Shale fault crack identification method and device based on deep learning

Abstract

The invention provides a shale fault crack identification method and device based on deep learning, and relates to the technical field of shale development engineering, and the method comprises the steps: obtaining the seismic data of a shale fault region; performing feature extraction and seismic inversion on the seismic data to obtain seismic features and prior features; constructing a graph structure based on the seismic features and the prior features to obtain a hierarchical graph structure; performing Monte Carlo perturbation on the hierarchical graph structure to obtain a perturbation sample graph; training the graph neural network by adopting the disturbance sample graph, and optimizing network parameters to obtain a crack recognition network; and acquiring actual seismic data of the target shale fault region, constructing an actual graph structure based on the actual seismic data, and performing crack identification on the actual graph structure through the crack identification network to obtain a crack identification result. According to the method, the problems of insufficient samples, rough graph construction mode and difficulty in ensuring the reliability and geometric consistency of the result in the current research on crack identification of the shale fault area are solved.

Description

Technical Field

[0001] This invention relates to the field of shale development engineering technology, and more specifically, to a method and apparatus for identifying shale faults and fractures based on deep learning. Background Technology

[0002] Shale is a sedimentary rock with a distinct bedding structure, and the development of internal fractures has a significant impact on oil and gas occurrence, shale gas development, and geological stability assessment. Especially near fault zones, due to tectonic stress concentration and lithological differences, fracture development patterns are more complex, exhibiting a heterogeneous distribution characteristic across multiple directions and scales. Accurately identifying the spatial distribution of fractures near faults is crucial for reservoir prediction and engineering safety.

[0003] Traditional fracture identification methods mainly rely on geological drilling, core analysis, or manual interpretation of seismic images, but these methods suffer from high costs, low resolution, or strong subjectivity. In recent years, deep learning has been increasingly applied to seismic data processing and subsurface structure identification tasks. Convolutional neural networks can extract edge and texture features from seismic images to some extent, but they have limitations in modeling spatial relationships and topological structures, making it difficult to meet the spatial topological modeling requirements in fault fracture identification.

[0004] To enhance spatial structure understanding, graph neural networks have been introduced to more effectively model spatial relationships with non-Euclidean structures in seismic data. Local regions within seismic traces or slices can be constructed as graph structures, utilizing the geometric relationships and physical properties between nodes to propagate spatial information, thereby improving the identification of complex fracture structures. However, current research on fracture identification in fault zones still faces challenges such as insufficient sample size, limited node feature representation, coarse graph construction methods, and the lack of explicit modeling of fracture connectivity and directionality, making it difficult to guarantee the reliability and geometric consistency of the results. Therefore, a novel fracture identification method that can integrate multi-scale spatial information, directional features, and reliability assessment mechanisms within a graph structure is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for identifying shale fault fractures based on deep learning, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a deep learning-based method for identifying shale fault fractures, including: Acquire seismic data in shale fault zones; Seismic data is subjected to feature extraction and seismic inversion to obtain seismic features and prior features. The seismic features include correlation, curvature value, instantaneous amplitude, instantaneous frequency, texture attributes and multi-scale features. The prior features include geostress-sensitive parameters and corresponding reliability indices. A graph structure is constructed based on seismic features and prior features, and a hierarchical graph structure is obtained through voxel mapping and high-level clustering. Monte Carlo perturbation is applied to the hierarchical graph structure, and the perturbation intensity is controlled by reliability indicators to obtain the perturbation sample graph; A graph neural network was trained using perturbed sample images, and the crack identification network was obtained by optimizing the network parameters. The actual seismic data of the target shale fault area is obtained, the actual map structure is constructed based on the actual seismic data, and the crack identification results are obtained by using a crack identification network to identify cracks in the actual map structure.

[0006] Secondly, this application also provides a deep learning-based shale fault fracture identification device, comprising: The acquisition module is used to acquire seismic data in shale fault zones; The extraction module is used to extract features and perform seismic inversion on seismic data to obtain seismic features and prior features. The seismic features include correlation, curvature value, instantaneous amplitude, instantaneous frequency, texture attributes and multi-scale features. The prior features include geostress-sensitive parameters and corresponding reliability indices. The construction module is used to build a graph structure based on seismic features and prior features, and obtain a hierarchical graph structure through voxel mapping and high-level clustering. The perturbation module is used to perform Monte Carlo perturbation on the hierarchical graph structure. The perturbation intensity is controlled by the reliability index to obtain the perturbation sample graph. The training module is used to train the graph neural network using perturbed sample maps, and obtain the crack recognition network by optimizing the network parameters; The identification module is used to acquire actual seismic data of the target shale fault area, construct an actual map structure based on the actual seismic data, and identify cracks in the actual map structure through a crack identification network to obtain crack identification results.

[0007] The beneficial effects of this invention are as follows: This invention integrates seismic characteristics and prior geological information to construct a hierarchical graph structure, and enhances the network's robustness to uncertain data through a perturbation mechanism, thereby improving the accuracy and stability of crack identification. In the identification stage, crack nodes are screened using reliability indicators, and crack segments are effectively eliminated through principal direction analysis and spatial aggregation and merging, effectively removing noise and misidentification results. Ultimately, a clear and continuous crack identification result is obtained, suitable for the practical engineering needs of complex structural areas.

[0008] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the deep learning-based shale fault fracture identification method described in this embodiment of the invention; Figure 2 This is a schematic diagram of the graph neural network structure described in an embodiment of the present invention. Detailed Implementation

[0011] 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, not all, of the embodiments of the present invention. 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. Therefore, the following detailed description of the embodiments of the present 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 present invention without inventive effort are within the scope of protection of the present invention.

[0012] It should be noted that similar reference numerals 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. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0013] Example 1: This embodiment provides a deep learning-based method for identifying shale fault fractures.

[0014] See Figure 1 The figure shows that the method includes steps S1, S2, S3, S4, S5, and S6.

[0015] Step S1: Obtain seismic data for the shale fault region; Understandably, acquiring seismic data requires seismic exploration. Specifically, this involves deploying seismic sources and geophone arrays in shale fault areas. Seismic waves are generated by exciting artificial sources (such as vibrating platforms, explosive sources, water hammers, etc.), and the geophones receive the reflected waves to obtain raw seismic data. The raw seismic data then undergoes preprocessing (denoising, DC removal, amplitude equalization, frequency filtering, static correction, etc.) and seismic imaging (stacking, migration) operations to obtain a three-dimensional seismic data volume used for fault fracture identification, i.e., seismic data.

[0016] The earthquake dataset used for training can also be obtained from publicly available seismic exploration databases, including 3D seismic reflection data, velocity models, or fracture label information.

[0017] Step S2: Perform feature extraction and seismic inversion on the seismic data to obtain seismic features and prior features. The seismic features include correlation, curvature value, instantaneous amplitude, instantaneous frequency, texture attributes and multi-scale features. The prior features include geostress-sensitive parameters and corresponding reliability indices. In this embodiment, seismic features directly reflect the changing behavior of seismic waves in underground structures, which helps to identify anomalies such as faults, cracks, and stratigraphic interfaces. Prior features are inferred parameters constructed based on the geophysical background and are used to assist in identification and reliability judgment. Especially in the identification of shale faults and cracks, rock mechanics and geostress often control the development location and morphology of cracks.

[0018] Therefore, seismic features are used to capture local differences in the propagation of seismic waves in different underground media, serving as the core basis for identifying tectonic boundaries and anomalies. Prior features are used to incorporate geological background constraints such as rock mechanics and geostress distribution. Combined with their reliability indices, the identification results can be screened and corrected, thereby improving the accuracy, interpretability, and practical engineering application value of crack identification.

[0019] In step S2, the steps for obtaining the prior features are as follows: Step A1: Perform seismic wave impedance inversion based on seismic data to obtain inversion results, which include P-wave impedance, S-wave impedance, and density; In this embodiment, the seismic data is converted from time to depth using an initial velocity model to obtain a depth-domain seismic profile (voxel data), ensuring that the inversion is based on the geological depth coordinate system. The initial velocity model can be obtained by assigning different velocity values ​​to different strata based on geological data and stratigraphic division, forming a layered velocity model, or by directly calculating the velocity within the well trajectory range using well logging curves obtained from drilling (such as sonic transit time, density logging, etc.). Combined with stratigraphic information, a velocity profile around the well is established, and then extended to the entire study area through spatial interpolation.

[0020] Obtain well logging data at the well location, and extract the initial model of the P-wave impedance of the target area as the initial value for inversion using the well logging data or empirical curves.

[0021] The reflection coefficients of each reflection interface (depth domain seismic profile) are calculated using seismic data and an initial velocity model to prepare the input parameters for inversion.

[0022] By employing appropriate inversion methods, such as AVO inversion based on a Bayesian framework or elastic wave impedance inversion, the initial P-wave impedance model is updated iteratively through the reflection coefficient. By fitting seismic data, the P-wave impedance, S-wave impedance, and density of each voxel are obtained through inversion.

[0023] Step A2: Combining prior well location information and geological background, spatial correction and uncertainty assessment are performed on the inversion results to obtain the corrected inversion results and corresponding reliability indicators; In this embodiment, the prior information on the well location and the geological background include well logging curve data of the well location (such as sonic transit time, density logging, porosity, etc.), as well as geological background data such as stratigraphic division, structural features, and sedimentary environment.

[0024] The inversion results are compared with the measured logging data in the well for error analysis. After obtaining the error, the error distribution of each inversion result (P-wave impedance, S-wave impedance, and density) is statistically analyzed to establish an error probability model, which can be either a normal distribution or a piecewise distribution. Based on the error variation trend of voxels in different regions, Kriging interpolation is used to fit the error estimate field of all voxels. The inversion results of all voxels are then spatially corrected using the error estimate field to obtain the corrected inversion results.

[0025] Based on the corrected error estimates, a reliability index is constructed for each voxel. This index measures the confidence level of the inversion results at that location, with a value ranging from 0 to 1. A higher value indicates a better fit between the inversion results and the well logging data, and a smaller error. The reliability index is obtained using error normalization, confidence interval probability, or a weighted fusion of three attributes.

[0026] Specifically, the error value for each voxel is normalized, mapping the error to the [0,1] interval. The error is then inverted and mapped to a preliminary reliability value using the following formula: ; In the formula, This indicates a preliminary reliability value. This represents the normalization error.

[0027] A confidence interval is established for the error probability distribution based on the error probability model, and the confidence probability value of each voxel is calculated, reflecting the probability that the error is within a reasonable range. The preliminary reliability values ​​of the three inversion results are weighted and fused to obtain the comprehensive reliability. The confidence probability value of each voxel is used as a correction factor to adjust the comprehensive reliability, resulting in the reliability index for the corresponding voxel.

[0028] Step A3: Calculate the geostress-sensitive parameters based on the corrected inversion results. The geostress-sensitive parameters include Poisson's ratio, brittleness index, and P-wave / S-wave velocity ratio.

[0029] In this embodiment, for each voxel, the velocity is calculated from impedance and density based on physical relationships, using the following formula: ; In the formula, Indicates the longitudinal wave velocity. Indicates the transverse wave velocity. This represents the corrected longitudinal wave impedance. This represents the corrected transverse wave impedance. This indicates the calibrated density.

[0030] Furthermore, Poisson's ratio is calculated using the relationship between P-wave velocity and S-wave velocity: ; In the formula, Represents Poisson's ratio. Indicates the longitudinal wave velocity. This indicates the velocity of the transverse wave.

[0031] Based on Poisson's ratio reflecting the brittleness of rocks, the brittleness index is obtained, and its simplified expression is: ; In the formula, Indicates the brittleness index. It represents Poisson's ratio.

[0032] Poisson's ratio, brittleness index, and the ratio of P-wave to S-wave velocity are used as three-dimensional attribute data and assigned to corresponding voxels for subsequent geological interpretation and fault fracture identification.

[0033] In step S2, the steps for obtaining the seismic features are as follows: Step B1: Perform coherence calculation on the seismic data. The coherence is obtained by using the sliding window cross-correlation method. It is understood that this embodiment uses voxels as basic units and employs the sliding window cross-correlation method to calculate the waveform similarity between each voxel and its neighboring voxels, thereby obtaining a three-dimensional coherence attribute volume (including the coherence of each voxel) that reflects the continuous changes in local seismic waveforms.

[0034] Specifically, a fixed-size three-dimensional sliding window (e.g., 11×11×5 voxels) is set. For each target voxel, neighborhood search segments are defined in multiple principal directions (e.g., horizontal, vertical, and depth directions) centered on it. For each neighboring voxel in each direction, its seismic amplitude sequence (waveform within a certain time or depth range) is extracted; for the central voxel, a reference waveform sequence of the same length is extracted. The waveform sequences of the central voxel and each neighboring voxel are normalized, and their similarity is calculated using cross-correlation coefficients. For the coherence values ​​calculated in each direction, the weighted value is taken as the final coherence index for that voxel. The coherence indices of all voxels are integrated to generate a voxel-level three-dimensional coherence attribute volume, with a value range of [0,1]. The lower the value, the more discontinuous the seismic waveform. When cracks (e.g., tension cracks, shear cracks) appear, they can cause abrupt changes, discontinuities, or energy attenuation in the waveform. Therefore, the coherence value will be significantly reduced in the crack boundary region, and the multi-directional weighting mechanism helps to locate asymmetric cracks or oblique crack structures.

[0035] Step B2: Extract structural surfaces based on seismic data. By extracting the maximum amplitude and peak values ​​of specific attributes as interfaces, the seismic reflection interface is obtained. In this embodiment, a processing window is first set up to read the original seismic amplitude value of the vertical profile where the voxel is located, and other specific attributes (such as the Hilbert envelope) are calculated at the same time to enhance the sensitivity of interface recognition.

[0036] In the vertical neighborhood of the voxel, the depth position corresponding to the maximum amplitude point is found, and the voxel is recorded as the reflection point at that horizontal position. This process is repeated sequentially for all voxel coordinates to obtain a three-dimensional interface voxel set consisting of the maximum amplitude and the peak position of specific attributes, i.e., the seismic reflection interface. To eliminate discrete noise, three-dimensional interpolation smoothing can be performed on the extracted interface points.

[0037] Step B3: Perform derivative calculations on the seismic reflection interface to obtain the interface curvature; In this step, based on the voxel set of the seismic reflection interface, the local curvature value of the reflection interface is calculated through derivative operations to obtain the interface curvature.

[0038] Step B4: Interpolate based on the interface curvature corresponding to the seismic reflection interface to obtain a curvature volume, wherein the curvature volume includes the curvature value corresponding to each voxel in the seismic data; In this step, based on the overall spatial extent of the seismic data volume, a target interpolation grid (consistent with the original seismic volume voxel grid) is set to ensure that the interpolation results can be mapped to each voxel location to obtain the curvature volume.

[0039] Curvature volumes, as one of the spatial derivative enhancement properties, can effectively reveal local geometric anomalies at seismic reflection interfaces. In areas with dense faults and fractures, curvature values ​​typically exhibit high values ​​or abrupt changes. Therefore, curvature volumes can be used to assist in identifying structural anomaly areas such as fault boundaries and fracture zones.

[0040] Step B5: Perform Hilbert transform and feature extraction on the seismic data to obtain instantaneous amplitude and instantaneous frequency; Step B6: Extract texture statistics based on seismic data slices, and calculate texture attributes by constructing a gray-level co-occurrence matrix. The texture attributes include GLCM energy and GLCM contrast. Understandably, areas with high GLCM energy indicate relatively uniform texture, potentially representing dense or homogeneous strata. Areas with high GLCM contrast indicate dramatic texture variations, often corresponding to faults, fractures, or complex geological structures. These texture attributes can be used to assist in fracture identification and geological body segmentation.

[0041] In this step, a 2D seismic amplitude data slice is extracted using a specific direction (horizontal XY plane, vertical XZ plane, or YZ plane). The seismic amplitude values ​​within the slice are normalized to a fixed grayscale range (0–255). For the normalized 2D grayscale image, spatial relationship parameters, including distance and direction, are set, and the number of times different grayscale levels co-occur under the above spatial relationship is counted to construct a 2D matrix. Therefore, for the entire slice, multiple GLCM matrices are obtained in multiple directions. Texture attributes are calculated using the GLCM matrix, and the energy and contrast attributes calculated in different directions are weighted and fused to obtain the comprehensive texture attribute value of the slice. The texture attributes of each pixel in the slice are mapped back to the corresponding voxel. At the same time, the processing results of multiple slices are superimposed or interpolated to generate a 3D voxel-level texture attribute volume, including the GLCM energy volume and the GLCM contrast volume.

[0042] Step B7: Perform multi-scale filtering and transformation on the seismic data to obtain multi-scale features.

[0043] In this step, multi-scale filters are used to capture detailed information at different scales in the seismic signal, taking into account both large-scale structural features and small-scale fracture characteristics. These multi-scale features are then mapped back to their corresponding three-dimensional voxel spaces to generate a three-dimensional multi-scale feature attribute volume containing rich scale information.

[0044] Step S3: Construct a graph structure based on seismic features and prior features, and obtain a hierarchical graph structure through voxel mapping and high-level clustering; In this embodiment, the lower-level nodes in the hierarchical graph structure describe voxel details, while the higher-level nodes abstract structural units, enabling the simultaneous capture of microscopic texture information of fractures and macroscopic spatial orientation of faults. The higher-level nodes fuse clustering results, weakening local anomaly noise, making it suitable for stable identification in fault regions with low signal-to-noise ratios and unstable data quality. Simultaneously, through spatial connection edges, the hierarchical graph structure maintains the physical adjacency and semantic coherence of geological bodies, improving the identification capability of complex geological structures (such as fault intersections and fracture propagation).

[0045] In step S3, obtaining the hierarchical graph structure includes: Step S31: Map each voxel of the seismic data to a low-level node; Step S32: Based on the position of the voxels, integrate the seismic features and prior features into the node feature vector of the corresponding low-level node; Understandably, mapping each voxel to a low-level node transforms the discrete voxels of seismic data into graph nodes, achieving a structured representation. Integrating seismic features and prior features into node feature vectors allows nodes to carry rich multidimensional attribute information, fusing seismic attributes such as amplitude, coherence, and curvature with geological priors, providing a comprehensive feature description for subsequent analysis. In this way, the data is transformed from raw three-dimensional voxels into graph nodes with multiple attributes, facilitating graph computation and feature fusion.

[0046] Step S33: Construct connecting edges between low-level nodes based on their proximity in space; Understandably, the adjacency relationships of voxels in space are complex, and individual nodes lack spatial context, making it difficult to express local geological continuity. Therefore, by constructing connecting edges between low-level nodes based on the spatial location of voxels, spatial adjacency relationships between voxels are established. Through the definition of edges, the graph structure can accurately reflect the physical adjacency and geological connections between voxels, promote feature propagation and information sharing, and facilitate the detection and identification of continuous anomalies (such as faults and fractures).

[0047] Step S34: Perform unsupervised clustering on the node feature vectors, and treat each cluster as a high-level node; Step S35: Connect each low-level node with the high-level node corresponding to its cluster through the membership relationship to construct cross-level connection edges; In this step, cross-layer connecting edges serve as information bridges between lower and higher layers, enabling the unified propagation of multi-scale information across the graph and ensuring that higher-level decisions can feed back into the detailed identification of lower layers. Meanwhile, edge weights are obtained by weighting spatial proximity and attribute similarity.

[0048] Step S36: Construct connection edges between high-level nodes based on spatial adjacency between high-level nodes to obtain a hierarchical graph structure containing low-level nodes, high-level nodes, and cross-level connection edges.

[0049] It should be noted that faults and fractures typically exhibit different geological characteristics at different spatial scales: small-scale voxel levels reflect detailed textures and waveform anomalies, while large-scale levels demonstrate the spatial extension of fault planes and fracture zones. Based on this, clustering is used to group spatially adjacent and similar low-level nodes into one class, achieving dimensionality reduction and semantic abstraction of the data. Cross-layer connection edges are constructed to connect low-level nodes with higher-level nodes in their respective clusters, enabling information transfer between nodes at different scales. This hierarchical partitioning reduces the number of nodes, improves computational efficiency, and preserves multi-scale information, allowing complex geological units to be expressed at a higher level.

[0050] Therefore, the hierarchical graph structure captures detailed voxel features through low-level nodes and abstracts geological units or fault clusters through high-level nodes, achieving multi-scale fusion and enhancing the comprehensive characterization and identification of fault fracture structures. Meanwhile, seismic data in fault areas often contains noise and anomalous fluctuations, making simple voxel-level identification susceptible to interference. The hierarchical graph effectively suppresses noise and improves the stability and accuracy of fracture identification by clustering and fusing multi-voxel information, and by smoothing and abstracting low-level information through high-level nodes.

[0051] Step S4: Apply Monte Carlo perturbation to the hierarchical graph structure, control the perturbation intensity through reliability indicators, and obtain the perturbation sample graph; It is understandable that seismic data itself has the characteristics of noise interference, strong non-stationarity of seismic features in fracture areas, and difficult-to-define boundaries. Therefore, relying solely on a single-level graph structure for training may cause the network to overfit to a certain version of the graph, ignore the inherent fluctuations and uncertainties of seismic data, and be unable to adapt to structural disturbances and characteristic noise in practical applications.

[0052] Therefore, by introducing different views through Monte Carlo perturbations, the network learns to identify stable features of cracks under various structural changes, reducing its dependence on individual connections or features. For example, with a real crack, even if some connection edges are removed by perturbations, the network can still identify its features from the context.

[0053] In step S4, the step of obtaining the perturbation sample map is as follows: Step S41: Construct a perturbation strategy based on reliability indicators, and add Gaussian noise to the node feature vectors of low-level nodes in the hierarchical graph structure through the perturbation strategy; In this embodiment, the formula for the perturbation strategy is: ; ; In the formula, Indicates the first The node feature vector after perturbation of the lower-level nodes. Indicates the first The node feature vectors of each low-level node. This indicates that the mean is 0 and the variance is 0. Gaussian noise, Indicates the first The characteristic perturbation standard deviation of each low-level node This represents the standard deviation of the largest characteristic perturbation. Indicates the first Reliability metrics for each low-level node.

[0054] Understandably, the perturbation strategy utilizes reliability metrics to achieve adaptive perturbation control, with low-reliability nodes being perturbed to a greater extent to enhance data diversity and robustness, which helps simulate uncertainties and noise interference scenarios in seismic properties.

[0055] Step S42: Calculate the perturbation probability for each connection edge in the hierarchical graph structure based on the reliability index; In this embodiment, the formula for calculating the perturbation probability is: ; In the formula, Representing an edge The probability of perturbation. This represents the probability of the maximum edge perturbation. Indicates the first Reliability metrics of low-level nodes Indicates the first Reliability metrics for each low-level node.

[0056] Step S43: Randomly generate a sampling probability for each connection edge and compare it with the perturbation probability. Perturb the connection edge according to the comparison result. In this embodiment, random generation , representing the sampling probability To randomly sample from [0,1]. Then, perturb that side: If the edge exists, it is randomly deleted; if the edge does not exist, it is randomly added according to the graph structure.

[0057] It should be noted that the above method is directly used for the connection edges between low-level nodes. For the connection edges between high-level nodes, the average reliability index of each high-level node is calculated, and then the perturbation is performed through the above steps.

[0058] Step S44: Obtain the confidence score of each low-level node belonging to each cluster, and obtain the highest confidence score of each low-level node; Step S45: If the highest confidence level is less than the confidence threshold, then obtain the second highest confidence level; Step S46: Perturb the cross-layer connection edges in the hierarchical graph structure based on the second-highest confidence and cross-layer perturbation probability, and reconnect the cross-layer connection edges; In this embodiment, the formula for calculating the cross-layer perturbation probability is as follows: ; In the formula, This represents the probability of cross-layer perturbation. Indicates the maximum disturbance coefficient. Indicates the first The probability values ​​of a low-level node belonging to its primary membership cluster, where the primary membership cluster is the cluster corresponding to the highest confidence level, that is, the cluster with the highest probability of the low-level node.

[0059] Randomly generated , representing the cross-layer sampling probability To randomly sample from [0,1]. If so, the cross-layer connection edge of the lower-level node is changed to connect to the second-highest-confidence cluster, which is the cluster corresponding to the second-highest confidence level.

[0060] Step S47: Generate a perturbation sample map based on the perturbed node feature vectors, connecting edges, and cross-layer connecting edges; Step S48: Repeat the perturbation process to generate multiple sets of perturbation sample images.

[0061] Step S5: Train the graph neural network using perturbed sample images, and obtain the crack recognition network by optimizing the network parameters; Step S5 includes: Step S51: Construct a training set based on multiple sets of perturbation sample images and corresponding crack annotation labels; Step S52: Initialize the network parameters of the graph neural network, which includes a feature encoding layer, multiple feature learning and enhancement layers, a multi-scale fusion layer and a fully connected layer connected in sequence; In this embodiment, the feature encoding layer includes a low-level node feature encoding layer and a high-level node feature encoding layer. The feature learning and enhancement layer includes a GAT layer (Graph Attention Network layer) and a cross-layer message passing layer. Both the low-level node feature encoding layer and the high-level node feature encoding layer employ fully connected layers and the ReLU activation function. Figure 2 As shown, in this embodiment, three sequentially connected feature learning and enhancement layers are selected.

[0062] Specifically, linear transformations and nonlinear activations are performed through the feature encoding layer. That is, the feature vectors of low-level nodes are encoded by the low-level node feature encoding layer to obtain low-dimensional embedded features. The feature vectors of high-level nodes are encoded by the high-level node feature encoding layer to obtain high-order semantic representations.

[0063] In each of the multiple feature learning and enhancement layers, the low-dimensional embedding features and high-order semantic representations are first processed by the GAT layer. The GAT layer includes a first GAT layer and a second GAT layer, specifically: ; ; In the formula, and They represent the first Layer and first The first low-level node feature representation of the layer. and They represent the first Layer and first The second high-level node feature representation of the layer. This represents the set of edges of the lower-level nodes. Represents the set of edges of higher-level nodes. and These represent the first GAT layer and the second GAT layer, respectively.

[0064] Then, through a cross-layer message passing layer, information exchange between low and high layers is performed based on an attention mechanism: ; ; In the formula, and They represent the first The feature representation of the second lower-level node and the feature representation of the second higher-level node in the layer. Represents the set of cross-layer connection edges. and Both represent fusion coefficients. This represents the attention mechanism.

[0065] Finally, residual connections are performed to obtain the output features of the feature learning and enhancement layers: ; ; In the formula, and They represent the first The feature representation of the third lower-level node and the feature representation of the third higher-level node.

[0066] The feature learning and the feature representations of the third low-level node and the third high-level node of the last layer of the enhancement layer are aggregated through a multi-scale fusion layer to obtain fused features, which are then passed through a fully connected layer to obtain the final result.

[0067] Understandably, this graph neural network captures fine-grained local features through feature encoding layers and aggregates local information to form high-order semantic representations. Employing multiple (3-layer) feature learning and enhancement layers progressively deepens message passing and feature fusion between nodes, which is beneficial for capturing more complex inter-node dependencies and enhancing the network's expressive power. A cross-layer message passing layer enables information interaction between low-level and high-level nodes through an attention mechanism, promoting the effective fusion of features at different levels. Furthermore, residual connections avoid the gradient vanishing problem in deep network training, improving training stability and convergence speed.

[0068] Step S53: Construct a joint loss function based on perturbation consistency, multi-scale spatial consistency, topological regularity, and weighted cross-entropy; It should be noted that, to improve the accuracy and robustness of shale fault fracture identification based on graph neural networks, a joint loss function was designed that comprehensively considers classification accuracy, perturbation stability, multi-scale spatial consistency, and topological constraints: ; In the formula, Denotes the joint loss function. This represents the weighted cross-entropy loss. This indicates the loss of consistency due to perturbation. This represents the loss of spatial consistency across multiple scales. Represents the loss for topology regularization. , and All of these represent weighting coefficients.

[0069] Specifically, weighted cross-entropy loss is a fundamental loss function in classification tasks, used to measure the difference between the network's predictions and the true labels. Since shale fault fractures typically constitute a small proportion of the overall structure, class imbalance exists. Node weights are introduced to adjust the contributions of different samples, preventing the network from favoring the majority class and thus improving the ability to identify fracture nodes. This loss quantifies the classification error by calculating the log-likelihood of the true label and predicted probability for each node.

[0070] Weighted cross-entropy loss The expression is: ; In the formula, Indicates the number of lower-level nodes. Indicates the first The weight of each low-level node, Indicates the first The true label of a low-level node Indicates the prediction of the first The probability that a low-level node is a crack.

[0071] To improve the network's robustness to input perturbations and avoid prediction instability caused by data noise or measurement errors, a perturbation consistency loss is designed. Specifically, multiple perturbations are applied to the same graph data, and the sum of squared deviations between the predicted probability of each node on each perturbation sample and the mean prediction of all perturbation samples for that node is calculated. This loss constrains the network output to remain consistent under perturbations, enhancing the network's stability and generalization ability.

[0072] Disturbance Consistency Loss The expression is: ; In the formula, This represents the number of perturbation samples. Indicates the first Predicting the th perturbation sample The probability that a low-level node is a crack. Indicates the first The predicted average value of each low-level node across all perturbed samples.

[0073] The spatial distribution of shale fault fractures exhibits a hierarchical structure, displaying multi-scale features at both low-level (local details) and high-level (global semantics) perspectives. The multi-scale spatial consistency loss, through the edge weights of the graph, promotes smoothing of the predicted probabilities of adjacent nodes at the same level, reflecting the spatial continuity of the geological structure. This loss helps suppress isolated anomalous predictions and enhances the spatial coherence of fracture identification.

[0074] Multi-scale spatial consistency loss The expression is: ; In the formula, This indicates hierarchy, including lower and higher levels. Indicates hierarchy The Middle The first low-level node and the first The edge weights of each low-level node. Denotes the Euclidean norm. and Representing levels China predicts the first The first low-level node and the first The probability that a low-level node is a crack.

[0075] The topological relationships of fracture nodes have prior knowledge in geological structures, and the topological similarity between nodes (topological prior) The topology regularization loss can be determined based on geological features or spatial proximity. It utilizes a weighted average of the prediction probability differences between adjacent nodes. Adjusting the penalty intensity encourages the network to output similar crack probabilities among nodes with high topological similarity, maintaining the rationality and physical consistency of the topology, and avoiding unreasonable prediction jumps.

[0076] Topology regularization loss The expression is: ; In the formula, Indicates the first The set of neighbors of each low-level node. This indicates taking the absolute value. Indicates the prediction of the first The probability that a low-level node is a crack. Indicates the first The first low-level node and the first Topological priors of the lower-level nodes.

[0077] Step S54: Jointly train the graph neural network using the training set, and update the network parameters based on the joint loss function, forward propagation, and back propagation; In this embodiment, the training set is divided into several batches, each batch containing multiple sets of perturbation sample images and corresponding labels. The perturbation sample image data from each batch is input into a graph neural network, sequentially passing through a feature encoding layer, multiple feature learning and enhancement layers, a multi-scale fusion layer, and a fully connected layer to obtain the crack prediction probability for each low-level node. Based on the prediction results and the true labels, the joint loss function is calculated. Backpropagation is performed based on the function value to calculate the gradient of the network parameters at each layer. The gradient descent optimization algorithm is used to update the network parameters, adjusting the weights and biases to minimize the joint loss function.

[0078] After each training round, network performance metrics (accuracy, recall, F1 score, etc.) can be calculated based on the validation set or cross-validation method to monitor training progress and prevent overfitting. When training reaches the preset number of iterations or the performance metrics converge, training is stopped, and the network parameters of the finally trained graph neural network are saved for subsequent crack recognition tasks.

[0079] Step S55: Obtain the crack recognition network using the updated network parameters and graph neural network.

[0080] Step S6: Obtain actual seismic data of the target shale fault area, construct an actual map structure based on the actual seismic data, and identify cracks in the actual map structure through a crack identification network to obtain crack identification results.

[0081] In this embodiment, the crack nodes identified by the graph neural network are actually a large number of points scattered in three-dimensional space. These points roughly reflect the possible locations of cracks, but they are discrete, numerous, and may contain noise. Therefore, steps such as spatial connectivity analysis, directional consistency elimination, and fragment merging are used to transform the point-like distribution into a linear crack structure, which facilitates the intuitive representation and understanding of the crack's morphology and direction in three-dimensional space, helping to identify and describe the actual cracks.

[0082] In step S6, the crack identification of the actual graph structure using a crack identification network to obtain crack identification results includes: Step S61: Use the crack identification network to identify cracks in the actual graph structure to obtain crack nodes; Step S62: Combine the reliability index of prior features to classify all crack nodes and obtain high-reliability nodes and low-reliability nodes; Step S63: Obtain the number of high-reliability nodes within a preset spatial range of low-reliability nodes; Step S64: If the number of highly reliable nodes is less than the preset number, then remove the corresponding low-reliability nodes; Step S65: Based on the high-reliability nodes and the retained low-reliability nodes, perform neighborhood connectivity analysis and extract the initial crack segments composed of connected crack nodes. Step S66: Calculate the principal direction of each initial crack segment, and remove abnormal segments with disordered directions based on the principal direction, while retaining valid segments with consistent directions; In this embodiment, each initial crack segment is composed of multiple crack nodes. Therefore, the center point of each initial crack segment is calculated, and then the covariance matrix is ​​constructed to calculate the eigenvector. The unit eigenvector corresponding to the largest eigenvalue is taken as the principal direction.

[0083] For each initial crack segment, if the variance of the direction generated by its main direction in its own node is greater than the direction threshold, it is considered to have chaotic direction and the segment is discarded.

[0084] Step S67: Calculate the spatial distance and orientation difference between valid segments: In this embodiment, for two valid segments, the distance between their center points is calculated as the spatial distance, and the directional difference is calculated through the main directions of the two valid segments.

[0085] Step S68: Merge effective segments based on spatial distance and directional differences to obtain crack identification results.

[0086] In this embodiment, a comprehensive similarity index is obtained by weighted fusion of spatial distance and directional differences. Using this index and a set merging threshold, valid segments with consistent orientation and spatial proximity are merged into a single complete rift. After merging, the main direction and center point of the merged segment are recalculated, and the segment set is updated. This process is repeated until all segments can no longer be merged.

[0087] In summary, this invention fully integrates seismic characteristics and prior geological information, constructs a hierarchical graph structure, and introduces a perturbation mechanism to enhance the network's adaptability to data uncertainty. Through node and edge perturbations controlled by reliability indices, the robustness of the graph neural network in identifying cracks in uncertain regions is effectively improved, especially in complex tectonic zones and seismic data with low signal-to-noise ratios, where stable identification is still achievable.

[0088] Furthermore, during the crack result screening stage, reliability indicators and connectivity analysis are combined to effectively eliminate misidentified nodes. Crack segments are merged through principal direction extraction and spatial relationship aggregation, ultimately obtaining crack identification results with good continuity and clear structure. This method not only improves the accuracy and completeness of crack identification but also has significant engineering application value.

[0089] Example 2: This embodiment provides a shale fault fracture identification device based on deep learning, the device comprising: The acquisition module is used to acquire seismic data in shale fault zones; The extraction module is used to extract features and perform seismic inversion on seismic data to obtain seismic features and prior features. The seismic features include correlation, curvature value, instantaneous amplitude, instantaneous frequency, texture attributes and multi-scale features. The prior features include geostress-sensitive parameters and corresponding reliability indices. The construction module is used to build a graph structure based on seismic features and prior features, and obtain a hierarchical graph structure through voxel mapping and high-level clustering. The perturbation module is used to perform Monte Carlo perturbation on the hierarchical graph structure. The perturbation intensity is controlled by the reliability index to obtain the perturbation sample graph. The training module is used to train the graph neural network using perturbed sample maps, and obtain the crack recognition network by optimizing the network parameters; The identification module is used to acquire actual seismic data of the target shale fault area, construct an actual map structure based on the actual seismic data, and identify cracks in the actual map structure through a crack identification network to obtain crack identification results.

[0090] The building module includes: A mapping unit is used to map each voxel of seismic data to a low-level node. The integration unit is used to integrate seismic features and prior features into node feature vectors for corresponding low-level nodes based on the location of voxels. The first building unit is used to construct connecting edges between low-level nodes based on their spatial proximity. The second building unit is used to perform unsupervised clustering on the node feature vectors, treating each cluster as a high-level node. The third building unit is used to connect each low-level node with the high-level node corresponding to its cluster through the membership relationship, and build cross-level connection edges. The fourth building unit is used to construct connection edges between high-level nodes based on the spatial adjacency between high-level nodes, resulting in a hierarchical graph structure containing low-level nodes, high-level nodes, and cross-level connection edges.

[0091] The disturbance module includes: The fifth building unit is used to construct a perturbation strategy based on reliability indicators, and to add Gaussian noise to the node feature vectors of low-level nodes in the hierarchical graph structure through the perturbation strategy. The first calculation unit is used to calculate the perturbation probability of each connection edge in the hierarchical graph structure based on the reliability index. The first perturbation unit is used to randomly generate a sampling probability for each connection edge, compare it with the perturbation probability, and perturb the connection edge according to the comparison result; The first acquisition unit is used to acquire the confidence level of each low-level node belonging to each cluster and to acquire the highest confidence level of each low-level node. The second acquisition unit is used to acquire the second highest confidence level if the highest confidence level is less than the confidence threshold. The second perturbation unit is used to perturb the cross-layer connection edges in the hierarchical graph structure based on the second highest confidence level and the cross-layer perturbation probability, and reconnect the cross-layer connection edges. The first generation unit is used to generate a perturbed sample graph based on the perturbed node feature vectors, connecting edges, and cross-layer connecting edges. The second generation unit is used to repeatedly execute the perturbation process and generate multiple sets of perturbation sample maps.

[0092] The identification module includes: The identification unit is used to identify cracks in the actual graph structure using the crack identification network to obtain crack nodes. The classification unit is used to classify all crack nodes by combining the reliability index of prior features, and obtain high-reliability nodes and low-reliability nodes. The third acquisition unit is used to acquire the number of high-reliability nodes within a preset spatial range of low-reliability nodes; The first filtering unit is used to remove the corresponding low-reliability nodes if the number of high-reliability nodes is less than a preset number. Connectivity unit is used to perform neighborhood connectivity analysis based on high-reliability nodes and retained low-reliability nodes, and to extract the initial crack segment composed of connected crack nodes. The second screening unit is used to calculate the main direction of each initial crack segment, remove abnormal segments with disordered directions based on the main direction, and retain valid segments with consistent directions. The second calculation unit is used to calculate the spatial distance and directional differences between valid segments: The merging unit is used to merge effective segments based on spatial distance and directional differences to obtain crack identification results.

[0093] It should be noted that the specific manner in which each module performs its operation in the apparatus described in the above embodiments has been described in detail in the embodiments of the method, and will not be elaborated here.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0095] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying shale fault fractures based on deep learning, characterized in that, include: Acquire seismic data in shale fault zones; Seismic data is subjected to feature extraction and seismic inversion to obtain seismic features and prior features. The seismic features include correlation, curvature value, instantaneous amplitude, instantaneous frequency, texture attributes and multi-scale features. The prior features include geostress-sensitive parameters and corresponding reliability indices. A graph structure is constructed based on seismic features and prior features, and a hierarchical graph structure is obtained through voxel mapping and high-level clustering. Monte Carlo perturbation is applied to the hierarchical graph structure, and the perturbation intensity is controlled by reliability indicators to obtain the perturbation sample graph; A graph neural network was trained using perturbed sample images, and the crack identification network was obtained by optimizing the network parameters. The actual seismic data of the target shale fault area is obtained, the actual map structure is constructed based on the actual seismic data, and the crack identification results are obtained by using a crack identification network to identify cracks in the actual map structure.

2. The shale fault fracture identification method based on deep learning according to claim 1, characterized in that, The steps for obtaining the prior features are as follows: Seismic wave impedance inversion is performed based on seismic data to obtain inversion results, which include P-wave impedance, S-wave impedance, and density. By combining prior information about the well location and the geological background, spatial correction and uncertainty assessment are performed on the inversion results to obtain the corrected inversion results and the corresponding reliability indicators. The geostress-sensitive parameters are calculated based on the corrected inversion results. These geostress-sensitive parameters include Poisson's ratio, brittleness index, and the ratio of P-wave to S-wave velocity.

3. The shale fault fracture identification method based on deep learning according to claim 1, characterized in that, The resulting hierarchical graph structure includes: Map each voxel of the seismic data to a low-level node; Based on the location of the voxels, the seismic features and prior features are integrated into the node feature vector of the corresponding low-level node; Based on the proximity relationships of low-level nodes in space, construct connecting edges between low-level nodes; Unsupervised clustering is performed on the node feature vectors, and each cluster is treated as a high-level node. Each low-level node is connected to the high-level node corresponding to its cluster by an edge based on the membership relationship, thus constructing a cross-level connection edge. Based on the spatial adjacency between high-level nodes, connection edges between high-level nodes are constructed, resulting in a hierarchical graph structure containing low-level nodes, high-level nodes, and cross-level connection edges.

4. The shale fault fracture identification method based on deep learning according to claim 1, characterized in that, The steps for obtaining the perturbation sample map are as follows: A perturbation strategy is constructed based on reliability indicators, and Gaussian noise is added to the node feature vectors of low-level nodes in the hierarchical graph structure through the perturbation strategy. The perturbation probability is calculated for each connection edge in the hierarchical graph structure based on reliability indicators; For each connecting edge, a sampling probability is randomly generated and compared with the perturbation probability. The connecting edge is then perturbed based on the comparison result. Obtain the confidence score of each low-level node belonging to each cluster, and obtain the highest confidence score of each low-level node; If the highest confidence level is less than the confidence threshold, then the second highest confidence level is obtained; Based on the second-highest confidence level and the cross-layer perturbation probability, the cross-layer connection edges in the hierarchical graph structure are perturbed and reconnected. A perturbation sample graph is generated based on the perturbation of node feature vectors, connecting edges, and cross-layer connecting edges. Repeat the perturbation process to generate multiple sets of perturbation sample images.

5. The shale fault fracture identification method based on deep learning according to claim 1, characterized in that, The process of training a graph neural network using perturbed sample images and optimizing the network parameters to obtain a crack identification network includes: A training set is constructed based on multiple sets of perturbation sample images and corresponding crack annotation labels; Initialize the network parameters of the graph neural network, which includes a feature encoding layer, multiple feature learning and enhancement layers, a multi-scale fusion layer, and a fully connected layer connected in sequence. A joint loss function is constructed based on perturbation consistency, multi-scale spatial consistency, topological regularity, and weighted cross-entropy. The graph neural network is jointly trained using the training set, and the network parameters are updated based on the joint loss function, forward propagation, and back propagation. The crack identification network is obtained by updating the network parameters and using a graph neural network.

6. The shale fault fracture identification method based on deep learning according to claim 1, characterized in that, The process of identifying cracks in the actual graph structure using a crack identification network to obtain crack identification results includes: The crack identification network is used to identify cracks in the actual graph structure to obtain crack nodes. By combining the reliability index of prior features, all crack nodes are classified into high-reliability nodes and low-reliability nodes. Obtain the number of highly reliable nodes within a preset spatial range of low-reliability nodes; If the number of highly reliable nodes is less than the preset number, the corresponding low-reliability nodes are removed. Based on high-reliability nodes and retained low-reliability nodes, neighborhood connectivity analysis is performed to extract the initial crack segments formed by connected crack nodes. Calculate the principal direction of each initial crack segment, and discard abnormal segments with disordered directions based on the principal direction, retaining valid segments with consistent directions; Calculate the spatial distance and directional differences between valid segments: Based on spatial distance and directional differences, effective segments are merged to obtain crack identification results.

7. A shale fault fracture identification device based on deep learning, characterized in that, include: The acquisition module is used to acquire seismic data in shale fault zones; The extraction module is used to extract features and perform seismic inversion on seismic data to obtain seismic features and prior features. The seismic features include correlation, curvature value, instantaneous amplitude, instantaneous frequency, texture attributes and multi-scale features. The prior features include geostress-sensitive parameters and corresponding reliability indices. The construction module is used to build a graph structure based on seismic features and prior features, and obtain a hierarchical graph structure through voxel mapping and high-level clustering. The perturbation module is used to perform Monte Carlo perturbation on the hierarchical graph structure. The perturbation intensity is controlled by the reliability index to obtain the perturbation sample graph. The training module is used to train the graph neural network using perturbed sample maps, and obtain the crack recognition network by optimizing the network parameters; The identification module is used to acquire actual seismic data of the target shale fault area, construct an actual map structure based on the actual seismic data, and identify cracks in the actual map structure through a crack identification network to obtain crack identification results.

8. The shale fault fracture identification device based on deep learning according to claim 7, characterized in that, The building module includes: A mapping unit is used to map each voxel of seismic data to a low-level node. The integration unit is used to integrate seismic features and prior features into node feature vectors for corresponding low-level nodes based on the location of voxels. The first building unit is used to construct connecting edges between low-level nodes based on their spatial proximity. The second building unit is used to perform unsupervised clustering on the node feature vectors, treating each cluster as a high-level node. The third building unit is used to connect each low-level node with the high-level node corresponding to its cluster through the membership relationship, and build cross-level connection edges. The fourth building unit is used to construct connection edges between high-level nodes based on the spatial adjacency between high-level nodes, resulting in a hierarchical graph structure containing low-level nodes, high-level nodes, and cross-level connection edges.

9. The shale fault fracture identification device based on deep learning according to claim 7, characterized in that, The disturbance module includes: The fifth building unit is used to construct a perturbation strategy based on reliability indicators, and to add Gaussian noise to the node feature vectors of low-level nodes in the hierarchical graph structure through the perturbation strategy. The first calculation unit is used to calculate the perturbation probability of each connection edge in the hierarchical graph structure based on the reliability index. The first perturbation unit is used to randomly generate a sampling probability for each connection edge, compare it with the perturbation probability, and perturb the connection edge according to the comparison result; The first acquisition unit is used to acquire the confidence level of each low-level node belonging to each cluster and to acquire the highest confidence level of each low-level node. The second acquisition unit is used to acquire the second highest confidence level if the highest confidence level is less than the confidence threshold. The second perturbation unit is used to perturb the cross-layer connection edges in the hierarchical graph structure based on the second highest confidence level and the cross-layer perturbation probability, and reconnect the cross-layer connection edges. The first generation unit is used to generate a perturbed sample graph based on the perturbed node feature vectors, connecting edges, and cross-layer connecting edges. The second generation unit is used to repeatedly execute the perturbation process and generate multiple sets of perturbation sample maps.

10. The shale fault fracture identification device based on deep learning according to claim 7, characterized in that, The identification module includes: The identification unit is used to identify cracks in the actual graph structure using the crack identification network to obtain crack nodes. The classification unit is used to classify all crack nodes by combining the reliability index of prior features, and obtain high-reliability nodes and low-reliability nodes. The third acquisition unit is used to acquire the number of high-reliability nodes within a preset spatial range of low-reliability nodes; The first filtering unit is used to remove the corresponding low-reliability nodes if the number of high-reliability nodes is less than a preset number. Connectivity unit is used to perform neighborhood connectivity analysis based on high-reliability nodes and retained low-reliability nodes, and to extract the initial crack segment composed of connected crack nodes. The second screening unit is used to calculate the main direction of each initial crack segment, remove abnormal segments with disordered directions based on the main direction, and retain valid segments with consistent directions. The second calculation unit is used to calculate the spatial distance and directional differences between valid segments: The merging unit is used to merge effective segments based on spatial distance and directional differences to obtain crack identification results.

Images