Geological structure spatial description method and device based on deep learning

Abstract

The application provides a geological structure space description method and device based on deep learning, and relates to the technical field of seismic exploration. The method first calculates the 3D dip angle domain offset gather horizontal slice at each imaging point based on the seismic observation data of the region to be imaged, then uses a first deep learning convolutional neural network to automatically identify the horizontal slice at the imaging point to distinguish between structural mutation points and continuous reflection points, and realizes fine discrimination of geological structure features. On this basis, a special second deep learning convolutional neural network and a third deep learning convolutional neural network are respectively used for different types of imaging points to accurately identify the local dip angle of the continuous reflection interface and the spatial distribution azimuth angle of the fracture in the image point dimension, thereby enhancing the spatial depiction precision of key geological bodies such as faults and fractures, and thus effectively improving the resolution of the geological structure space description.

Description

Technical Field

[0001] This invention relates to the field of seismic exploration technology, and in particular to a method and apparatus for spatial description of geological structures based on deep learning. Background Technology

[0002] In seismic exploration, seismic imaging, which involves migrating surface seismic observation data to obtain information about complex subsurface structures, is primarily used in industrial applications for describing subsurface structures and lithology in oil and gas exploration. However, with the shale oil and gas revolution driving energy structure transformation, the industry has increasingly higher requirements for describing micro-structures such as fractures and small faults in reservoirs. Therefore, the low accuracy of seismic imaging for continental shale has long been a core problem hindering the efficient exploitation of shale oil and gas. The resolution limit of traditional reflected wave imaging is one-quarter of the dominant frequency wavelength; even with various frequency domain protection or extension methods, its resolution still cannot meet the needs of actual production. Therefore, there is an urgent need for a technical solution that can improve the accuracy of subsurface micro-structure imaging. Summary of the Invention

[0003] The purpose of this invention is to provide a method and apparatus for spatial description of geological structures based on deep learning, so as to alleviate the technical problem of low accuracy in the imaging of underground microstructures in the prior art.

[0004] In a first aspect, the present invention provides a deep learning-based method for spatial description of geological structures, comprising: acquiring seismic observation data of an area to be imaged; performing migration imaging of the imaging space of the area to be imaged based on the seismic observation data to obtain a 3D dip-domain migration gather of the imaging space, and extracting horizontal slices of the 3D dip-domain migration gather at each imaging point in the imaging space; using a first deep learning convolutional neural network to perform type identification on the horizontal slices of the 3D dip-domain migration gather at the target imaging point, and obtaining classification labels for the horizontal slices; the target imaging point represents any imaging point in the imaging space; the classification labels for the horizontal slices include: horizontal slices of tectonic abrupt change points and horizontal slices of continuous reflection points. Slices are used to identify the location of stable phase points in the 3D dip-domain offset gather horizontal slices at the target imaging point. The coordinates of the stable phase points are used as the local dip angle of the continuous reflection interface at the target imaging point. In the case of horizontal slices labeled as tectonic abrupt change points, a third deep learning convolutional neural network is used to identify the fracture orientation in the 3D dip-domain offset gather horizontal slices at the target imaging point. The spatial distribution azimuth of the fracture at the target imaging point is determined based on the fracture orientation identification results. Based on the identification results of all horizontal slices, a quantitative spatial description of the geological structure of the area to be imaged is constructed.

[0005] In an optional implementation, the first deep learning convolutional neural network includes a lightweight convolutional neural network (CNN), which includes a feature extraction module and a classification decision module. The feature extraction module is used to extract features from the horizontal slices of the 3D tilt-domain offset gather to obtain a high-dimensional feature vector. The classification decision module is used to map the high-dimensional feature vector to a binary classification class space and use the class with the highest probability as the classification label of the horizontal slice of the 3D tilt-domain offset gather.

[0006] In an optional implementation, the second deep learning convolutional neural network includes: an improved U-Net network, which includes: a first encoder and a first decoder; the first encoder consists of multiple convolutional blocks with LeakyReLU activation functions and a convolutional downsampling module with a stride of 2; the first decoder is skip-connected to the feature maps of the corresponding layers of the first encoder; the first encoder is used to extract multi-scale spatial features of the horizontal slices of the 3D tilt domain offset gather layer by layer; the first decoder is used to restore the spatial details of the horizontal slices of the 3D tilt domain offset gather layer by layer based on the multi-scale spatial features, generate a pixel-level probabilistic heatmap, and use the extreme point positions in the pixel-level probabilistic heatmap as the stable phase point positions.

[0007] In an optional implementation, the third deep learning convolutional neural network includes a Swin-Unet network, which includes a second encoder, a bottleneck layer, and a second decoder. The second decoder and the feature maps of the corresponding layers of the second encoder are connected by skip connections. The second encoder is used to extract deep semantic features from the horizontal slices of the 3D tilt domain offset gather layer by layer. The bottleneck layer is used to abstract the deep semantic features into a compact high-dimensional semantic representation to obtain high-dimensional semantic features. The second decoder is used to generate a probability hotspot map indicating the phase change region based on the high-dimensional semantic features and the high-resolution feature maps of the corresponding layers of the encoding path.

[0008] In an optional implementation, the fracture orientation identification result includes: a probability heatmap indicating the phase change region; determining the spatial distribution azimuth of the fracture at the target imaging point based on the fracture orientation identification result, including: binarizing the probability heatmap to obtain the spatial location of the phase change; using the least squares fitting method to perform linear fitting on the spatial location of the phase change to obtain the phase change line and the spatial orientation it represents; and using the spatial orientation represented by the phase change line as the spatial distribution azimuth of the fracture at the target imaging point.

[0009] In an optional implementation, the first deep learning convolutional neural network, the second deep learning convolutional neural network, and the third deep learning convolutional neural network are all trained under supervised supervision; the training samples of the first deep learning convolutional neural network include: 3D tilt domain offset gather horizontal slices with classification labels of construction mutation point horizontal slices and 3D tilt domain offset gather horizontal slices with classification labels of continuous reflection point horizontal slices; the training samples of the second deep learning convolutional neural network include: 3D tilt domain offset gather horizontal slices with stable phase point position labels; the training samples of the third deep learning convolutional neural network include: 3D tilt domain offset gather horizontal slices with phase mutation line labels.

[0010] In an optional implementation, the training samples for the first deep learning convolutional neural network are 3D tilt domain offset gather horizontal slices corresponding to imaging points in drastically changing regions and continuous reflective surface regions.

[0011] Secondly, the present invention provides a deep learning-based device for spatial description of geological structures, comprising: an acquisition module for acquiring seismic observation data of an area to be imaged; an imaging module for performing migration imaging of the imaging space of the area to be imaged based on the seismic observation data, obtaining a 3D dip-domain migration gather of the imaging space, and extracting a horizontal slice of the 3D dip-domain migration gather at each imaging point in the imaging space; and a first identification module for performing type identification of the horizontal slice of the 3D dip-domain migration gather at the target imaging point using a first deep learning convolutional neural network, and obtaining a classification label for the horizontal slice; the target imaging point represents any imaging point in the imaging space; the classification labels for the horizontal slices include: horizontal slices of tectonic abrupt change points and horizontal slices of continuous reflection points. The system comprises four modules: a first module for identifying the location of a stable phase point in a 3D dip-domain offset gather horizontal slice labeled as a continuous reflection point; a second module for identifying the location of a stable phase point in a 3D dip-domain offset gather horizontal slice labeled as a continuous reflection point, using a second deep learning convolutional neural network to identify the location of a fracture in a 3D dip-domain offset gather horizontal slice labeled as a tectonic abrupt change point, using a third deep learning convolutional neural network to identify the location of a fracture in a 3D dip-domain offset gather horizontal slice labeled as a tectonic abrupt change point, using the location of a fracture in a 3D dip-domain offset gather horizontal slice labeled as a tectonic abrupt change point, and determining the spatial distribution azimuth of a fracture in a 3D dip-domain offset gather horizontal slice labeled as a tectonic abrupt change point; and a construction module for constructing a quantitative spatial description of the geological structure of the area to be imaged based on the identification results of all horizontal slices.

[0012] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the deep learning-based spatial description method for geological structures as described in any of the foregoing embodiments.

[0013] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the deep learning-based spatial description method for geological structures as described in any of the foregoing embodiments.

[0014] The deep learning-based spatial description method for geological structures provided by this invention first calculates 3D dip-domain migration gather horizontal slices at each imaging point based on seismic observation data of the area to be imaged. Then, a first deep learning convolutional neural network is used to automatically identify the horizontal slices at the imaging points to distinguish between tectonic abrupt changes and continuous reflection points, achieving refined discrimination of geological structural features. Building upon this, dedicated second and third deep learning convolutional neural networks are employed for different types of imaging points to accurately identify the local dip angles of continuous reflection interfaces and the spatial distribution azimuth angles of fractures at the image point dimension. This enhances the spatial characterization accuracy of key geological bodies such as faults and fractures, thus effectively improving the resolution of the spatial description of geological structures. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a deep learning-based spatial description method for geological structures, provided as an embodiment of the present invention;

[0017] Figure 2 A schematic diagram of the Fresnel band morphology on a horizontal slice corresponding to the imaging points at different structural locations in a three-dimensional tilt gather provided by an embodiment of the present invention;

[0018] Figure 3 A schematic diagram of a lightweight convolutional neural network (CNN) provided in an embodiment of the present invention;

[0019] Figure 4 A schematic diagram of training samples and prediction results of a second deep learning convolutional neural network provided in an embodiment of the present invention;

[0020] Figure 5 A schematic diagram of an improved U-Net network structure provided in an embodiment of the present invention;

[0021] Figure 6 A schematic diagram of training samples and prediction results of a third deep learning convolutional neural network provided in an embodiment of the present invention;

[0022] Figure 7 This is a schematic diagram of the network structure of an improved Swin-Unet network provided in an embodiment of the present invention;

[0023] Figure 8 In order to be in Figure 6 A schematic diagram showing the phase transition lines and their corresponding spatial distribution azimuth angles marked on the original horizontal slice;

[0024] Figure 9 A functional module diagram of a deep learning-based spatial description device for geological structures provided in an embodiment of the present invention;

[0025] Figure 10 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0028] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0029] Example 1

[0030] Traditional reflected wave imaging has a resolution limit of one-quarter of the dominant wavelength, which is insufficient for practical production needs. While existing seismic imaging methods (such as Kirchhoff migration and reverse-time migration) can provide images of subsurface structures, they suffer from problems like blurring and artifacts in areas sensitive to local dip changes (such as faults and fold transition zones). Furthermore, manual interpretation is inefficient and highly subjective. Therefore, this invention proposes a spatial description method for geological structures based on 3D dip-domain migration gather horizontal slicing and deep learning convolutional neural network collaborative identification, enabling end-to-end extraction of quantitative geological structural information from raw seismic observation data.

[0031] Figure 1A flowchart illustrating a deep learning-based spatial description method for geological structures, as provided in this embodiment of the invention, is shown below. Figure 1 As shown, the method specifically includes the following steps:

[0032] Step S102: Obtain seismic observation data of the area to be imaged.

[0033] At the outset of the method, users need to collect raw seismic observation data covering the target exploration area (i.e., the area to be imaged mentioned above) to provide the input basis for subsequent migration imaging. Seismic observation data refers to the surface signals received by a seismic detector array deployed in the field, which are reflected / refracted back by the subsurface medium after being excited by a seismic source. Generally, the raw seismic observation data needs to undergo preprocessing, including conventional seismic data processing procedures such as denoising, static correction, amplitude compensation, and dynamic correction, before being used in subsequent steps to ensure that its signal-to-noise ratio meets the imaging requirements.

[0034] Step S104: Based on the seismic observation data, perform migration imaging of the imaging space of the area to be imaged to obtain the 3D dip domain migration gather of the imaging space, and extract the horizontal slice of the 3D dip domain migration gather at each imaging point in the imaging space.

[0035] After acquiring seismic observation data, the first step is to perform high-precision migration imaging of the imaging space of the area to be imaged in order to generate a 3D dip domain migration gather for that imaging space, or simply a three-dimensional dip gather.

[0036] Specifically, if the imaging space contains 10 imaging lines, each imaging line containing 10 imaging lines. There are 10 Common Depth Points (CDPs), each containing 10 CDPs. There are n imaging points, where the coordinates of any one imaging point are represented as... Then its 3D tilt gather migration imaging result is: .

[0037] in, , When representing the two-way travel of seismic waves, and These are the seismic wave self-shot points. propagation to imaging point and propagation from the imaging point to the receiving point During the trip, This represents the first derivative of the seismic data sequence for the j-th seismic trace, where j is the trace number and k represents the total number of input seismic traces. Representing the imaging points respectively The corresponding dip angles in the x and y directions are calculated using the following formula: , , This represents the root mean square velocity of the imaging point.

[0038] Based on the expression of the 3D tilt gather migration imaging results, it can be seen that... It is a five-dimensional data volume, with dimensions of [sizes to be filled in]. , , , , .in, , These represent the number of preset tilt angles in the x-direction and the number of preset tilt angles in the y-direction, respectively. If... Unchanged, only The change refers to obtaining the imaging results at the same CDP, which is a three-dimensional dip gather of a seismic trace. This dip gather is a three-dimensional data volume. If All remain unchanged, only Changes can be used to obtain imaging points. A horizontal slice (also called a depth slice, which is essentially a two-dimensional image) of a three-dimensional dip gather will show a complete and smooth concentric ellipse on the horizontal slice if the subsurface interface is continuous. However, the Fresnel zone of the reflected wave at the tectonic abrupt change will show a spatial phase abrupt change on the horizontal slice of the three-dimensional dip gather. For example, there may be a phase reversal within the Fresnel zone or it may be a typical two-part unconformity ellipse. Figure 2 This is a schematic diagram of the Fresnel band morphology on a horizontal slice corresponding to the imaging points at different structural locations, provided as an embodiment of the present invention.

[0039] Step S106: Use the first deep learning convolutional neural network to perform type recognition on the 3D tilt domain offset gather horizontal slice at the target imaging point to obtain the classification label of the horizontal slice; the target imaging point represents any imaging point in the imaging space.

[0040] After extracting the 3D tilt-domain offset gather horizontal slices at all imaging points, this embodiment of the invention first uses a pre-trained first deep learning convolutional neural network to perform type identification on each horizontal slice to determine the classification label of each horizontal slice. The classification labels for the horizontal slices include: horizontal slices with constructed abrupt change points and horizontal slices with continuous reflection points.

[0041] Dip gathers are pre-stack gathers generated during migration imaging by partially superimposing all participating data according to the stratigraphic dip. In the 3D dip domain, the reflection phase axis is a convex surface, with its stable phase point focused near the dip angle of the reflection interface. In horizontal slices, the Fresnel bands of continuous reflection waves are concentric ellipses, while the Fresnel bands at tectonic abrupt changes exhibit spatial phase abruptness, especially at faults, often displaying typical spatial distribution characteristics. Furthermore, the direction of this spatial phase abrupt change directly points to the fault strike. The first deep learning convolutional neural network identifies the classification label of a horizontal slice by recognizing the morphology of the Fresnel bands within the slice, that is, determining whether the imaging point to which the horizontal slice belongs is a "tectonic abrupt change point" or a "continuous reflection point".

[0042] Step S108: When the classification label is a horizontal slice of continuous reflection point, the second deep learning convolutional neural network is used to identify the stable phase point position of the horizontal slice of 3D tilt domain offset gather at the target imaging point, so as to use the coordinates of the stable phase point as the local tilt angle of the continuous reflection interface at the target imaging point.

[0043] Step S110: When the classification label is a horizontal slice of a mutation point, the third deep learning convolutional neural network is used to identify the fracture orientation of the 3D tilt domain offset gather horizontal slice at the target imaging point, so as to determine the spatial distribution azimuth of the fracture at the target imaging point based on the fracture orientation identification result.

[0044] For horizontal slices with two different classification labels, this embodiment of the invention trains two deep learning convolutional networks to further describe micro-structural features. Specifically, for horizontal slices with continuous reflection points, deep learning is used to extract local dip angles, and for horizontal slices with structural abrupt changes, fault strike analysis is performed to improve the accuracy of seismic exploration in describing local underground structures.

[0045] When the target imaging point is determined to be a continuous reflection point, this embodiment of the invention uses a pre-trained second deep learning convolutional network to locate the stable phase point in a horizontal slice at that imaging point. The Fresnel bands in this horizontal slice are concentric ellipses, and the stable phase point location is essentially the location of the center of the ellipse in the image. Stable phase point coordinates. The tilt angle values ​​(pairs) represent the local tilt angle of the continuous reflective interface at the imaging point.

[0046] When a target imaging point is identified as a structural abrupt change point, this embodiment of the invention uses a pre-trained third deep learning convolutional network to identify the fracture orientation in a horizontal slice at that imaging point. In this horizontal slice, the Fresnel bands exhibit a spatial phase abrupt change, and the direction of this spatial phase abrupt change points towards the fracture direction. Therefore, the spatial distribution azimuth of the fracture at the target imaging point can be determined based on the fracture orientation identification result.

[0047] Step S112: Construct a quantitative spatial description of the geological structure of the area to be imaged based on the identification results of all horizontal slices.

[0048] By integrating all the classification labels, local dip angles, and spatial distribution azimuth angles of fractures obtained from the above steps, a quantitative spatial description of the geological structure of the area to be imaged can be constructed.

[0049] The deep learning-based spatial description method for geological structures provided in this invention first calculates 3D dip-domain migration gather horizontal slices at each imaging point based on seismic observation data of the area to be imaged. Then, a first deep learning convolutional neural network is used to automatically identify the horizontal slices at the imaging points to distinguish between tectonic abrupt changes and continuous reflection points, achieving refined discrimination of geological structural features. Furthermore, dedicated second and third deep learning convolutional neural networks are employed for different types of imaging points to accurately identify the local dip angles of continuous reflection interfaces and the spatial distribution azimuth angles of fractures at the image point dimension. This enhances the spatial characterization accuracy of key geological bodies such as faults and fractures, thus effectively improving the resolution of the spatial description of geological structures.

[0050] In one alternative implementation, the first deep learning convolutional neural network, the second deep learning convolutional neural network, and the third deep learning convolutional neural network are all trained under supervised supervision.

[0051] The training samples for the first deep learning convolutional neural network include: 3D tilt domain offset gather horizontal slices with classification labels of constructing mutation point horizontal slices and 3D tilt domain offset gather horizontal slices with classification labels of continuous reflection point horizontal slices.

[0052] The training samples for the second deep learning convolutional neural network include: horizontal slices of 3D tilt domain offset gathers with stable phase point location annotations.

[0053] The training samples for the third deep learning convolutional neural network include: horizontal slices of 3D tilt domain offset gathers labeled with phase change lines.

[0054] Optionally, the training samples for the first deep learning convolutional neural network are horizontal slices of 3D tilt domain offset gathers corresponding to imaging points in the region of dramatic change and the region of continuous reflective surface.

[0055] As described above, the embodiments of the present invention use three specialized convolutional neural networks, which are used to identify structural abrupt change points / continuous reflection points, locate stable phase points, and detect phase abrupt change lines, respectively, to achieve accurate extraction of key geological structural features in the horizontal slices of 3D dip-domain offset gathers.

[0056] The first deep learning convolutional neural network is essentially a binary classification convolutional neural network, designed to automatically distinguish whether an imaging point is located in a continuous reflection structure or a fracture structure (such as a fault, fracture zone, stratigraphic pinch-out) based on a 3D dip-domain offset gather horizontal slice. During network training, the input data is a 3D dip-domain offset gather horizontal slice, with pixel-level classification labels used for training. The network output is the predicted classification label. To ensure the model's classification accuracy, the classification labels are manually labeled by professional technicians. Furthermore, samples with classification labels for horizontal slices representing structural abrupt changes are selected from slices in areas of dramatic geological change, such as near known faults, fold cores, and stratigraphic flexure zones. Samples with classification labels for horizontal slices representing continuous reflection points are selected from slices in layered regions (i.e., continuous reflection surface regions) far from faults and with stable sedimentation.

[0057] In one alternative implementation, the first deep learning convolutional neural network includes a lightweight convolutional neural network (CNN), which includes a feature extraction module and a classification decision module.

[0058] The feature extraction module is used to extract features from the horizontal slices of the 3D tilt domain offset gather to obtain high-dimensional feature vectors.

[0059] The classification decision module is used to map high-dimensional feature vectors to a binary classification category space and use the highest probability category as the classification label for the horizontal slice of the 3D tilt domain offset gather.

[0060] Figure 3 This diagram illustrates a lightweight convolutional neural network (CNN) according to an embodiment of the present invention. A deep learning CNN is a deep learning framework used for image classification and pattern recognition tasks. It extracts hierarchical features from input data through multiple convolutional layers and nonlinear activation functions. In the task of classifying horizontal slices for two different scenarios—reflection points (i.e., continuous reflection points mentioned above) and diffraction points (i.e., structural abrupt change points mentioned above)—the number of classification categories required is small, and the feature differences are significant. Therefore, only three convolutional blocks are needed, resulting in a smaller number of parameters, making it a custom lightweight CNN. It mainly consists of a feature extraction module and a classification decision module. The feature extraction module is composed of multiple convolutional layers, batch normalization layers, activation functions, and pooling layers stacked together. It progressively extracts spatial local features from seismic data through multi-scale convolutional kernels and uses max pooling to downsample the spatial dimension. The classification decision module consists of fully connected layers (FC) and dropout regularization layers, mapping the final high-dimensional feature vector to the category space to complete the classification prediction.

[0061] The second deep learning convolutional neural network aims to identify stable phase points in horizontal slices. A stable phase point is the location of an extreme path where the seismic wave path satisfies Fermat's principle; in migration imaging, it represents the optimal focusing location of the true subsurface reflection point. The network's training samples are horizontal slices of continuous reflection points selected from the training samples of the first deep learning convolutional neural network. The network input data is: horizontal slices of 3D dip-domain migration gathers; the training label is: manually labeled stable phase point locations (i.e., ellipse center locations); and the network output data is: predicted stable phase point locations. Figure 4 This is a schematic diagram illustrating training samples and prediction results of a second deep learning convolutional neural network provided in an embodiment of the present invention. Figure 4 The horizontal and vertical axes of the three views represent the corresponding coordinate values ​​in the binary tilt angle gather coordinate system, that is, the pseudo-tilt angle coordinate values ​​in the x and y directions; the label data refers to the target center point position marked in the training set, which is stored in coordinate form. Figure 4 The left-hand view shows the original horizontal slice shape, which shows that it has a smooth and complete elliptical phase. The middle view shows the result of manually marking the stable phase point position when it was used as a training sample. The right-hand view shows the stable phase point position predicted after network recognition. The tilt angle value corresponding to the position is the local tilt angle of the reflection interface at that image point (represented by the x and y components).

[0062] In one optional implementation, the second deep learning convolutional neural network includes: an improved U-Net network, which includes: a first encoder and a first decoder; the first encoder consists of multiple convolutional blocks with LeakyReLU activation functions and convolutional downsampling modules with a stride of 2; the first decoder is skip-connected to the feature maps of the corresponding layers of the first encoder.

[0063] The first encoder is used to extract multi-scale spatial features of the horizontal slices of the 3D tilt domain offset gather layer by layer.

[0064] The first decoder is used to reconstruct the spatial details of the horizontal slice of the 3D tilt domain offset gather layer by layer based on multi-scale spatial features, generate a pixel-level probabilistic heat map, and use the extreme point positions in the pixel-level probabilistic heat map as the stable phase point positions.

[0065] Figure 5This is a schematic diagram of the network structure of an improved U-Net network provided in an embodiment of the present invention. U-Net is a deep learning framework applied to image semantic segmentation tasks, achieving pixel-level accurate classification through an encoder-decoder structure. In the task of automatically identifying the location of stable phase points in horizontal slices of three-dimensional dip gathers, this embodiment of the present invention employs an improved U-Net framework, mainly composed of a first encoder and a first decoder. The first encoder consists of multiple convolutional blocks and downsampling modules stacked together, used to extract multi-scale spatial features of seismic data layer by layer. The first decoder recovers spatial details through an upsampling module and performs skip connections with the feature maps of corresponding levels during the encoding process, effectively fusing local texture and global structural information. Furthermore, to prevent the loss of negative features, all convolutional blocks in the network use LeakyReLU as the activation function; to avoid the problem of local information loss caused by direct pooling, a convolutional layer with a stride of 2 is used to process the feature map during downsampling.

[0066] The third deep learning convolutional neural network aims to identify phase abrupt change lines in horizontal slices. The network's training samples are horizontal slices constructed from the training samples of the first deep learning convolutional neural network, representing abrupt change points. The network input data consists of horizontal slices of 3D tilt-domain offset gathers, with training labels consisting of manually labeled phase abrupt change lines. The network output data is a probability heatmap indicating phase abrupt change regions. Figure 6 This is a schematic diagram illustrating training samples and prediction results of a third deep learning convolutional neural network provided in an embodiment of the present invention. Figure 6 The left-hand view shows the original horizontal slice shape, which shows obvious phase abrupt changes or inconsistencies in the elliptical phase shape. The middle view shows the result of manually marking phase abrupt change lines when used as training samples. The right-hand view shows the probability heat map of the phase abrupt change region predicted after network recognition. The spatial angle corresponding to its shape represents the spatial direction of the break at that image point.

[0067] In one alternative implementation, the third deep learning convolutional neural network includes a Swin-Unet network, which includes a second encoder, a bottleneck layer, and a second decoder; the feature maps of the corresponding layers of the second decoder and the second encoder are skip-connected.

[0068] The second encoder is used to extract deep semantic features of the horizontal slices of the 3D tilt domain offset gather layer by layer.

[0069] The bottleneck layer is used to abstract deep semantic features into a compact high-dimensional semantic representation, thus obtaining high-dimensional semantic features.

[0070] The second decoder is used to generate a probability hotspot map indicating phase transition regions based on high-dimensional semantic features and high-resolution feature maps corresponding to the encoding path levels.

[0071] Figure 7 This is a schematic diagram of the network structure of an improved Swin-Unet network provided in an embodiment of the present invention. The Swin-Unet network contains several key modules, including Swin Transformer (shifting window-transformer), Patch Partition, Patch Merging, Patch Expanding, Linear Embedding, and Linear Projection. Its core advantage lies in the organic integration of the powerful modeling capabilities of visual Transformers with the multi-scale skip connection mechanism of U-Net, thereby achieving efficient modeling of long-range dependencies while preserving spatial details. Compared with traditional CNN-like networks, Swin-Unet does not rely on convolutional kernels with fixed receptive fields, but instead dynamically models the global correlation between pixels through a self-attention mechanism, making it suitable for recognizing complex and non-stationary geological structure features in tilt channels. Among them, the Swin Transformer, as the core driving engine of Swin-Unet, is the key to the performance leap of this architecture. It is a visual Transformer architecture based on a windowed self-attention mechanism, which exhibits superior performance to traditional convolutional networks in image classification, detection, and segmentation tasks.

[0072] For the task of identifying azimuth indicator features in horizontal slices of inclined gathers, the Swin-Unet network used in this embodiment of the invention mainly consists of three parts: a second encoder, a bottleneck layer, and a second decoder. The second encoder is composed of stacked multi-level Swin Transformer blocks, each containing a multi-head self-attention mechanism based on local windows and shift window attention, capturing cross-regional spatial context information while reducing computational complexity. Then, the Patch Merging module progressively downsamples, achieving spatial compression and channel expansion of the feature map, extracting deep semantic features layer by layer. The bottleneck layer, located at the end of the encoding path, is composed of the last-level Swin Transformer block and is responsible for further abstracting the multi-scale downsampled features into a compact high-dimensional semantic representation. This bottleneck layer contains a multi-head window self-attention mechanism, capable of modeling global context information with limited computational overhead. The second decoder upsamples through the Patch Expanding module and fuses high-resolution feature maps from corresponding levels of the encoding path to form a skip connection structure, effectively restoring spatial details.

[0073] In an optional embodiment, the fracture orientation identification result includes: a probability heatmap indicating the phase abrupt change region; in step S110 above, determining the spatial distribution azimuth angle of the fracture at the target imaging point based on the fracture orientation identification result specifically includes the following steps:

[0074] Step S1101: Binarize the probability heatmap to obtain the spatial location of the phase abrupt change.

[0075] Step S1102: Use the least squares fitting method to perform linear fitting on the spatial location of the phase change to obtain the phase change line and the spatial orientation it represents.

[0076] Step S1103: The spatial orientation represented by the phase change line is taken as the spatial distribution azimuth angle of the break at the target imaging point.

[0077] Using the above steps to Figure 6 After processing the probability heatmap indicating the phase abrupt change region, the phase abrupt change line in the horizontal slice and the spatial orientation it represents can be obtained. Figure 8 In order to be in Figure 6 A schematic diagram showing the phase transition lines and their corresponding spatial distribution azimuth angles marked on the original horizontal slice.

[0078] The embodiments of the present invention utilize second and third deep learning convolutional neural networks to extract the local tilt angle or spatial orientation of the fracture at the continuous reflective interface, with a resolution reaching the pixel dimension, which has higher accuracy compared with traditional reflected wave imaging methods.

[0079] In summary, this invention introduces deep learning methods based on horizontal slicing of three-dimensional dip gathers, enabling the determination of local structural properties and the extraction of spatial distribution morphological parameters at imaging points. Processors only need to collect the structural morphology and spatial distribution features corresponding to imaging points at typical structural locations, construct samples, and train the network. Then, using deep learning neural networks, they can achieve detailed descriptions of local structural properties and spatial distribution at any imaging point, significantly improving the accuracy of describing small-scale subsurface structures while avoiding a massive workload. This method extracts high-resolution small-scale subsurface structural distribution information from surface seismic observation data, which has significant application value for oil and gas reservoir exploration and development.

[0080] Example 2

[0081] This invention also provides a deep learning-based geological structure spatial description device, which is mainly used to execute the deep learning-based geological structure spatial description method provided in Embodiment 1 above. The following is a detailed description of the deep learning-based geological structure spatial description device provided in this invention.

[0082] Figure 9 A functional block diagram of a deep learning-based spatial description device for geological structures is provided in an embodiment of the present invention, as shown below. Figure 9As shown, the device mainly includes: an acquisition module 11, an imaging module 12, a first recognition module 13, a second recognition module 14, a third recognition module 15, and a construction module 16, wherein:

[0083] The acquisition module 11 is used to acquire seismic observation data of the area to be imaged.

[0084] The imaging module 12 is used to perform migration imaging of the imaging space of the area to be imaged based on seismic observation data, obtain the 3D dip domain migration gather of the imaging space, and extract the horizontal slice of the 3D dip domain migration gather at each imaging point in the imaging space.

[0085] The first identification module 13 is used to perform type identification on the 3D tilt domain offset gather horizontal slice at the target imaging point using a first deep learning convolutional neural network, and obtain the classification label of the horizontal slice; the target imaging point represents any imaging point in the imaging space; the classification label of the horizontal slice includes: horizontal slice with abrupt change point and horizontal slice with continuous reflection point.

[0086] The second identification module 14 is used to identify the stable phase point position of the 3D tilt domain offset gather horizontal slice at the target imaging point by using a second deep learning convolutional neural network when the classification label is a continuous reflection point horizontal slice, so as to use the coordinates of the stable phase point as the local tilt angle of the continuous reflection interface at the target imaging point.

[0087] The third identification module 15 is used to identify the fracture orientation of the 3D tilt domain offset gather horizontal slice at the target imaging point by using a third deep learning convolutional neural network when the classification label is a horizontal slice of the construction mutation point, so as to determine the spatial distribution azimuth of the fracture at the target imaging point based on the fracture orientation identification result.

[0088] Module 16 is used to construct a quantitative spatial description of the geological structure of the area to be imaged based on the identification results of all horizontal slices.

[0089] The geological structure spatial description device based on deep learning provided in this invention first calculates 3D dip-domain migration gather horizontal slices at each imaging point based on seismic observation data of the area to be imaged. Then, a first deep learning convolutional neural network is used to automatically identify the horizontal slices at the imaging points to distinguish between tectonic abrupt changes and continuous reflection points, achieving refined discrimination of geological structural features. Furthermore, dedicated second and third deep learning convolutional neural networks are used for different types of imaging points to accurately identify the local dip angles of continuous reflection interfaces and the spatial distribution azimuth angles of fractures at the image point dimension. This enhances the spatial characterization accuracy of key geological bodies such as faults and fractures, thus effectively improving the resolution of the geological structure spatial description.

[0090] Optionally, the first deep learning convolutional neural network includes a lightweight convolutional neural network (CNN), which includes a feature extraction module and a classification decision module.

[0091] The feature extraction module is used to extract features from the horizontal slices of the 3D tilt domain offset gather to obtain high-dimensional feature vectors.

[0092] The classification decision module is used to map high-dimensional feature vectors to a binary classification category space and use the highest probability category as the classification label for the horizontal slice of the 3D tilt domain offset gather.

[0093] Optionally, the second deep learning convolutional neural network includes: an improved U-Net network, which includes: a first encoder and a first decoder; the first encoder consists of multiple convolutional blocks with LeakyReLU activation functions and convolutional downsampling modules with a stride of 2; the first decoder is skip-connected to the feature maps of the corresponding layers of the first encoder.

[0094] The first encoder is used to extract multi-scale spatial features of the horizontal slices of the 3D tilt domain offset gather layer by layer.

[0095] The first decoder is used to reconstruct the spatial details of the horizontal slice of the 3D tilt domain offset gather layer by layer based on multi-scale spatial features, generate a pixel-level probabilistic heat map, and use the extreme point positions in the pixel-level probabilistic heat map as the stable phase point positions.

[0096] Optionally, the third deep learning convolutional neural network includes: a Swin-Unet network, which includes: a second encoder, a bottleneck layer, and a second decoder; the feature maps of the corresponding layers of the second decoder and the second encoder are skip-connected.

[0097] The second encoder is used to extract deep semantic features of the horizontal slices of the 3D tilt domain offset gather layer by layer.

[0098] The bottleneck layer is used to abstract deep semantic features into a compact high-dimensional semantic representation, thus obtaining high-dimensional semantic features.

[0099] The second decoder is used to generate a probability hotspot map indicating phase transition regions based on high-dimensional semantic features and high-resolution feature maps corresponding to the encoding path levels.

[0100] Optionally, the fracture orientation identification result includes: a probability heatmap indicating the phase abrupt change region; the third identification module 15 is also used for:

[0101] Binarize the probability heatmap to obtain the spatial location of phase abrupt changes.

[0102] The least squares fitting method is used to fit a straight line to the spatial location of the phase change, and the phase change line and the spatial orientation it represents are obtained.

[0103] The spatial orientation represented by the phase change line is taken as the spatial distribution azimuth angle of the break at the target imaging point.

[0104] Optionally, the first deep learning convolutional neural network, the second deep learning convolutional neural network, and the third deep learning convolutional neural network are all trained under supervised supervision.

[0105] The training samples for the first deep learning convolutional neural network include: 3D tilt domain offset gather horizontal slices with classification labels of constructing mutation point horizontal slices and 3D tilt domain offset gather horizontal slices with classification labels of continuous reflection point horizontal slices.

[0106] The training samples for the second deep learning convolutional neural network include: horizontal slices of 3D tilt domain offset gathers with stable phase point location annotations.

[0107] The training samples for the third deep learning convolutional neural network include: horizontal slices of 3D tilt domain offset gathers labeled with phase change lines.

[0108] Optionally, the training samples for the first deep learning convolutional neural network are horizontal slices of 3D tilt domain offset gathers corresponding to imaging points in the region of dramatic change and the region of continuous reflective surface.

[0109] Example 3

[0110] See Figure 10 This invention provides an electronic device, which includes a processor 60, a memory 61, a bus 62, and a communication interface 63. The processor 60, the communication interface 63, and the memory 61 are connected via the bus 62. The processor 60 is used to execute executable modules, such as computer programs, stored in the memory 61.

[0111] The memory 61 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 63 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0112] Bus 62 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 10The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0113] The memory 61 is used to store programs. After receiving an execution instruction, the processor 60 executes the program. The method executed by the apparatus defined by the process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor 60 or implemented by the processor 60.

[0114] Processor 60 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 60 or by instructions in software form. Processor 60 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 61. Processor 60 reads the information in memory 61 and, in conjunction with its hardware, completes the steps of the above method.

[0115] The computer program product of the deep learning-based method and apparatus for spatial description of geological structures provided in this embodiment of the invention includes a computer-readable storage medium storing non-volatile program code executable by a processor. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0116] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0117] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0118] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0119] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use. They are only for the convenience of describing this 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, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0120] Furthermore, terms such as "horizontal," "vertical," and "sag" do not imply that components must be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal relative to "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0121] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A spatial description method for geological structures based on deep learning, characterized in that, include: Acquire seismic observation data of the area to be imaged; Based on the seismic observation data, the imaging space of the area to be imaged is migrated to obtain a 3D dip domain migration gather of the imaging space, and a horizontal slice of the 3D dip domain migration gather at each imaging point in the imaging space is extracted from it. The first deep learning convolutional neural network is used to perform type recognition on the horizontal slices of the 3D tilt domain offset gather at the target imaging point to obtain the classification label of the horizontal slices; The target imaging point represents any imaging point in the imaging space; The classification labels for the horizontal slices include: horizontal slices with constructed mutation points and horizontal slices with continuous reflection points; When the classification label is a horizontal slice of continuous reflection point, the second deep learning convolutional neural network is used to identify the stable phase point position of the horizontal slice of 3D tilt domain offset gather at the target imaging point, so as to use the coordinates of the stable phase point as the local tilt angle of the continuous reflection interface at the target imaging point. When the classification label is a horizontal slice of a mutation point, the fracture orientation is identified by the 3D tilt domain offset gather horizontal slice at the target imaging point using a third deep learning convolutional neural network, so as to determine the spatial distribution azimuth of the fracture at the target imaging point based on the fracture orientation identification result. Based on the identification results of all horizontal slices, a quantitative spatial description of the geological structure of the area to be imaged is constructed. The fracture orientation identification result includes: a probability heatmap indicating the phase abrupt change region; and determining the spatial distribution azimuth angle of the fracture at the target imaging point based on the fracture orientation identification result, including: The probability heatmap is binarized to obtain the spatial location of the phase abrupt change; The spatial location of the phase change was fitted with a straight line using the least squares fitting method to obtain the phase change line and the spatial orientation it represents. The spatial orientation represented by the phase change line is taken as the spatial distribution azimuth angle of the fracture at the target imaging point.

2. The deep learning-based spatial description method for geological structures according to claim 1, characterized in that, The first deep learning convolutional neural network includes: a lightweight convolutional neural network CNN, wherein the lightweight convolutional neural network CNN includes: a feature extraction module and a classification decision module; The feature extraction module is used to extract features of the horizontal slices of the 3D tilt domain offset gather to obtain a high-dimensional feature vector. The classification decision module is used to map the high-dimensional feature vector to a binary classification category space, and use the highest probability category as the classification label of the horizontal slice of the 3D tilt domain offset gather.

3. The deep learning-based spatial description method for geological structures according to claim 1, characterized in that, The second deep learning convolutional neural network includes an improved U-Net network, which includes a first encoder and a first decoder; the first encoder consists of multiple convolutional blocks with LeakyReLU activation functions and convolutional downsampling modules with a stride of 2; the first decoder makes skip connections to the feature maps of the corresponding layers of the first encoder. The first encoder is used to extract multi-scale spatial features of horizontal slices of 3D tilt domain offset gather layer by layer; The first decoder is used to reconstruct the spatial details of the horizontal slice of the 3D tilt domain offset gather layer by layer based on the multi-scale spatial features, generate a pixel-level probability heat map, and use the extreme point position in the pixel-level probability heat map as the stable phase point position.

4. The deep learning-based spatial description method for geological structures according to claim 1, characterized in that, The third deep learning convolutional neural network includes a Swin-Unet network, which includes a second encoder, a bottleneck layer, and a second decoder; the feature maps of the corresponding layers of the second decoder and the second encoder are skipped connections. The second encoder is used to extract deep semantic features of horizontal slices of 3D tilt domain offset gathers layer by layer; The bottleneck layer is used to abstract the deep semantic features into a compact high-dimensional semantic representation, thereby obtaining high-dimensional semantic features; The second decoder is used to generate a probability hotspot map indicating phase transition regions based on the high-dimensional semantic features and the high-resolution feature map corresponding to the encoding path level.

5. The deep learning-based spatial description method for geological structures according to claim 1, characterized in that, The first deep learning convolutional neural network, the second deep learning convolutional neural network, and the third deep learning convolutional neural network are all trained under supervised supervision; The training samples of the first deep learning convolutional neural network include: 3D tilt domain offset gather horizontal slices with classification labels of constructing mutation point horizontal slices and 3D tilt domain offset gather horizontal slices with classification labels of continuous reflection point horizontal slices. The training samples for the second deep learning convolutional neural network include: horizontal slices of 3D tilt domain offset gathers with stable phase point location annotations; The training samples for the third deep learning convolutional neural network include: horizontal slices of 3D tilt domain offset gathers labeled with phase change lines.

6. The deep learning-based spatial description method for geological structures according to claim 5, characterized in that, The training samples of the first deep learning convolutional neural network are horizontal slices of 3D tilt domain offset gathers corresponding to imaging points in the region of drastic change and the region of continuous reflective surface.

7. A spatial description device for geological structures based on deep learning, characterized in that, include: The acquisition module is used to acquire seismic observation data of the area to be imaged; The imaging module is used to perform migration imaging of the imaging space of the area to be imaged based on the seismic observation data, obtain the 3D dip domain migration gather of the imaging space, and extract the horizontal slice of the 3D dip domain migration gather at each imaging point in the imaging space. The first identification module is used to perform type identification on the horizontal slice of the 3D tilt domain offset gather at the target imaging point using the first deep learning convolutional neural network, and obtain the classification label of the horizontal slice. The target imaging point represents any imaging point in the imaging space; The classification labels for the horizontal slices include: horizontal slices with constructed mutation points and horizontal slices with continuous reflection points; The second identification module is used to identify the stable phase point position of the 3D tilt domain offset gather horizontal slice at the target imaging point by using the second deep learning convolutional neural network when the classification label is a continuous reflection point horizontal slice, so as to use the coordinates of the stable phase point as the local tilt angle of the continuous reflection interface at the target imaging point. The third identification module is used to identify the fracture orientation of the 3D tilt domain offset gather horizontal slice at the target imaging point by using a third deep learning convolutional neural network when the classification label is a horizontal slice of the construction mutation point. Based on the fracture orientation identification result, the spatial distribution azimuth of the fracture at the target imaging point is determined. A construction module is used to construct a quantitative spatial description of the geological structure of the area to be imaged based on the identification results of all horizontal slices; The fracture orientation identification results include: a probability heatmap indicating the phase abrupt change region; the third identification module is also used for: The probability heatmap is binarized to obtain the spatial location of the phase abrupt change; The spatial location of the phase change was fitted with a straight line using the least squares fitting method to obtain the phase change line and the spatial orientation it represents. The spatial orientation represented by the phase change line is taken as the spatial distribution azimuth angle of the fracture at the target imaging point.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the deep learning-based spatial description method for geological structures as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the deep learning-based spatial description method for geological structures as described in any one of claims 1 to 6.

Images