Three-dimensional seismic fault identification and visualization method based on improved U-Net

Abstract

The invention provides a three-dimensional seismic fault identification and visualization method based on improved U-Net, and belongs to the technical field of seismic data interpretation. The method comprises the steps of performing standardization processing and data enhancement operation on input three-dimensional seismic volume data, and dividing the input three-dimensional seismic volume data into a training set and a verification set; training set data are input into an improved three-dimensional U-Net model, cavity space pyramid pooling (ASPP) is introduced into a bottleneck layer to extract multi-scale features, and Guide Decoder branches are arranged on a second layer encoder and a third layer encoder to enhance middle layer supervision and improve boundary definition and continuity; the model outputs a probability prediction result obtained through Sigmoid function activation, and a fault mask is generated through thresholding; and finally, selecting a three-way angle section to realize three-dimensional fusion visualization with an original seismic body. Compared with the prior art, the method gives consideration to the recognition of long faults and small cracks, and remarkably improves the precision and robustness of fault prediction.

Description

Technical Field

[0001] This invention relates to the field of seismic data interpretation technology, and in particular to a three-dimensional seismic fault identification and visualization method based on an improved U-Net. Background Technology

[0002] Seismic exploration is an important tool for oil and gas exploration and subsurface structural interpretation. Fault identification is crucial for delineating oil and gas-bearing structures, analyzing structural evolution, and determining well locations. Current fault identification methods are mainly divided into two categories: traditional seismic attribute analysis methods and automatic identification methods based on deep learning.

[0003] Traditional seismic attribute analysis methods primarily identify fault strikes and distributions by calculating the discontinuities of seismic phase axes in the coherence volume. Simultaneously, curvature attributes are used to analyze the degree of stratigraphic bending to infer fault development zones. In practical applications, these methods combine manual interpretation or semi-automatic calculation to perform thresholding, attribute overlay, and visualization of the attribute volume, thereby marking the location and morphology of faults. However, these methods rely heavily on human interpretation, resulting in low processing efficiency. Furthermore, when seismic data is noisy, stratigraphic structures are complex, or fault characteristics are not prominent, incomplete fault boundary identification can easily occur, leading to misjudgments and missed faults.

[0004] The development of automatic fault identification methods based on deep learning has been rapid in recent years, especially those based on 3D U-Net. These methods leverage the feature extraction and segmentation capabilities of convolutional neural networks to achieve voxel-level prediction of 3D seismic bodies through end-to-end feature learning, automatically generating fault prediction results. However, existing methods still have shortcomings: on the one hand, the bottleneck layer convolutional receptive field of standard U-Net is limited, lacking the ability to simultaneously model long-range extended faults and localized fine cracks; on the other hand, although skip connections can fuse shallow and deep features during the decoding stage, the mid-layer features themselves are not directly supervised. During training, the model tends to over-rely on skip connections while neglecting mid-layer edge and contour information, resulting in blurred and discontinuous predicted fault boundaries.

[0005] To address the aforementioned shortcomings, existing research has introduced dilated convolutions in 3D seismic fault identification to expand the receptive field and enhance the ability to capture long-range faults. Attention mechanisms have also been employed to suppress irrelevant noise and highlight fault-related regions during the feature fusion stage, thus improving prediction accuracy and robustness to some extent. However, these methods generally suffer from complex network structures, high computational costs, and high training overhead. Furthermore, they remain insufficient in supervising mid-level features, resulting in unclear depictions of fault boundaries and difficulty in simultaneously achieving comprehensive identification of multi-scale structures and maintaining boundary continuity. Summary of the Invention

[0006] This invention provides a three-dimensional seismic fault identification and visualization method based on an improved U-Net. This method introduces a hollow spatial pyramid pooling (ASPP) module into the traditional three-dimensional U-Net network structure to enhance multi-scale feature extraction capabilities. Simultaneously, a guided decoder branch is set in the second and third layer encoders to supervise the mid-layer features during the training phase, thereby improving the problems of boundary ambiguity and fault discontinuity. This invention aims to solve the problems of ambiguous fault boundary prediction, insufficient continuity, and inadequate multi-scale structure identification capabilities in existing methods.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for three-dimensional seismic fault identification and visualization based on an improved U-Net includes the following steps:

[0009] Step 1: Preprocess the input 3D seismic volume data. The training and validation data used are from the FaultSeg3D dataset, which was built and released open-source on GitHub by researchers including XinmingWu from the University of Texas at Austin. This dataset contains large-scale synthetic seismic volumes and their corresponding fault annotations. First, the mean and standard deviation of the original data are calculated, and the data is standardized using the zero-mean unit variance normalization method. Then, the normalized data is sliced ​​and patched to obtain sub-voxels suitable for 3D convolutional networks. During the training phase, the sub-voxels are further augmented, including random flipping along the longitudinal, lateral, and depth directions, translation of 5 voxels each along the X, Y, and Z axes, and 90° rotation around the X, Y, and Z axes, thereby improving the model's robustness and generalization ability. Finally, the processed data is divided into training and validation sets in a 6:4 ratio. During the inference and visualization phase, real Netherlands F3 earthquake data is introduced. This dataset, collected by the Netherlands Geological Survey in the North Sea, is widely used as a generalization evaluation test set for fault identification models.

[0010] Step 2 involves inputting the preprocessed 3D seismic data into an improved U-Net network model. This network consists of a four-layer encoder, a decoder, and a bottleneck layer. In the encoder stage, input features are progressively extracted through multi-layer 3D convolution and downsampling. The first two encoder layers focus on capturing edge and contour information, while subsequent layers extract deeper semantic features. An ASPP module is set in the bottleneck layer. This module includes a one-dimensional convolutional branch, three-dimensional convolutional branches with dilation rates of 1, 2, and 3, and a global context branch. The outputs of these branches are concatenated along the channel dimension and then fused using one-dimensional convolution to obtain multi-scale features that combine local details and global context. This step provides multi-scale semantic information for subsequent decoding and fine-grained fault boundary reconstruction.

[0011] Step 3: The decoder upsamples the features output by the ASPP module layer by layer and concatenates them with the corresponding encoder features along the channel dimension. Spatial resolution is restored through convolutional fusion, thus obtaining the main output probability map of the fault. Simultaneously, Guide Decoder branches are derived from the second and third encoder layers. Each feature path is processed by 3D and 1D convolutions to form a low-resolution prediction map, which is then upsampled to restore the same spatial size as the main output. During the training phase, the two guide prediction maps and the main output prediction map participate in the weighted joint loss calculation, thereby providing multi-scale supervision in the middle layers, avoiding excessive reliance on skip connections, and improving the clarity and continuity of fault boundaries.

[0012] Step 4: The main output probability map is activated by the Sigmoid function to generate fault probability results, and the fault mask is obtained by binarization with a fixed threshold of 0.5. Based on the triaxial tangent index, the fault mask is fused with the original seismic body and displayed. The fault is then visualized in three dimensions to present its spatial location, strike and extension characteristics.

[0013] Compared with existing technologies, this invention effectively solves the problem of insufficient multi-scale modeling capability of traditional 3D U-Net by introducing the ASPP module in the bottleneck layer, enabling the model to simultaneously identify long-range faults and small cracks. By simultaneously introducing guided decoders from the second and third layers and performing joint optimization during the training phase, the lack of mid-layer feature supervision is effectively compensated for, improving the problems of fuzzy fault boundary prediction and poor continuity. Combined with data standardization enhancement and 3D angular section visualization, not only is the accuracy and robustness of fault identification improved, but the intuitiveness and interpretability of the prediction results are also enhanced. Attached Figure Description

[0014] Figure 1 This is an overall flowchart of a three-dimensional seismic fault identification and visualization method based on an improved U-Net according to an embodiment of the present invention;

[0015] Figure 2 This is a network structure diagram of the improved U-Net in an embodiment of the present invention;

[0016] Figure 3 This is a network structure diagram of the ASPP module in an embodiment of the present invention;

[0017] Figure 4 This is a network structure diagram of the Guide Decoder module in an embodiment of the present invention;

[0018] Figure 5 This is a diagram illustrating the fault prediction effect in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0020] This embodiment proposes a 3D seismic fault identification and visualization method based on an improved U-Net to address the shortcomings of existing fault identification methods, such as insufficient multi-scale feature modeling, blurred fault boundaries, and poor prediction continuity. The method improves upon the traditional 3D U-Net framework by introducing an ASPP module at the bottleneck layer. It employs 1×1×1 convolutions, parallel 3D convolutional branches with different dilation rates, and a global context branch to achieve multi-scale feature extraction and fusion, thus simultaneously recognizing both long-range faults and small fractures. Furthermore, a GuideDecoder branch is derived from the second and third layer encoders. During the training phase, it generates a guided prediction map through convolution and upsampling, which is then jointly used with the main output prediction map in loss optimization. This avoids excessive reliance on skip connections, effectively supervising mid-layer edges and shape structures, and significantly improving the problems of blurred and discontinuous fault boundaries. Through these innovative designs, this embodiment can improve the accuracy and robustness of fault prediction while maintaining model efficiency, and enhance the intuitiveness and interpretability of the results.

[0021] This embodiment presents a three-dimensional seismic fault identification and visualization method based on an improved U-Net, comprising the following steps:

[0022] Step 1. Obtain and preprocess the publicly available dataset to construct a dataset for training the improved U-Net-based 3D seismic fault identification in Step 2 below.

[0023] This embodiment selects a publicly released synthetic earthquake fault dataset as the training and validation data source. This dataset was generated by researchers based on seismic wave propagation theory and tectonic geological modeling methods. By introducing fault structures of different scales, orientations, and extension characteristics into a three-dimensional seismic body, it simulates the distribution of faults in a real underground geological environment. Compared to measured seismic data, this dataset has advantages such as clear structural features, well-defined fault boundaries, lower noise interference, and stronger controllability. Therefore, it is widely used in research on automatic earthquake fault identification and training of deep learning models. This dataset can effectively support the stable convergence of the model in the early training stage and provide a reliable guarantee for the subsequent evaluation of fault identification performance.

[0024] In the data preprocessing stage, the original 3D seismic volume data is first standardized. Specifically, the mean and standard deviation of each voxel value of the seismic volume are calculated, and amplitude correction is performed using a zero-mean unit variance normalization method to eliminate amplitude differences between different data blocks and ensure that the input data are on the same numerical scale. Subsequently, the normalized 3D seismic volume is divided into multiple sub-blocks according to a fixed volume size using a 128×128×128 sliding window slicing method, with each sub-block serving as an independent sample. A 25% window overlap rate is set during the segmentation process to increase the number of samples and avoid the loss of fault information due to boundary truncation. For fault labeling data, a strict correspondence is maintained with the seismic volume segmentation to ensure a one-to-one match between input and label.

[0025] After the sample split, the dataset is further grouped according to the ratio of training set to validation set, using a 6:4 ratio. This ensures that the training set is used for model parameter learning, the validation set is used for hyperparameter tuning and overfitting monitoring, and the Netherlands F3 dataset is used for independent performance evaluation as the test set. The final preprocessed dataset can be represented as X∈R N ×D×H×W Where N is the number of samples, and D, H, and W represent the depth, height, and width of the samples, respectively; the corresponding fault label is denoted as Y∈{0,1}. N×D×H×W , where 1 represents the fault location and 0 represents the non-fault location.

[0026] Step 2. Build a 3D seismic fault identification model based on the improved U-Net.

[0027] like Figure 2 As shown, the network model constructed in this embodiment is based on a three-dimensional U-Net structure, including an encoder, an ASPP module (bottom layer), a decoder, and a GuideDecoder branch. The input three-dimensional seismic volume data is denoted as X. patch ∈R 1 ×D′×H′×W′ First, the feature is mapped to an 8-channel feature C1 through the first convolutional block, and then P1 is obtained through max pooling. Then, it enters the second convolutional block, which outputs a 16-channel feature C2, which is pooled to obtain P2. Next, it enters the third convolutional block, which outputs a 32-channel feature C3, which is downsampled to obtain P3. This feature C3 also serves as the input to guide the decoder branch. Finally, it enters the fourth convolutional block to obtain a 64-channel feature C4, which is pooled to obtain P4 and then fed into the bottleneck layer.

[0028] At the bottleneck layer, the input feature P4 is transformed into a 128-channel feature C5 by the fifth convolutional block and then fed into the ASPP module. ASPP consists of a one-dimensional convolutional branch, multiple three-dimensional convolutional branches with different dilation rates d∈{1,2,3}, and a global pooling branch. After concatenation, the branches are fused by a 1×1×1 convolution to obtain the multi-scale feature representation F. ASPP This feature combines local details with global structural information, providing support for subsequent decoding stages.

[0029] During the decoding phase, F ASPP First, after deconvolution and upsampling, the feature is concatenated with C4 to obtain feature C6; then, after convolution and further upsampling, it is concatenated with C3 to obtain feature C7; then, it is concatenated with C2 to obtain C8, and finally concatenated with C1 to obtain C9. At the highest resolution, C9 undergoes a 1×1×1 convolution and a sigmoid activation function to output the main prediction result Y. main =σ(f 1×1 (C9))∈[0,1 1×D′×H′×W′ The value represents the probability that each voxel belongs to a fault.

[0030] Meanwhile, guide decoder branches are set on the second and third layer features C2 and C3, respectively. Each feature path is first processed by 3×3×3 three-dimensional convolution and ReLU activation to obtain intermediate features G1 and G2, and then by 1×1×1 convolution and Sigmoid activation to generate a low-resolution prediction result Y. guide1 Y guide2 Then, after trilinear interpolation and upsampling, the result is compared with the master prediction result Y. main Same spatial dimensions. During the training phase, Y... guide With Y main They jointly participate in loss optimization, thereby introducing supervision in the middle stage to help the model better learn fault boundary and contour information.

[0031] like Figure 3 As shown, this embodiment introduces the ASPP module into the network bottleneck layer to enhance the multi-scale modeling capability of features. Let the input feature map of the bottleneck layer be C5∈R. N×C×D×H×W Where N is the batch size and C = 128 is the number of channels. The ASPP module consists of four types of branches: the first is a 1×1×1 convolutional branch, whose output feature is denoted as F. 1×1 Secondly, there are multiple three-dimensional convolutional branches with different dilation rates d∈{1,2,3}, whose outputs are denoted as F. d＝1 F d＝2 F d＝3 The first is used to extract local and mesoscale features under different receptive fields; the second is the global context branch, which is to obtain the global feature F by performing three-dimensional adaptive average pooling on the input feature C5 and then performing a 1×1×1 convolution transformation. gpThen, the outputs are upsampled to the same spatial size as the other branches using trilinear interpolation.

[0032] F concat =Concat(F 1×1 ,F d＝1 ,F d＝2 ,F d＝3 ,F gp )

[0033] The dimension is R. N×C′×D×H×W The fusion features are obtained, where C′ is the number of channels after concatenation. Then, a 1×1×1 convolution is used for channel compression and mapping to obtain the final multi-scale enhanced feature F. ASPP .

[0034] like Figure 4 As shown, this embodiment introduces GuideDecoder branches on the second and third layer encoder features C2 and C3 to achieve multi-scale supervision of the intermediate layer features. For the second layer feature C2, it is first transformed into an intermediate feature G1 by a 3×3×3 three-dimensional convolution and ReLU activation, and then the predicted result Y is obtained by a 1×1×1 convolution and Sigmoid activation. guide1 For the third layer feature C3, it is also processed by 3×3×3 convolution and ReLU activation to obtain G2, and then by 1×1×1 convolution and Sigmoid activation to obtain the predicted result Y. guide2 The calculation formula is as follows:

[0035] Y guidei =σ(f 1×1×1 (δ(f 3×3×3 (C i+1 )))),i=1,2.

[0036] Where f 3×3×3 (·) represents a 3D convolution operation with a kernel size of 3×3×3, δ(·) represents the ReLU activation function, and σ(·) is the Sigmoid activation function. Subsequently, Y guide1 Y guide2 Upsampled to the master prediction result Y using trilinear interpolation respectively main Same space dimensions:

[0037]

[0038] During the model training phase, the main prediction result and the two guided prediction results jointly participate in the optimization of the loss function, thereby achieving multi-scale joint supervision of mid-level features. To this end, this embodiment designs a joint loss function based on adaptive weights. Let the true tomographic label be Y. true The main output prediction is Y. mainThe two guidance predictions are respectively The corresponding branch loss is defined using the binary cross-entropy loss:

[0039]

[0040] Where Loss(·) represents the voxel-level segmentation loss function. The final total loss is defined using an adaptive weighted form as follows:

[0041] L=αL main +β1L guide1 +β2L guide2

[0042] Here, α, β1, and β2 are dynamic adaptive weight coefficients that can be automatically adjusted according to the convergence of each branch during training to balance the contributions of the main output and the intermediate guiding branches. This approach avoids the limitations of fixed weights, allowing the main branch prediction and guiding supervision to achieve reasonable optimization at different stages, thereby improving the clarity and continuity of fault boundaries. This design can both constrain the learning of intermediate features and improve the clarity and continuity of fault boundaries.

[0043] like Figure 5 As shown, this is a demonstration of the fault identification results of the method in this embodiment on the test dataset. (1) is a triaxial sectional view of the original seismic body, which can be used to show the distribution characteristics of seismic energy on the structural plane; (2) is a fault prediction result diagram output by the model of this invention. It can be clearly seen that the model has a high response intensity at the fault boundary and can maintain good consistency with the fault strike and distribution characteristics. Compared with the original seismic data, it can be seen that this method can accurately extract the main fault structure and maintain good continuity and clarity in the details, thereby verifying the role of the network structure design and the mid-level supervision mechanism in improving the effectiveness of fault identification.

[0044] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the embodiments listed above. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A three-dimensional seismic fault identification and visualization method based on improved U-Net, characterized in that, The method comprises: Step one, standardizing preprocessing and data enhancement on the input three-dimensional synthetic seismic volume data, generating normalized voxel data, and dividing it into a training set and a validation set; Step two, inputting the normalized voxel data into a three-dimensional U-Net neural network model with a guided decoder and an empty space pyramid pooling (ASPP) module for processing to obtain a prediction result tensor, and generating a fault probability prediction result through a Sigmoid activation function; Step three, threshold binarization processing of the prediction result to extract potential fault areas; Step four, based on the data set established in step one, training the seismic fault prediction model built in step two, binarizing the probability map through a fixed threshold of 0.5 to form a fault mask, obtaining the trained seismic fault prediction model, and using the trained seismic fault prediction model to predict faults in the original seismic volume data; Step five, based on three-way section index, performing three-dimensional angular section visualization on the original seismic volume data and the fault prediction result.

2. The method of claim 1, wherein, The step one comprises: The training and validation data used are derived from the FaultSeg3D data set constructed by Xinming Wu et al. of the University of Texas at Austin and open source published on the GitHub platform, which contains large-scale synthetic seismic volume and its corresponding fault annotation; the original data is first calculated for mean and standard deviation, and normalized by zero mean unit variance normalization method; then data enhancement is performed, specifically including random flipping along the longitudinal, transverse and depth directions, translation of 5 voxels along X, Y and Z directions, and 90° rotation around X, Y and Z coordinate axes, to improve the robustness and generalization ability of the model to various fault shapes; then the processed seismic volume data is sliced to generate sub-voxel blocks with a size of 64x64x64 suitable for input of three-dimensional convolution network; to improve the robustness and generalization ability of the model to various fault shapes; finally, the processed data is divided into a training set and a validation set in a ratio of 6:4, the training set is enabled with enhancement strategy, the validation set maintains the original form for generalization performance evaluation, and NetherlandsF3 real seismic data is introduced in the inference and visualization stage, which is widely used as a generalization evaluation test set for fault identification model, collected by the Netherlands Geological Survey in the North Sea, to verify the effectiveness of the prediction model in real seismic scenarios.

3. The method of claim 1, wherein, The step two comprises that the three-dimensional U-Net neural network model is composed of four layers of encoder modules and four layers of decoder modules, the encoder part sequentially performs convolution and pooling operations on the input seismic body data to extract features layer by layer, wherein the first to fourth encoding layers are respectively composed of two three-dimensional convolution layers and ReLU activation functions, and down-sampling is realized through three-dimensional maximum pooling; after feature extraction in the bottleneck layer, the ASPP structure is introduced, the multi-scale hollow convolution branch and the global pooling branch are fused in parallel to realize the extraction and enhancement of multi-scale context features; the decoder part adopts layer-by-layer deconvolution up-sampling and jump connection with the corresponding encoding layer features, restores the spatial resolution through convolution layer fusion, and gradually generates the fault prediction result; in addition, an auxiliary guided decoding branch is introduced at the second layer and the third layer encoding features respectively, which is up-sampled to the same size as the main output after convolution and Sigmoid activation, used for auxiliary training and middle layer supervision, so as to improve the convergence efficiency and fault boundary recognition accuracy of the model.

4. The method of claim 3, wherein, The three-dimensional U-Net neural network model is provided with a guided decoder branch outside the main decoding path, the guided decoder is introduced from the feature maps of the second layer encoder and the third layer encoder respectively, restored to the original size by convolution and up-sampling, and compared with the real fault mask pixel by pixel, the binary cross entropy loss is calculated to introduce supervision constraint in the middle layer stage, and the continuity of the fault boundary and the clarity of the prediction are improved.

5. The method of claim 3, wherein, The three-dimensional U-Net neural network model is provided with an ASPP module at the bottleneck layer, the module includes a 1x1x1 convolution branch, three-dimensional convolution branches with expansion rates of 1, 2 and 3, and a global context branch, multi-scale feature representation is obtained by feature fusion through feature splicing, so as to enhance the recognition ability of the model to different scale fault structures.

6. The method of claim 1, wherein, The step five comprises that in the visualization stage, the X-direction middle section, the Y-direction middle section and the Z-direction middle section are selected as three-way angular sections, the fault result predicted by the seismic fault prediction model is fused and displayed with the original seismic body data, and three-dimensional image drawing and presentation are carried out based on the Matplotlib drawing library, so as to realize intuitive identification and spatial structure visualization of potential fault area.

Images