, The velocity model for prediction.

[0032] Optionally, the boundary loss is specifically:

[0033]

[0034] in, is the boundary loss, W(v) ij is the boundary reflectivity adjustment factor, is the sum of the absolute differences of the gradient map, v is the true value, is the predicted velocity model, (i, j) is the coordinate of the pixel in the velocity model, and n, m is the number of pixels on the horizontal and vertical coordinates in the velocity model.

[0035] The present invention has the following beneficial effects:

[0036] The present invention performs multi-scale analysis through a wavelet convolution layer in the encoder to enhance low-frequency features; then extracts discriminative information by utilizing the spatial relationship of features in the decoder to improve the resolution; finally, introduces a reflection coefficient into the loss function to emphasize the penalty for low-contrast boundaries, thereby improving the accuracy of full waveform inversion. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of the process of a full waveform inversion method for enhancing boundary perception;

[0038] Figure 2 This is a schematic diagram of the seismic waveform inversion network model structure;

[0039] Figure 3 Schematic diagram of the spatial attention module structure;

[0040] Figure 4 It is a schematic diagram of the calculation process of the reflection coefficient adjustment factor;

[0041] Figure 5 Schematic diagram of experimental results on the FlatFaultA dataset;

[0042] Figure 6 Schematic diagram of experimental results on the CurveVelA dataset;

[0043] Figure 7Schematic diagram of experimental results on the CurveFaultA dataset;

[0044] Figure 8 It is a schematic diagram of the comparison results between the present invention and the prior art. DETAILED DESCRIPTION

[0045] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0046] Reference Figure 1 , a full waveform inversion method for enhancing boundary perception provided by an embodiment of the present invention includes the following steps S1 to S5:

[0047] S1. Obtain seismic data;

[0048] S2, constructing a seismic waveform inversion network model; the seismic waveform inversion network model includes an encoder, a decoder and a skip connection;

[0049] S3, using an encoder to resize the seismic data and extract high-level features that retain low-frequency information to obtain a multi-scale encoding feature map;

[0050] S4, using skip connections to input the multi-scale encoded feature maps extracted by the encoder into different levels of the decoder;

[0051] S5. Use the decoder to extract discriminative information of the spatial relationship contained in the multi-scale encoded feature map to generate a prediction speed model.

[0052] In an alternative embodiment of the present invention, referring to Figure 2 , the encoder includes:

[0053] A scale adjustment module, a first encoding module, a second encoding module, a third encoding module and a fourth encoding module are arranged in sequence.

[0054] In the encoder, the resizer (grey cube) resizes the input to a roughly square size to facilitate square convolution operations. Max pooling reduces the spatial dimensionality and retains the most salient features.

[0055] The feature extractor consists of an ordered repetition of convolution, normalization, and PReLU. Normalization balances the impact of different feature dimensions on the network and speeds up the convergence process.

[0056] PReLU introduces nonlinearity while allowing negative input activation values, helping the network to better learn complex data patterns and features. Ordered repetitive operations enable the feature extractor to better learn the input data features and enhance the ability to recognize complex patterns.

[0057] In an alternative embodiment of the present invention, referring to Figure 2 , the first encoding module, the second encoding module, the third encoding module and the fourth encoding module all set a wavelet convolution layer after the convolution layer.

[0058] The wavelet convolution layer first uses a weight tensor to perform a convolution operation on the input feature map to obtain a convolution feature map; then it uses multi-layer wavelet decomposition on the input feature map. During the decomposition process, wavelet transform is used on the low-frequency feature map separated in the previous stage to obtain a reduced low-frequency feature map and a high-frequency feature map, and then these feature maps of different frequencies are convolved to obtain a low-frequency convolution feature map and a high-frequency convolution feature map; then an inverse wavelet transform is used on the low-frequency convolution feature map and the high-frequency convolution feature map to obtain a transformed feature map; finally, the transformed feature map and the convolution feature map are cascaded to obtain an output feature map.

[0059] In the encoder, repeated convolution and pooling operations are used to reduce the data dimension and extract important features. In addition, a WTconv layer is introduced before each feature extraction stage. This approach expands the receptive field of convolution, preserves important low-frequency information, and helps to recover deep geological structures.

[0060] The WTconv layer is designed to preserve low-frequency information. First, the input feature map x is assigned to Then, a k×k weight tensor w is used to perform a convolution operation on x. The multi-layer wavelet decomposition is performed, and the number of layers l in the present invention is 1. In the decomposition process, the wavelet transform is applied to the low-frequency features separated in the previous stage to obtain the reduced low-frequency features. and high frequency features Then convolution operations are performed on these different frequency graphs to obtain and The algorithm performs inverse wavelet transform on the high and low frequency components after convolution. Finally, the inverse transform result z (1) And the convolution result Cascade to get the final output z (0) This process emphasizes low-frequency components during the calculation, which is particularly beneficial for accurately inverting deep geological structures.

[0061] This embodiment sets a WTconv layer in the encoder to perform multi-scale analysis of seismic data in the wavelet domain, effectively detecting signal discontinuities and capturing features at different scales. This layer also enhances key low-frequency features, which is helpful for the inversion of deep geological structures.

[0062] In an alternative embodiment of the present invention, referring to Figure 2 , the decoder includes:

[0063] A first decoding module, a second decoding module, a third decoding module and a fourth decoding module are arranged in sequence.

[0064] Among them, the first decoding module, the second decoding module, the third decoding module and the fourth decoding module all set a spatial attention module after the convolution layer.

[0065] The spatial attention module first performs average pooling and maximum pooling operations on the input feature map, then concatenates the pooled feature maps and performs a convolution operation, and then performs a weighted fusion of the convolution feature map with the input feature map after passing the activation function to generate a spatial attention feature map.

[0066] In the decoder, deconvolution is used to recover the detailed features of the velocity map. In addition, a spatial attention module (SAM) is integrated in the decoder after each deconvolution to increase the resolution. This approach recovers the previously neglected spatial information and helps achieve accurate boundary localization and clearer segmentation in deep learning full waveform inversion (DL-FWI).

[0067] Reference Figure 3 ,The spatial attention module (SAM) focuses on identifying which parts are information-rich regions. SAM consists of four components: spatial feature integration through average pooling and maximum pooling, connection feature fusion, sigmoid feature weighting, and spatial attention map generation.

[0068] In full waveform inversion, contextual information is crucial. It helps model spatial context and sharpen boundaries. It also enables the use of neighborhood information to weight important areas, thereby enhancing the focus on geological boundary details.

[0069] In this embodiment, SAM is set in the decoder to assign different weights to different spatial positions in the feature map, so that the decoder can pay more attention to important boundary areas, thereby improving the accuracy and effectiveness of feature selection. In addition, SAM enhances the decoder's understanding of spatial relationships by focusing on features at different positions, thereby better integrating spatial information.

[0070] In an optional embodiment of the present invention, pixel-level loss and boundary loss are used as the total loss when training the seismic waveform inversion network model, specifically:

[0071]

[0072] in, is the total loss, is the pixel-level loss, is the boundary loss, α is the weight, v is the true value, The velocity model for prediction.

[0073] Among them, the pixel level loss is specifically:

[0074]

[0075] in, is the pixel-level loss, is the L1 norm, is the L2 norm, λ is the pixel loss weight, v is the true value, The velocity model for prediction.

[0076] Among them, the boundary loss is specifically:

[0077]

[0078] in, is the boundary loss, W(v) ij is the boundary reflectivity adjustment factor, is the sum of the absolute differences of the gradient map, v is the true value, is the predicted velocity model, (i, j) is the coordinate of the pixel in the velocity model, and n, m is the number of pixels on the horizontal and vertical coordinates in the velocity model.

[0079] Most DL-FWI methods adopt a loss function, or Minimize the difference between the predicted value and the true value. However, they usually lead to blurred boundaries. Here, we propose an RCTB loss function that includes pixel-level loss and boundary loss:

[0080]

[0081] The pixel-level loss considers the difference between the predicted velocity model and the true value, i.e.

[0082]

[0083] The boundary loss aims to optimize the representation to improve the accuracy of boundary inversion, i.e.

[0084]

[0085] in

[0086]

[0087] It is a function of the sum of the absolute differences of the two directional gradient maps, and W is a function of the boundary reflectivity adjustment factor.

[0088] In the loss function design of this embodiment, a reflection coefficient is introduced into the reflection coefficient adjusted boundary (RCTB) loss to adjust the gradient map weight of the low contrast area, so as to focus on the boundary area of ​​the speed transition and reduce the error.

[0089] This embodiment introduces a penalty on the boundary in the RCTB loss function, and takes pixel-level fitting into consideration, especially assigning a higher weight to low-contrast boundaries, thereby optimizing the inversion result.

[0090] Reference Figure 4 , there are three steps to calculate W. It starts with the true value v, such as Figure 4 As shown in (a).

[0091] In a first step, the present embodiment locates the boundary and its transition area. Figure 4 (b) The true value boundary obtained by the Canny algorithm, that is,

[0092] e=E(v).

[0093] The existing DL-FWI method does not achieve ideal performance on the boundary. The real geological boundary shows transition velocity changes due to diffusion motion. Therefore, this embodiment uses the dilation operation To locate the transition area:

[0094]

[0095] Where (i, j) represents the coordinates of the pixel in the velocity model, D(k) represents the scope of the structural element k, and when k is a 3×3 matrix, D(k) = {-1, 0, 1} 2 Therefore, the dilation operation takes the maximum value in the 3×3 neighborhood of the target pixel as the new value. Figure 4 Middle (c), also called the transition region.

[0096] In the second step, the embodiment calculates the reflectivity coefficient of the velocity map. The contrast of the boundary is positively correlated with the amplitude of the reflectivity coefficient. The embodiment calculates the reflectivity coefficient and uses a dilation operation to connect similar regions. Figure 4 (d) is the reflection coefficient image. It is calculated by the following formula:

[0097]

[0098] Where ρ represents the density map, v represents the velocity map, ρ i,j represents the density value at coordinate (i,j) and v i,jr represents the velocity value of the coordinate (i, j). It is usually assumed that ρ is a constant matrix. Therefore, it is equivalent to

[0099]

[0100] Therefore, the smaller the velocity difference, the smaller the value of the absolute reflectance. Boundaries with small reflectance are also the most challenging weak-contrast regions in DL-FWI. Figure 4 (e) is the dilation result of the reflectance image. Large values ​​will spread to the neighborhood, making them more prominent at the boundary. In order to penalize only the boundary, the reflectance near the boundary is effective, such as Figure 4 As shown in (f).

[0101] In the third step, the present embodiment generates a boundary reflectivity adjustment factor. Weak contrast boundaries are more likely to be underestimated. The present embodiment uses a step function to assign greater weights to weak contrast areas, i.e., areas with smaller reflectivity coefficients. The step function is

[0102]

[0103] Among them, r is the processed reflectivity map, r ij is the value at coordinate (i, j), and T is the threshold.

[0104] Boundary reflectivity adjustment factor

[0105] The penalty range of the boundary loss is narrowed to the vicinity of the boundary and the penalty weight of the low-contrast boundary is increased. This embodiment applies a reflectivity adjustment factor to the gradient difference between the predicted and true velocity models to reduce the inversion error at the boundary. In summary, the RCTB loss introduces a penalty for the boundary while considering pixel-level fitting. In particular, it emphasizes a higher weight on the low-contrast boundary, thereby improving the inversion results.

[0106] The method proposed by the present invention is experimentally verified below.

[0107] All experiments were conducted on a system equipped with an NVIDIA RTX 3060 GPU and a 13th-generation Intel Core TM The training was conducted on an i9-13900K workstation using the PyTorch GPU framework. The RCTB loss function was used during training with parameters set to α=0.5, λ=0.5, and T=0.05. The structural element k in the dilation operation is a 3×3 matrix. All methods use the Adam optimizer with the following parameters: β1=0.9, β2=0.999, ∈=10 -8 , and a momentum term m = 0.9. All modules are initialized using the default method of PyTorch.

[0108] In the experiments conducted on the OpenFWI dataset, the velocity model is normalized to the range [0,1][0,1] and the learning rate is set to 10 -4 The parameter configuration of DD-Net70 follows the parameters specified in the original publication. The batch size of InversionNet and VelocityGAN is adjusted to 20. This adjustment improves the performance, as shown in the corresponding tables and figures. The batch size of ABA-FWI is also set to 20. ABA-FWI converges in about 120 epochs.

[0109] The present invention uses eight indicators to evaluate the performance of the algorithm: peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), mean square error (MSE), mean absolute error (MAE), universal image quality (UIQ), learning perceptual image block similarity (LPIPS), boundary mean square error (BMSE) and boundary mean absolute error (BMAE). PSNR is used to measure the quality of reconstruction. SSIM, UIQ and LPIPS are used to evaluate the similarity between two images. MSE and MAE are used to evaluate pixel differences. BMSE and BMAE are intended to evaluate the inversion accuracy near the boundary. Since the first four indicators are usually used, the present invention only presents the last four.

[0110] UIQ is a full-reference image quality assessment metric that calculates the quality score by comparing the feature differences between the predicted image and the reference image. Unlike structural similarity (SSIM), UIQ requires the original image as a reference. It describes the structural information of the image in a way that is more consistent with human perception. The calculation formula of UIQ is:

[0111]

[0112] in, and They measure the degree of linear correlation, the distance of average pixel brightness, and the similarity of contrast respectively. The value of the first component is in the range of [-1,1][-1,1], and the value range of the last two is [0,1][0,1]. The higher the product of these three values, the greater the similarity.

[0113] LPIPS is a learning-based perceptual image patch similarity metric. Unlike traditional image quality assessment metrics such as SSIM, it uses a pre-trained convolutional neural network (usually a VGG network) to extract features and calculate differences. By comparing the similarity of high-level feature representations in images, LPIPS enhances human perception of differences. The calculation formula for LPIPS is:

[0114]

[0115] where v1 and v2 represent the actual and predicted velocity maps, respectively, and φ i represents the i-th layer of the convolutional neural network, N is the total number of layers, Measures the difference between two feature representations. The final score of LPIPS is the sum of the differences across all layers. The smaller the score, the more similar the two images are.

[0116] BMAE and BMSE reduce the error calculation range from global to near edge. BMAE takes into account the predicted velocity model The difference between the true value b near the edge is:

[0117]

[0118] The calculation method of b is

[0119]

[0120] The calculation method is

[0121]

[0122] BMSE is

[0123]

[0124] The smaller the value, the smaller the error near the boundary.

[0125] The present invention uses the OpenFWI dataset for testing.

[0126] Figure 5 Experimental results on the FlatFaultA dataset. Inversion results of four methods on two FlatFaultA examples. The first row of each example represents the inversion result. The second row shows the residual between the predicted value and the true value. Columns 2 to 5 correspond to the four methods, and the last column corresponds to the noisy data. For visual clarity, we limit the color scale of the residual to the range of [-100,100][-100,100]. It is challenging to identify fault locations and trends on FlatFaultA, especially in deep geological layers. ABA-FWI is able to obtain clear boundaries in low-contrast areas (indicated by red arrows). Incorrect boundary positioning, such as vertical or horizontal offset, will result in rough boundaries in the residual image. The larger the error, the rougher the boundary. The red dashed line indicates that our method is able to accurately invert the fault trend. The white dashed line shows that ABA-FWI is still able to clearly depict the precise location of the fault in the deep subsurface layer.

[0127] Figure 6Experimental results on the CurveVelA dataset. Inversion results of four methods on two CurveVelA examples. The first row of each example represents the inversion result. The second row shows the residual between the predicted value and the true value. The 2nd to 5th columns correspond to the four methods, and the last column corresponds to the noisy data. For visual clarity, we limit the color scale of the residual to the range of [-100,100][-100,100]. Performance on CurveVelA The key to inverting CurveVelA is to depict a clear and accurate interface. Figure 6 This shows that ABA-FWI produces accurate inversion results. Figure 6 As shown in (a), InversionNet and DD-Net70 invert blurred edges (marked by red arrows) in the depth range of 0.3 to 0.6 km.

[0128] VelocityGAN produces incorrect results (marked by red rectangles) with more layers than the real situation. Figure 6 As shown in (b), the red arrows indicate that ABA-FWI can obtain clearer and more accurate interfaces in deep subsurface layers. DD-Net70 also obtains clear boundaries, but there are obvious artifacts inside the formation, resulting in significant errors. The yellow arrows indicate that ABA-FWI inverts clearer interfaces in low-contrast areas.

[0129] Figure 7 Experimental results on the CurveFaultA dataset. Inversion results of four methods on two CurveFaultA examples. The first row of each example represents the inversion result. The second row shows the residual between the predicted value and the true value. The 2nd to 5th columns correspond to the four methods, and the last column corresponds to the noisy data. For visual clarity, we limit the color scale of the residual to the range of [-100,100][-100,100]. Performance on CurveFaultA CurveFaultA is the most challenging task, which requires clear inversion of the curve interface and accurate positioning of the fault. Figure 7 The inversion results of the interior and boundary of geological layers are shown. The red arrows indicate that ABA-FWI is able to depict clearer interface shapes at fault and curve locations. The white arrows highlight that DD-Net70 produces artifacts inside geological layers, although the boundaries are relatively accurate. The improved accuracy of ABA-FWI in boundary delineation is demonstrated by the finer boundaries in the residual image. Our method can maintain smaller horizontal and vertical offsets even in deeper layers.

[0130] Figure 8 Whether to add feature maps of valid modules (WTconv layers and SAM) to the network. Figure 8(a) and (b) correspond to the feature maps of a typical example in FlatFaultA with and without the effective module, respectively, while (c) and (d) correspond to the results of CurveFaultA. For each example, the left and right columns represent the feature map outputs from the penultimate and final decoding stages, respectively (only the first 12 channels are shown). The results show that the effective module is retained to present accurate edges, while the effective module is removed to show rougher edges and artifacts at the same feature extraction stage. Therefore, the WTconv layer and SAM integrate key information to generate edges earlier and more effectively.

[0131] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0134] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0135] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A full waveform inversion method for enhancing boundary perception, characterized in that: The following steps are involved: Acquisition of seismic data; Constructing a seismic waveform inversion network model; the seismic waveform inversion network model includes an encoder, a decoder and a skip connection; The encoder is used to resize the seismic data and extract high-level features that retain low-frequency information to obtain a multi-scale encoded feature map; Use skip connections to input the multi-scale encoded feature maps extracted by the encoder into different levels of the decoder; The decoder is used to extract discriminative information about the spatial relationship contained in the multi-scale encoded feature map to generate a prediction speed model.

2. A full waveform inversion method for enhancing boundary perception according to claim 1, characterized in that the encoder include: A scale adjustment module, a first encoding module, a second encoding module, a third encoding module and a fourth encoding module are arranged in sequence.

3. A full waveform inversion method for enhancing boundary perception according to claim 2, characterized in that: The first encoding module, the second encoding module, the third encoding module and the fourth encoding module are all provided with a wavelet convolution layer after the convolution layer.

4. A full waveform inversion method for enhancing boundary perception according to claim 3, characterized in that: The wavelet convolution layer first uses a weight tensor to perform a convolution operation on the input feature map to obtain a convolution feature map; then it uses multi-layer wavelet decomposition on the input feature map. During the decomposition process, wavelet transform is used on the low-frequency feature map separated in the previous stage to obtain a reduced low-frequency feature map and a high-frequency feature map, and then these feature maps of different frequencies are convolved to obtain a low-frequency convolution feature map and a high-frequency convolution feature map; then an inverse wavelet transform is used on the low-frequency convolution feature map and the high-frequency convolution feature map to obtain a transformed feature map; finally, the transformed feature map and the convolution feature map are cascaded to obtain the output feature map.

5. The full waveform inversion method for enhancing boundary perception according to claim 1, characterized in that: The decoder includes: A first decoding module, a second decoding module, a third decoding module and a fourth decoding module are arranged in sequence.

6. The full waveform inversion method for enhancing boundary perception according to claim 1, characterized in that: The first decoding module, the second decoding module, the third decoding module and the fourth decoding module all set a spatial attention module after the convolution layer.

7. A full waveform inversion method for enhancing boundary perception according to claim 6, characterized in that: The spatial attention module first performs average pooling and maximum pooling operations on the input feature map, then concatenates the pooled feature maps and performs a convolution operation, and then performs a weighted fusion of the convolution feature map with the input feature map after passing the activation function to generate a spatial attention feature map.

8. The full waveform inversion method for enhancing boundary perception according to claim 1, characterized in that: When training the seismic waveform inversion network model, pixel-level loss and boundary loss are used as the total loss, specifically: in, is the total loss, is the pixel-level loss, is the boundary loss, α is the weight, v is the true value, The velocity model for prediction.

9. A full waveform inversion method for enhancing boundary perception according to claim 8, characterized in that: The pixel-level loss is: in, is the pixel-level loss, is the L1 norm, is the L2 norm, λ is the pixel loss weight, v is the true value, The velocity model for prediction.

10. The full waveform inversion method for enhancing boundary perception according to claim 8, characterized in that: The boundary loss is specifically: in, is the boundary loss, W(v) ij is the boundary reflectivity adjustment factor, is the sum of the absolute differences of the gradient map, v is the true value, is the predicted velocity model, (i, j) is the coordinate of the pixel in the velocity model, and n, m is the number of pixels on the horizontal and vertical coordinates in the velocity model.

Images