Seismic data fracture identification method based on DenseVNet multi-scale deep learning
By using the DenseVNet multi-scale deep learning method, a nonlinear mapping model between seismic data and fracture labels is established, which solves the problem of insufficient accuracy in small fracture identification in existing technologies and realizes accurate identification and automated extraction of fractures at different scales.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- EAST CHINA UNIV OF TECH
- Filing Date
- 2026-01-31
- Publication Date
- 2026-05-12
AI Technical Summary
Existing fracture identification technologies are not sensitive to small fractures, have insufficient identification accuracy, and low automation, making it difficult to effectively identify small fractures in seismic data.
A fault identification method based on DenseVNet multi-scale deep learning is adopted for seismic data. Through multi-scale feature extraction, encoder and decoder modules, a nonlinear mapping relationship model between seismic data and fault labels is established. The DenseVNet multi-scale deep learning network structure is used to extract and identify fault features at different scales.
It achieves accurate identification of large, medium and small faults, improves the identification accuracy and automation of small faults, makes full use of the spatial and local features of seismic data, and adapts to the local high resolution of small faults and the global low resolution features of large and medium faults.
Smart Images
Figure CN122017988A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of seismic data fracture identification technology, specifically a seismic data fracture identification method based on DenseVNet multi-scale deep learning. Background Technology
[0002] Fault identification in seismic data is a crucial step in seismic interpretation and oil and gas exploration. Fault systems not only control tectonic evolution and sedimentary patterns but also directly influence hydrocarbon migration, accumulation, and preservation. Large and medium-sized faults determine the tectonic trend, attitude, and shape of oil-bearing fault blocks, while small faults further control local micro-scale structures, segment oil-bearing fault blocks, and complicate oil-water relationships within them. Large faults have a lateral extension exceeding 1 km and a displacement greater than 20 m; medium-sized faults have a lateral extension not exceeding 1 km and a displacement less than 20 m; and small faults typically have a lateral extension not exceeding 500 m and a displacement less than 10 m.
[0003] As oilfield exploration and development deepens, the focus of fracture identification has shifted to small fractures or fracture zones, aiming to identify as many small fractures as possible while still identifying large and medium-sized fractures. Small fractures have three key characteristics: first, they are difficult to detect using conventional fracture identification standards in seismic data from routine exploration; second, they are relatively small in scale in terms of displacement and extension length; and third, they have short horizontal extensions, allowing them to transport oil and gas or laterally seal them, thus controlling reservoir formation. On seismic profiles, small fractures typically exhibit characteristics such as minute displacements, distortion of the same phase axis, or weakened amplitude, making them highly concealed and difficult to identify. These concealed fracture systems have a significant impact on the distribution of remaining oil, the stimulation of unconventional reservoirs, and well network deployment.
[0004] Currently, fracture identification methods are mainly divided into three categories: The first category is traditional identification methods based on seismic attributes, mainly including techniques such as coherence volume, curvature, variance, and ant volume. These methods detect faults based on the geometric, amplitude, or phase characteristics of seismic signals. For example, coherence volume identifies discontinuous regions by calculating the similarity between adjacent traces; curvature attributes infer fault locations using changes in stratigraphic curvature; and ant volume technology tracks fault trajectories by simulating ant colony foraging behavior. Although these methods are helpful for fault detection to some extent, they still have significant limitations. Ant volume technology simulates ant foraging, and while it is relatively sensitive to small faults, it is also most susceptible to interference, such as significant noise interference and multiple solutions. Coherence volume, curvature, and variance are based on the differences between seismic phase axes corresponding to faults, and the calculation of these differences requires the use of time windows. The smoothing effect of time windows affects the resolution and accuracy of identifying small fault seismic response features such as minor shifts in phase axes, resulting in poor sensitivity to small faults.
[0005] The second category is fracture recognition methods based on edge detection and image processing. These methods treat seismic data as images and employ techniques such as edge detection, gradient enhancement, and structural tensors to enhance fractures. While they may have some visual effect, they often lack the spatial continuity constraints in a geological sense and struggle to distinguish fracture responses from lithological changes, noise, and other interference.
[0006] The third category is fracture identification methods based on deep learning. With the development of deep learning technology, structures such as Convolutional Neural Networks (CNN) and U-Net have been initially applied to earthquake interpretation and fracture detection. These methods can automatically learn features from data and have shown superior performance compared to traditional methods on some publicly available datasets. However, existing deep learning fracture identification methods still have the following problems: insufficient detection capability for small targets (such as small fractures) and serious loss of details.
[0007] In summary, while existing fracture identification methods have improved the effectiveness and accuracy of fracture identification, they require high-quality seismic data, are not sensitive to small fractures, and are not effective in identifying small fractures with small displacements. Summary of the Invention
[0008] The purpose of this invention is to provide a seismic data fracture identification method based on DenseVNet multi-scale deep learning, in order to solve the problems of insufficient accuracy, low automation, and limited ability to detect small fractures in existing fracture identification technologies.
[0009] The technical solution of this invention is: A seismic data fracture identification method based on DenseVNet multi-scale deep learning includes the following steps: Acquire actual seismic data volume of the work area and input the actual seismic data volume of the work area into a pre-trained nonlinear mapping relationship model to output fracture identification results; The nonlinear mapping model includes an input layer, a multi-scale feature extraction module, an encoder module, a decoder module, and an output layer connected in sequence. The multi-scale feature extraction module is used to extract fracture features at different scales from the actual seismic data volume of the work area input through the input layer. The fracture features at different scales are spliced together, and the encoder module mines and reduces the dimensionality of the spliced multi-scale features to generate multi-level encoded features containing semantic information at different levels. The decoder module is used to restore and fuse the generated multi-level encoded features respectively, and finally outputs the fracture identification results through the output layer.
[0010] Preferably, as a further improvement of the present invention, the multi-scale feature extraction module includes a first-scale convolutional layer, a second-scale convolutional layer, and a third-scale convolutional layer arranged in parallel. The first-scale convolutional layer is used to extract a first fracture feature, the second-scale convolutional layer is used to extract a second fracture feature, and the third-scale convolutional layer is used to extract a third fracture feature. The sizes of the first, second, and third fracture features increase sequentially. The kernel sizes of the first, second, and third-scale convolutional layers are 3, 5, and 7, respectively, the stride is 2, the number of filters is 32, and the activation function is ReLU.
[0011] Preferably, as a further improvement of the present invention, the encoder module includes three dense layers and two sampling layers. The first dense layer is used to receive spliced multi-scale features. The first downsampling layer performs average pooling on the output of the first dense layer to obtain a first feature. The first feature is input into the second dense layer. The second downsampling layer performs average pooling on the output of the second dense layer again to obtain a second feature. The second feature is input into the third dense layer to obtain a third feature. The decoder module includes four convolutional blocks and three upsampling layers. The first feature is connected to the first convolutional block to obtain the fourth feature. The second feature is connected to the second convolutional block and then to the first upsampling layer to obtain the fifth feature. The third feature is connected to the third convolutional block and then to the second upsampling layer to obtain the sixth feature. The fourth, fifth, and sixth features are concatenated, and the concatenated result is connected to the fourth convolutional block. The output of the fourth convolutional block is connected to the third upsampling layer, and the upsampled result is connected to the output layer.
[0012] Preferably, as a further improvement of the present invention, the first dense layer contains 2 convolutional loops and has 64 filters; the second dense layer contains 3 convolutional loops and has 128 filters; the third dense layer contains 4 convolutional loops and has 256 filters; the kernel size of the inner convolutional layers of the first, second, and third dense layers is 3, and skip connections are used.
[0013] Preferably, as a further improvement of the present invention, the first dense layer contains 2 convolutional loops and has 64 filters; the second dense layer contains 3 convolutional loops and has 128 filters; the third dense layer contains 4 convolutional loops and has 256 filters; the kernel size of the inner convolutional layers of the first, second, and third dense layers is 3, and skip connections are used.
[0014] Preferably, as a further improvement of the present invention, the pooling scale of the first downsampling layer and the second downsampling layer is 2, the upsampling scale of the first upsampling layer and the third upsampling layer is 2, and the upsampling scale of the second upsampling layer is 4.
[0015] Preferably, as a further improvement of the present invention, each of the first to fourth convolutional blocks is composed of two convolutional layers connected sequentially. The first convolutional block is used to refine the first feature, and its two convolutional layer parameters are as follows: the first convolutional layer has 64 filters, 1 kernel, 1 stride, and ReLU activation function; the second convolutional layer has 64 filters, 3 kernels, 1 stride, and ReLU activation function. The second convolutional block is used to refine the second feature, and its two convolutional layer parameters are as follows: the first convolutional layer has 128 filters, 1 kernel, 1 stride, and ReLU activation function; the second convolutional layer has 128 filters and 3 kernels. The first convolutional block has a stride of 1 and an activation function of ReLU. The second convolutional block is used to refine the third feature. Its two convolutional layer parameters are as follows: the first convolutional layer has 256 filters, 1 kernel, a stride of 1, and an activation function of ReLU; the second convolutional layer has 256 filters, 3 kernels, a stride of 1, and an activation function of ReLU. The third convolutional block is used to refine the concatenated features of the fourth, fifth, and sixth features. Its two convolutional layer parameters are as follows: the first convolutional layer has 32 filters, 1 kernel, a stride of 1, and an activation function of ReLU; the second convolutional layer has 32 filters, 3 kernels, a stride of 1, and an activation function of ReLU.
[0016] Preferably, as a further improvement of the present invention, the nonlinear mapping relationship model is trained based on the fracture synthetic seismic dataset. During training, the fracture synthetic seismic dataset is first divided into a training dataset and a validation dataset according to a preset ratio. After training the nonlinear mapping relationship model based on the training dataset, it is trained again using the validation dataset. When the loss function of the nonlinear mapping relationship model meets the preset accuracy requirement, the training ends and the trained nonlinear mapping relationship model is obtained.
[0017] Preferably, as a further improvement of the present invention, the process of acquiring the fracture synthetic seismic dataset includes the following steps: firstly, acquiring fracture synthetic seismic data volume to form an initial fracture synthetic seismic dataset; then, performing normalization and data augmentation processing on the initial fracture synthetic seismic dataset to obtain the fracture synthetic seismic dataset after preprocessing; wherein, the data augmentation processing adopts at least one of noise addition, inversion, and mirroring.
[0018] Preferably, as a further improvement of the present invention, the loss function is a balanced cross-entropy loss function, which is determined based on the following formula: ; In the formula, This represents the Focal cross-entropy loss value. To balance the cross-entropy loss value, The loss value is the result of fusing the Focal cross-entropy loss value and the balanced cross-entropy loss value. The weights are the Focal cross-entropy loss values. To balance the weights of the cross-entropy loss value Preferably, as a further improvement of the present invention, the training of the DenseVNet multi-scale deep learning network structure adopts the Adam optimizer, the learning rate is set to 1e-4, the batch size is 1, the number of iterations is greater than or equal to 100, and when the loss function of the validation set has been iterated for 10 consecutive times and the value of the loss function has not decreased, the training is stopped and the current nonlinear mapping relationship model is saved.
[0019] Compared with the prior art, the beneficial effects of the present invention are: 0. Based on the characteristics of fractures, a DenseVNet multi-scale deep learning method was designed. The deep learning network structure, model parameters and loss function were studied to form a systematic identification and characterization of fractures in the study area. A nonlinear mapping relationship model between seismic data and fracture label data was established based on the DenseVNet multi-scale deep learning network structure and applied to fracture identification of seismic data. It can make full use of the spatial and local features of seismic data to realize intelligent identification and extraction of fractures at multiple scales.
[0020] 2. The established nonlinear mapping relationship model generates local high-resolution features suitable for small fracture identification and global low-resolution features suitable for large and medium fracture identification through the cooperation of multi-scale feature extraction module and encoder module. Then, the feature resolution is unified through decoder module, and finally the accurate identification of fractures of different scales (especially hidden small fractures) is achieved. Attached Figure Description
[0021] Figure 1 This is a flowchart illustrating a seismic data fracture identification method based on DenseVNet multi-scale deep learning according to the present invention.
[0022] Figure 2 This is a schematic diagram of the DenseVNet multi-scale deep learning network structure in this invention.
[0023] Figure 3 This is a schematic diagram of the quality control of model training parameters in this invention.
[0024] Figure 4(a) is a schematic diagram of the seismic data profile along the main measuring line of the present invention.
[0025] Figure 4(b) is a schematic diagram of the fracture identification results under the main test line of the present invention.
[0026] Figure 4(c) is a schematic diagram of the coherent body fracture identification results at the location of the survey line corresponding to Figure 4(b).
[0027] Figure 5(a) is a schematic diagram of the seismic data profile along the connection line of the present invention.
[0028] Figure 5(b) is a schematic diagram of the breakage identification results of the connecting line of the present invention.
[0029] Figure 5(c) is a schematic diagram of the coherent body fracture identification results at the location of the survey line corresponding to Figure 5(b).
[0030] Figure 6(a) is a schematic diagram of the horizontal slice of seismic data in this invention.
[0031] Figure 6(b) is a schematic diagram of the horizontal slice of the fracture identification results of the present invention.
[0032] Figure 6(c) is a schematic diagram of the coherent body fracture identification results corresponding to the horizontal slice position in Figure 6(b).
[0033] Figure 7(a) is a schematic diagram of the seismic data of carbonate rock formations in this invention.
[0034] Figure 7(b) is a schematic diagram of the fracture identification results of carbonate rock formations in this invention.
[0035] Figure 7(c) is a schematic diagram of the fracture identification results of the carbonate rock strata coherent body at the location corresponding to Figure 7(b). Detailed Implementation
[0036] The following is combined Figure 1 Figure 7 illustrates a detailed description of specific embodiments of the present invention. In the description of the invention, it should be understood that the terms "center," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.
[0037] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature; in the description of the invention, unless otherwise stated, "a plurality of" means two or more.
[0038] Example like Figure 1 As shown in Figure 7, this embodiment of the invention provides a seismic data fracture identification method based on DenseVNet multi-scale deep learning, including the following steps: S1. Obtain the seismic data volume, which includes the fault-synthesized seismic data volume and the actual seismic data volume of the work area.
[0039] The fault-synthetic seismic data volume is a data volume generated through forward modeling of fault patterns. During acquisition, a geological model is established based on existing fault patterns, and the observed seismic response data is simulated using physical equations (such as wave equations) to obtain the fault-synthetic seismic data volume, which is used for model construction and training.
[0040] The acquisition of actual seismic data in the work area comprises two stages. The first stage involves data acquisition: seismic sources (such as explosives, vibration vehicles, or air guns) and receivers (geophones) are deployed in the work area (land or sea). The seismic sources generate seismic waves, which are reflected by the underground rock strata and recorded by the receivers. This yields raw single-shot data, recording the most primitive vibration signals; the data volume is large and the signal-to-noise ratio is low. The second stage involves indoor seismic data processing: the acquired single-shot data undergoes a series of processing steps, including defining the observation system, denoising, deconvolution, static correction, velocity analysis, and migration, to form seismic data that reflects the underground structural morphology. This data is typically in Segy format. The actual seismic data volume in the work area serves as the input parameter for the model, which outputs fracture identification results.
[0041] S2. A nonlinear mapping model between seismic data and fault labels is established using the DenseVNet multi-scale deep learning network structure.
[0042] Specifically, large and medium-sized faults determine the structural trend, attitude, and shape of oil-bearing fault blocks, while small faults further control local micro-scale structures, segment oil-bearing fault blocks, and complicate the oil-water relationship within them. With the deepening of oilfield exploration and development, the focus of fault identification has shifted to small faults or fault fracture zones, aiming to identify as many small faults as possible while still identifying large and medium-sized faults. Large and medium-sized faults refer to those that can be identified using conventional geophysical methods. Small faults have three aspects: first, they are difficult to detect using standard fault identification criteria in conventional seismic data; second, they are relatively small in scale in terms of displacement and extension length; and third, they have short horizontal extensions, allowing them to transport oil and gas or laterally block them, controlling reservoir formation. On seismic profiles, small faults typically exhibit characteristics such as minute displacements, in-phase axis distortion, or weakened amplitude, making them highly concealed. Small fractures typically have a lateral extension length of no more than 500m and a displacement of less than 10m, making them difficult to identify. Medium fractures have a lateral extension length of no more than 1km and a displacement of less than 20m. Large fractures have a lateral extension length of more than 1km and a displacement of more than 20m. DenseVNet multi-scale deep learning is a convolutional neural network architecture that improves network feature extraction capabilities, information utilization efficiency, and multi-level target extraction resolution through dense connections. It is suitable for identifying large, medium, and small targets, and for fracture identification, it is suitable for simultaneously identifying large, medium, and small fractures. The input data for the DenseVNet multi-scale deep learning network structure is a fracture synthetic seismic dataset, consisting of downsampling contraction paths and upsampling expansion paths. Specifically, it includes sequentially connected input layers, multi-scale feature extraction modules, encoder modules, decoder modules, and output layers (the principle is as follows...). Figure 2(As shown). The DenseVNet multi-scale deep learning downsampling subnetwork is a sequence of three dense layers connected by downsampling average pooling layers; each skip connection is a single convolution of the corresponding dense layer output, and the upsampling network includes upsampling of the final segmentation resolution. The storage efficiency and batch space discarding of dense layers enable deeper networks to have higher resolution, which is beneficial for segmenting smaller structures. Downsampling can produce local high-resolution features in early layers as well as global low-resolution features. Global low-resolution features can identify large and medium-sized fractures, while local high-resolution features help identify small fractures. Therefore, after model training or generalization, large, medium, and small fractures can all be identified. By designing the network structure and optimizing the parameter configuration in a targeted manner, the model can better adapt to the inherent characteristics of seismic data and effectively meet the specific needs of fracture identification tasks. The DenseVNet multi-scale deep learning network structure is described as follows: First, input the fracture synthetic seismic data volume. Then, connect the input data to three convolutional layers of different scales (First-scale convolutional layer: filter 32, kernel 3, stride 2, ReLU activation function; this corresponds to small-scale target extraction, i.e., small fracture identification; Second-scale convolutional layer: filter 32, kernel 5, stride 2, ReLU activation function; this corresponds to medium-scale target extraction, i.e., medium fracture identification; Third-scale convolutional layer: filter 32, kernel 7, stride 2, ReLU activation function; this corresponds to large-scale target extraction, i.e., large fracture identification). Then, concatenate the calculation results of these three convolutional layers. The output of this layer passes through the first dense layer (two layers of looped, skip connections, filter 64, kernel 3; in this dense layer, the calculation result of the first layer is used as the input of the second layer). Finally, the data calculated by the dense layer is obtained. According to the data, the output of the first dense layer is then processed by average pooling (the first downsampling layer) with a pooling scale of 2. The result of this average pooling is denoted as the first feature. The first feature is then input into the second dense layer (3 layers of recurrent and skip connections, 128 filters, 3 convolutional kernels; in this dense layer, the result of the first layer is used as the input of the second layer, the result of the second layer is used as the input of the third layer, and finally the data calculated by the dense layer is obtained). The output of the second dense layer is then processed by average pooling again (the second downsampling layer) with a pooling scale of 2. The result of this average pooling is denoted as the second feature. The second feature is then input into the third dense layer (4 layers of recurrent and skip connections, 256 filters, 3 convolutional kernels; in this dense layer, the result of the first layer is used as the input of the second layer, the result of the second layer is used as the input of the third layer, and the result of the third layer is used as the input of the fourth layer, and finally the data calculated by the dense layer is obtained). The result of this layer is denoted as the third feature.Meanwhile, the first feature is connected to the first convolutional block (consisting of 1 convolutional layer [filter 64, kernel 1, stride 1, activation function ReLU] + 1 convolutional layer [filter 64, kernel 3, stride 1, activation function ReLU]), and the calculated result is denoted as the fourth feature; the second feature is connected to the second convolutional block (consisting of 1 convolutional layer [filter 128, kernel 1, stride 1, activation function ReLU] + 1 convolutional layer [filter 128, kernel 3, stride 1, activation function ReLU]), and based on this, in the first upsampling layer with a scale of 2, after upsampling, the calculated result is denoted as the fifth feature; the third feature is connected to the third convolutional block (consisting of 1 convolutional layer [filter 256, kernel 1, stride 1, activation function ReLU] + 1 convolutional layer [filter 256, kernel 3, stride 1, activation function ReLU]), and based on this, in the second upsampling layer with a scale of 4, after upsampling, the calculated result is denoted as the sixth feature. Based on the above, the fourth, fifth, and sixth features are concatenated, and the concatenated data is connected to the fourth convolutional block (consisting of one convolutional layer [filter 32, kernel 1, stride 1, activation function ReLU] + one convolutional layer [filter 32, kernel 3, stride 1, activation function ReLU]). The data calculated by this convolutional block is upsampled using a third upsampling layer with a scale of 2. The upsampled result is then connected to the output layer (a convolutional layer, filter 1, stride 1, activation function Sigmoid) to obtain the output data volume.
[0043] S3. Train the established nonlinear mapping relationship model.
[0044] Specifically, the collected fracture synthetic seismic data volume is used to form an initial fracture synthetic seismic dataset; then, the initial fracture synthetic seismic dataset is normalized and data augmented to obtain the fracture synthetic seismic dataset after preprocessing; wherein, the data augmentation process adopts at least one of the following methods: adding noise, inverting, and mirroring.
[0045] During training, the fracture synthetic earthquake dataset was first divided into training and validation datasets in a 7:3 ratio. Using the deep learning library in the Python platform, the model was trained based on the relationship model between earthquake and fracture data established by the DenseVNet multi-scale deep learning network structure in this invention. The model optimizer was Adam, the learning rate was 1e-4, the batch size was 1, the number of layers in the DenseVNet multi-scale network was 3, the loss function was the balanced cross-entropy loss function, and the number of iterations was 100. When the validation set loss did not decrease for 10 consecutive iterations, training was stopped, and the current optimal model was saved.
[0046] The balanced cross-entropy loss function is used to address the problem of extremely imbalanced fault samples in seismic data, improving the model's sensitivity to fault points. Its construction process is as follows: The formula for the balanced cross-entropy loss function is as follows: (1) (2) The formula for the Focal cross-entropy loss function is as follows: (3) The formula for focal_cross_entropy_balanced (balanced cross-entropy loss function), which combines the Focal cross-entropy loss function and the balanced cross-entropy loss function, is as follows: (4) In the formula: This represents the Focal cross-entropy loss value. To balance the cross-entropy loss value, The loss value is the result of fusing the Focal cross-entropy loss value and the balanced cross-entropy loss value. The weights are the Focal cross-entropy loss values. To balance the weights of the cross-entropy loss values, To balance the weighting factors of positive and negative samples, Adjustable focus parameters (usually) ), For the probability of fracture prediction, For the first A broken label, This represents the number of earthquake data samples. The first representative of earthquake data The position index of each sample point This represents the ratio of unbroken data to total data. This indicates the proportion of broken data in the total data. Increase This allows the model to focus more on learning from fewer labeled samples during training, making it suitable for imbalanced positive and negative samples. In the training set, the number of non-breakpoints (label 0) is much greater than or equal to the number of breakpoints (label 1), thus enabling its application... It can enhance the model's ability to learn about fracture points, thus enabling better fracture identification.
[0047] The formula for establishing the nonlinear relationship between seismic data and fault data using the DenseVNet multi-scale network structure is as follows: (5) The multi-scale neural network structure established during model building. For earthquake data, These are the network parameters for the DenseVNet multi-scale network structure. The results are for fracture identification data predicted by the DenseVNet multi-scale network.
[0048] After training the nonlinear mapping relationship model based on the training dataset, it is trained again using the validation dataset. When the loss function of the nonlinear mapping relationship model meets the preset accuracy requirements, the training ends and the trained nonlinear mapping relationship model is obtained.
[0049] S4. Input the actual seismic data of the work area into the trained nonlinear mapping relationship model, and output the fracture identification results.
[0050] Specifically, the trained DenseVNet multi-scale network structure model is applied to seismic data containing fractures to be identified, outputting fracture identification results as fracture probabilities. The model's performance is evaluated using quality control plots (such as loss curves). After training, the model's training parameters are checked for suitability. If unsuitable, the number of iterations during training is adjusted, and the model is retrained. If the training parameters are acceptable, the model's training results are applied to actual seismic data. Finally, visualized fracture identification results (profiles, horizontal slices, etc.) are output. Quality control of model training parameters is as follows: Figure 3 As shown, from Figure 3 As can be seen from the figure, during the model training process, the highest accuracy of both training and validation exceeded 93%, rising from 60% to over 93%, with the accuracy curves showing a smooth and stable increase. The model training and validation loss value decreased from 0.4 to approximately 0.02, with the loss curve showing a smooth and stable decrease. During model training, there were no instances of non-convergence or gradient explosion; the model converged and underwent effective training. The figure demonstrates that the model and the parameters used in this invention are of high quality, and the seismic data obtained through this invention... The identified fracture results are shown in Figures 4, 5, 6, and 7. As can be seen from Figures 4-7, the fracture results identified by the present invention, as shown in Figures 4(b), 5(b), 6(b), and 7(b), match the original seismic data, as shown in Figures 4(a), 5(a), 6(a), and 7(a), to a high degree. The fracture features identified by the present invention are clear and obvious. In contrast, the fractures identified by the coherence volume in Figures 4(c), 5(c), 6(c), and 7(c) are blurry and the features are not obvious. This demonstrates that the method of the present invention is effective in actual work areas.
[0051] In summary, this invention, based on a fully convolutional neural network deep learning method using artificial intelligence technology, provides a new approach to fracture identification. Based on fracture characteristics, a DenseVNet multi-scale deep learning method was designed, and the deep learning network structure, model parameters, and loss function were studied to form a systematic identification and characterization of fractures in the study area. Applying DenseVNet multi-scale deep learning to fracture identification in seismic data can fully utilize the spatial and local features of seismic data to achieve intelligent identification and extraction of multi-scale fractures (large, medium, and small fractures). This fully convolutional neural network deep learning method based on artificial intelligence technology provides a new approach to fracture identification.
[0052] The above-disclosed embodiments are merely preferred embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.
Claims
1. A seismic data fracture identification method based on DenseVNet multi-scale deep learning, characterized in that, Includes the following steps: Acquire actual seismic data volume of the work area and input the actual seismic data volume of the work area into a pre-trained nonlinear mapping relationship model to output fracture identification results; The nonlinear mapping model includes an input layer, a multi-scale feature extraction module, an encoder module, a decoder module, and an output layer connected in sequence. The multi-scale feature extraction module is used to extract fracture features at different scales from the actual seismic data volume of the work area input through the input layer. The fracture features at different scales are spliced together, and the encoder module mines and reduces the dimensionality of the spliced multi-scale features to generate multi-level encoded features containing semantic information at different levels. The decoder module is used to restore and fuse the generated multi-level encoded features respectively, and finally outputs the fracture identification results through the output layer.
2. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 1, characterized in that, The multi-scale feature extraction module includes a first-scale convolutional layer, a second-scale convolutional layer, and a third-scale convolutional layer set in parallel. The first-scale convolutional layer is used to extract the first fracture feature, the second-scale convolutional layer is used to extract the second fracture feature, and the third-scale convolutional layer is used to extract the third fracture feature. The size of the first fracture feature, the second fracture feature, and the third fracture feature increases sequentially. The kernel sizes of the first-scale convolutional layer, the second-scale convolutional layer, and the third-scale convolutional layer are 3, 5, and 7, respectively, the stride is 2, the number of filters is 32, and the activation function is ReLU.
3. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 2, characterized in that, The encoder module includes three dense layers and two sampling layers. The first dense layer is used to receive spliced multi-scale features. The first downsampling layer performs average pooling on the output of the first dense layer to obtain the first feature. The first feature is then input into the second dense layer. The second downsampling layer performs average pooling on the output of the second dense layer again to obtain the second feature. The second feature is then input into the third dense layer to obtain the third feature. The decoder module includes four convolutional blocks and three upsampling layers. The first feature is connected to the first convolutional block to obtain the fourth feature. The second feature is connected to the second convolutional block and then to the first upsampling layer to obtain the fifth feature. The third feature is connected to the third convolutional block and then to the second upsampling layer to obtain the sixth feature. The fourth, fifth, and sixth features are concatenated, and the concatenated result is connected to the fourth convolutional block. The output of the fourth convolutional block is connected to the third upsampling layer, and the upsampled result is connected to the output layer.
4. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 3, characterized in that, The first dense layer contains 2 convolutional loops and has 64 filters. The second dense layer contains 3 convolutional loops and has 128 filters. The third dense layer contains 4 convolutional loops and has 256 filters. The convolutional kernel size of the inner convolutional layers of the first, second, and third dense layers is 3, and skip connections are used.
5. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 3, characterized in that, The pooling scale of the first downsampling layer and the second downsampling layer is 2, the upsampling scale of the first upsampling layer and the third upsampling layer is 2, and the upsampling scale of the second upsampling layer is 4.
6. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 3, characterized in that, Each of the first to fourth convolutional blocks consists of two convolutional layers connected sequentially. The first convolutional block refines the first feature, with the following parameters for its two convolutional layers: 64 filters, 1 kernel, 1 stride, and ReLU activation function; and 64 filters, 3 kernels, 1 stride, and ReLU activation function. The second convolutional block refines the second feature, with the following parameters for its two convolutional layers: 128 filters, 1 kernel, 1 stride, and ReLU activation function; and 128 filters, 3 kernels, 1 stride, and ReLU activation function. The first convolutional block uses ReLU as the activation function; the second convolutional block is used to refine the third feature, and its two convolutional layer parameters are as follows: the first convolutional layer has 256 filters, 1 kernel, 1 stride, and ReLU activation function; the second convolutional layer has 256 filters, 3 kernels, 1 stride, and ReLU activation function; the third convolutional block is used to refine the concatenated features of the fourth, fifth, and sixth features, and its two convolutional layer parameters are as follows: the first convolutional layer has 32 filters, 1 kernel, 1 stride, and ReLU activation function; the second convolutional layer has 32 filters, 3 kernels, 1 stride, and ReLU activation function.
7. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 1, characterized in that, The nonlinear mapping relationship model is trained based on the fracture synthetic seismic dataset. During training, the fracture synthetic seismic dataset is first divided into a training dataset and a validation dataset according to a preset ratio. After training the nonlinear mapping relationship model based on the training dataset, it is trained again using the validation dataset. When the loss function of the nonlinear mapping relationship model meets the preset accuracy requirements, the training ends and the trained nonlinear mapping relationship model is obtained.
8. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 7, characterized in that, The process of acquiring the fracture synthetic seismic dataset includes the following steps: first, collecting fracture synthetic seismic data to form an initial fracture synthetic seismic dataset; then, normalizing and data augmenting the initial fracture synthetic seismic dataset to obtain the fracture synthetic seismic dataset after preprocessing; wherein, the data augmentation process employs at least one of noise addition, inversion, and mirroring.
9. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 7, characterized in that, The loss function is the balanced cross-entropy loss function, which is determined based on the following formula: ； In the formula, This represents the Focal cross-entropy loss value. To balance the cross-entropy loss value, The loss value is the result of fusing the Focal cross-entropy loss value and the balanced cross-entropy loss value. The weights are the Focal cross-entropy loss values. To balance the weights of the cross-entropy loss values.
10. The seismic data fracture identification method based on DenseVNet multi-scale deep learning according to claim 9, characterized in that, The nonlinear mapping model is trained using the Adam optimizer with a learning rate of 1e-4, a batch size of 1, and an iteration count of 100 or more. Training stops and the current nonlinear mapping model is saved when the loss function on the validation set has not decreased after 10 consecutive iterations.