An intelligent classification method based on an amphibious crack in-situ collection device

Abstract

The application provides an intelligent classification method based on an amphibious crack in-situ acquisition device, relates to the technical fields of geological survey, oil and gas exploration and machine vision, and the application obtains original digital images containing geological structure surfaces returned by the amphibious acquisition device; a preset deep neural network model is inputted, and a two-dimensional binary mask image of cracks is outputted; skeletonization processing is performed on the images to construct a topological skeleton network, topological key nodes are extracted based on the connection degrees of skeleton pixels, and spatially independent crack samples are separated; for each crack sample, a multi-dimensional feature space vector is constructed, and an expert decision tree is inputted; the geological genesis type and evolution period of each crack sample are determined according to preset condition thresholds, and a visual classification result is outputted. Through the composite chassis design of the track and the retractable vector propeller, in combination with the pressure-resistant and waterproof streamline layered cabin, the walking obstacle-crossing performance and image acquisition stability under the amphibious complex flow field environment are improved.

Description

Technical Field

[0001] This invention relates to the fields of geological surveying, oil and gas exploration and machine vision, and in particular to an intelligent classification method based on an amphibious fracture in-situ acquisition device. Background Technology

[0002] In the exploration and development of oil and gas, especially unconventional oil and gas, the occurrence, development degree, and genetic type of natural faults and microfractures directly determine the reservoir's storage space and fluid seepage characteristics. In the operation and maintenance of water conservancy projects and geotechnical engineering investigations, they directly affect the overall mechanical stability of the rock mass. Currently, the acquisition of geological fracture information largely relies on traditional borehole photography or underwater robot scanning methods. However, these technologies still face many technical challenges in practical engineering applications.

[0003] First, the hardware acquisition layer suffers from poor environmental adaptability and insufficient acquisition stability. Conventional exploration equipment mostly adopts wheeled propulsion structures or fixed-wing buoyancy control methods, with limited attitude control modes and weak obstacle-crossing capabilities, making it unsuitable for complex operating environments with high obstacles, such as irregular narrow holes, blind alleys, or soft sediment surfaces on land. Furthermore, in water-related operations, dam base areas, and complex flow field environments, traditional acquisition equipment is highly susceptible to disturbances from dynamic water pressure, making it difficult to maintain a relatively stable hovering and anchoring attitude around the target rock wall. This ultimately leads to motion blur and image distortion in the acquired crack images, compromising the integrity and accuracy of the raw data.

[0004] Secondly, image interpretation and intelligent recognition algorithms have significant shortcomings. When processing rock wall images contaminated with mud or subject to underwater scattering interference, existing conventional convolutional neural networks are prone to irreversible loss of high-frequency edge details of micro-cracks due to downsampling operations, drastically reducing the accuracy of identifying minute cracks. Simultaneously, current crack recognition technologies are limited to extracting static geometric parameters such as crack length and width, failing to establish a quantitative mapping relationship between crack morphology and underlying geological genesis and multi-stage evolutionary timelines. This hinders refined, source-tracing analysis of geological cracks and fails to meet the application requirements of high-precision geological engineering evaluation. Summary of the Invention

[0005] The main objective of this invention is to improve the walking and obstacle-crossing performance and image acquisition stability in complex amphibious flow field environments by using a composite chassis design of tracks and retractable vector propellers, combined with a pressure-resistant and waterproof streamlined layered cabin. At the same time, an algorithm combining frequency domain feature preservation and multi-dimensional topological-geometric feature conditional decision-making is proposed to achieve automated classification and interpretation of the formation stages of microcracks. Therefore, an intelligent classification method based on an amphibious crack in-situ acquisition device is proposed.

[0006] Another objective of this invention is to provide an amphibious in-situ fissure sampling device.

[0007] To achieve the above objectives, a first aspect of the present invention proposes an intelligent classification method based on an amphibious fissure in-situ acquisition device, comprising: S1, acquire the original digital image containing the geological structure surface transmitted back by the amphibious fissure in-situ acquisition device; S2, input the original digital image into a preset deep neural network model, use frequency domain decomposition to retain high-frequency edge features and combine spatial attention mechanism to suppress background noise, and output a two-dimensional binarized mask image of the crack. S3, skeletonize the two-dimensional binary mask image to construct a topological central axis network, extract key topological nodes based on the connectivity degree of the skeleton pixels, and separate spatially independent crack samples. S4. For each fracture sample, calculate its global topological connectivity index, dimensionless aperture variation coefficient, strike angle, tortuosity, and cutting priority to construct a multidimensional feature space vector. Input the multidimensional feature space vector into the expert decision tree and determine the geological genesis type and evolution stage of each fracture sample according to the preset condition threshold. Output the visualized classification results.

[0008] In one embodiment of the present invention, S2 includes: In the feature downsampling path of the encoder, the discrete wavelet transform module is used to decompose the input feature map into a low-frequency approximation component and three high-frequency detail components through a preset orthogonal filter. The low-frequency approximation component and the three high-frequency detail components are concatenated and stitched together in the channel dimension to achieve lossless transmission of high-frequency edge texture features. In the skip connection path between the encoder and decoder, a spatial attention gating module is embedded. It uses deep semantic features to generate a spatial attention weight matrix to recalibrate shallow features, so as to adaptively suppress background noise and output a two-dimensional binarized mask image.

[0009] In one embodiment of the present invention, the step of using a discrete wavelet transform module to decompose the input feature map into a low-frequency approximation component and three high-frequency detail components through a preset orthogonal filter includes: The orthogonal filtering matrix of the Haar wavelet basis is used to perform cross product operations of low-pass and high-pass basis functions in the horizontal and vertical directions, which transforms the input spatial domain feature map into four frequency domain independent sub-bands with half the spatial resolution. The four frequency domain independent sub-bands include low-frequency approximate components, horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components.

[0010] In one embodiment of the present invention, the step of concatenating and splicing the low-frequency approximation component with the three high-frequency detail components in the channel dimension includes: The low-frequency approximation component and the three high-frequency detail components are fully concatenated along the channel axis by the Concat concatenation operation to obtain the concatenated feature tensor. A convolutional layer with a kernel size of 1×1 is used to perform cross-channel information interaction and dynamic weighting on the cascaded feature tensor, compressing the number of channels to a specified 2C dimension, and outputting the processed feature data.

[0011] In one embodiment of the present invention, S3 includes: A skeletonization algorithm is performed on the two-dimensional binary mask image to obtain a topological skeleton network with a single pixel width; The connection degree of each skeleton pixel in the topological skeleton network is calculated using 3×3 neighborhood convolution, and the endpoints with a degree of 1, the bifurcation points with a degree of 3, and the cross-cutting points with a degree of 4 are extracted as topological key nodes. Spatially independent crack samples were separated based on the key nodes of the topology using a connected component algorithm.

[0012] In one embodiment of the present invention, the global topological connectivity index is calculated as follows: the total number of crack bifurcation points and intersection cutting points is counted, and then the total number of all crack endpoints, bifurcation points and intersection cutting points is combined to calculate the global topological connectivity index. The dimensionless aperture variation coefficient is extracted by using Euclidean distance transformation to extract the aperture data sequence of the entire crack skeleton, calculating the fluctuation range and average value of the aperture data, and thus obtaining the corresponding coefficient; at the same time, a small correction value is added to avoid the situation where the denominator is empty during the calculation process.

[0013] In one embodiment of the present invention, the method for determining the cutting priority is as follows: When there are intersecting points in the crack sample, the pixel coordinates of the local skeleton window centered on the intersecting points are extracted for principal component analysis, and the ratio of the first eigenvalue to the second eigenvalue of its covariance matrix is ​​calculated as the local interpenetration linearity. If the local interleaving linearity is greater than a preset linearity threshold, then a cutting priority is set. , indicating late-stage generation; If the local interleaving linearity is less than or equal to a preset linearity threshold, then a cutting priority is set. This indicates that the node was generated early and cut off by a later crack; if there are no intersecting nodes, a cutting priority is set. .

[0014] In one embodiment of the present invention, the orientation angle is extracted by using principal component analysis to extract the main orientation of the individual fracture and calculating the angle between the main orientation and the reference bedding. The method for extracting the tortuosity is as follows: the length of the straight trace is evaluated using the long side of the smallest bounding rectangle, and the tortuosity is the ratio of the actual skeleton pixel length to the length of the straight trace.

[0015] In one embodiment of the present invention, the expert decision tree includes the following judgment rules: Artificial artifact filtering rules: When the tortuosity and dimensionless aperture coefficient of variation of the sample object are both lower than the first preset threshold, and the global topological connectivity index is zero, it is judged as a non-geological artificial artifact and removed. Shear fracture surface determination rule: When the cutting priority is late generation level and the dimensionless aperture variation coefficient is lower than the second preset threshold, it is determined to be a late master-controlled high penetration shear fracture surface. Multi-stage tension crack determination rules: When the cutting priority is at the early generation level and the dimensionless aperture variation coefficient is higher than the third preset threshold, it is determined to be an early-evolution multi-stage staggered stress tension crack; it also includes: Multi-stage superimposed highly connected networks under the strong heterogeneous composite fracture network criteria; highly efficient expansion tectonic tension fractures belonging to high-angle development and fluid overpressure-driven hydraulic expansion; primary bedding-parallel detachment fractures of the earliest developed syn-sedimentary sequence, strictly restricted by weak bedding planes; and diagenetic local contraction microfractures with extremely small scale and free, non-connected ends.

[0016] To achieve the above objectives, a second aspect of the present invention provides an amphibious in-situ acquisition device for fissures, comprising: a power module, an in-situ image acquisition and processing module, and a composite cabin, wherein: Power module: Tracked chassis with symmetrical track mechanisms on both sides of the chassis, and four road wheels under each track to provide power for walking on irregular geological surfaces or in narrow holes. The retractable dynamic positioning assembly is symmetrically embedded in the central area of ​​the chassis and includes two sets of retractable propellers. The retraction mechanism controls the opening or closing of the propellers, corresponding to the underwater working mode and the land working mode, respectively. Image in-situ acquisition and processing module: Equipped with a high-pixel lens device and a ring light illumination device, it is integrated at the front of the first-layer cabin to provide real sample data under different water and land conditions. The returned data is temporarily stored in the data transmission module in the second layer, and then the image data is transmitted to the terminal for visualization and analysis. In addition, the intelligent control module will control the movement and sampling of the control device under different water and land conditions. Composite Cabin: A double-layer composite cabin, fixedly mounted on top of the chassis. The first layer is a pressure-resistant and waterproof streamlined cabin, with an in-situ image acquisition and processing module facing the direction of travel at its front end. The second layer is stacked on top of the first layer and has pressure-resistant and waterproof functions. It integrates a data transmission module and an intelligent control module.

[0017] The intelligent classification method and device based on the amphibious in-situ crack acquisition device of the present invention effectively improves the stability and clarity of crack image acquisition in complex amphibious environments, and realizes automated and refined classification and interpretation of the formation stages of microcracks.

[0018] The beneficial effects of this invention are as follows: Significantly improved multi-terrain obstacle crossing performance and attitude maintenance capability: The walking mechanism adopts a tank-like dual-track chassis with four road wheels on each side, which enhances the adhesion and obstacle crossing height on soft land surfaces and irregular hard rock surfaces; the two sets of retractable propellers in the middle of the chassis open outward in wading or complex flow fields, and the thrust is dynamically fine-tuned by the intelligent control module to effectively counteract water flow disturbances. At the same time, it can maintain a stable floating state when used in water, ensuring the shooting effect of the device, and achieving highly stable attitude maintenance and efficient operation near the structural surface.

[0019] The cabin structure features high penetration resistance and functional independence: In the double-layer cabin design, the streamlined pressure-resistant and waterproof structure of the lower layer reduces fluid resistance and isolates high-pressure water, protecting the front-end visual perception module; the upper communication cabin is independently arranged, reducing the shielding interference of the power circuit and the waterproof cabin on electromagnetic signals, and ensuring real-time high-bandwidth data transmission.

[0020] Microcrack interpretation has physical and mechanical interpretability: the algorithm introduces dimensionless aperture variation coefficient and linearity analysis of local eigenvalues ​​based on cross nodes, transforming the apparent geometric morphology into a physical vector that can characterize the stress state and cutting sequence of geological movement. The fine interpretation of the genetic period is achieved through a fully closed expert decision tree.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating an intelligent classification method based on an amphibious fissure in-situ acquisition device, provided in an embodiment of the present invention; Figure 2 A three-dimensional top view schematic diagram of an amphibious in-situ fissure acquisition device is provided for embodiments of the present invention; Figure 3 A three-dimensional side view schematic diagram of an amphibious fissure in-situ acquisition device is provided for embodiments of the present invention; Figure 4 A bottom schematic diagram of the amphibious fissure in-situ data acquisition device is provided for embodiments of the present invention; Figure 5 The present invention provides an image in-situ acquisition and processing module for an amphibious fissure in-situ acquisition device. Figure 6 A diagram showing the stowed state of the retractable propeller in the middle of the chassis is provided for an embodiment of the present invention. Figure 7 A diagram showing the deployed state of the retractable propeller thruster in the middle of the chassis is provided for an embodiment of the present invention. Figure 8 This invention provides a demonstration of the main design of a retractable propeller thruster located in the middle of the chassis, as shown in the embodiments of the present invention. Figure 9 This is a perspective curve diagram of the device; Figure 10 This is a flowchart illustrating the overall process of data acquisition and classification interpretation provided in this embodiment of the invention. Figure 11 This is a pre-defined deep neural network architecture topology diagram in an embodiment of the present invention; Figure 12 This is a schematic diagram illustrating the eigenvalue decomposition and lossless cascading principle of the Discrete Wavelet Transform (DWT) module in an embodiment of the present invention. Figure 13 A flowchart of the multidimensional topological-geometric feature-based decision classification tree constructed for embodiments of the present invention; Figure 14 This is a schematic diagram of a visual interpretation result with category annotations output by an embodiment of the present invention; in, Power module (1): Tracked chassis (11); Road wheels (12); Propeller (13); Retraction and extension actuator (14). Image in-situ acquisition and processing module (2): High-pixel lens device (21); Ring light illumination device (22); Data transmission module (23); Intelligent control module (24); Composite cabin (3): First layer cabin (31); Second layer cabin (32). Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] The following describes, with reference to the accompanying drawings, an intelligent classification method based on an amphibious fissure in-situ acquisition device, according to an embodiment of the present invention.

[0026] Example 1 This embodiment provides an intelligent classification method based on an amphibious in-situ fissure acquisition device. For example... Figure 1 As shown, it includes: S1, acquire the original digital image containing the geological structure surface transmitted back by the amphibious in-situ fracture acquisition device.

[0027] Specifically, acquiring raw digital images containing geological structural surfaces transmitted back by an amphibious acquisition device is the data foundation for subsequent crack identification and classification interpretation. This step aims to acquire raw image data that accurately reflects the morphology, distribution, and texture characteristics of micro-cracks on the rock wall surface using a dedicated acquisition device deployed in the target exploration area. The acquisition device is adaptable to amphibious environments and can operate stably in various scenarios, including irregular terrestrial surfaces, narrow openings, and complex underwater flow fields. Its front-end image acquisition module takes directional photos of the target rock wall during movement or hovering, generating digital images containing crack information. The acquired raw digital images must have sufficient spatial resolution and illumination uniformity to ensure that the edge details and grayscale contrast of the micro-cracks are fully preserved, providing high-quality input data for subsequent high-precision image segmentation and feature extraction. This image data is transmitted back in real-time to an external processing terminal via a wired or wireless link through the communication transmission module integrated within the acquisition device, or temporarily stored in local storage for subsequent batch export.

[0028] In one implementation, the original digital images are captured by an in-situ image acquisition module integrated at the front of the first layer of the composite cabin. This module includes a high-pixel lens device and a ring light illumination device, which can provide uniform supplementary lighting under different water and land lighting conditions. The acquired image data is transmitted back to the terminal via a data transmission module in the second layer of the cabin.

[0029] This step, by acquiring high-quality, high-fidelity raw digital images, provides a reliable input data source for subsequent crack segmentation and classification, ensuring the information integrity and analytical accuracy of the entire intelligent classification method. It is a prerequisite for ensuring the effective operation of subsequent algorithms.

[0030] S2, the original digital image is input into a preset deep neural network model, frequency domain decomposition is used to preserve high-frequency edge features and spatial attention mechanism is combined to suppress background noise, and a two-dimensional binarized mask image of the crack is output.

[0031] Specifically, the original digital image is input into a pre-set deep neural network model. This model adopts an encoder-decoder architecture, the core of which is to achieve high-precision pixel-level segmentation of crack regions in the image through the synergistic effect of frequency domain decomposition and spatial attention mechanism.

[0032] Specifically, in the feature downsampling stage of the encoder, a discrete wavelet transform module is used to decompose the input feature map in the frequency domain, transforming it from the spatial domain to the frequency domain. This separates the low-frequency approximate components representing the macroscopic background from multiple high-frequency components containing high-frequency details such as microcrack edges and sharp corner directions. This process reduces the spatial resolution of the feature map while avoiding the irreversible loss of high-frequency information caused by traditional pooling operations, achieving lossless transfer of crack edge texture features. Subsequently, the decomposed frequency components are concatenated and stitched along the channel dimension, and feature fusion is performed through cross-channel information interaction to form a more information-rich concatenated feature tensor. In the skip connection path of the decoder, a spatial attention gating module is embedded. This module uses deep semantic features from the decoder as a gating signal to recalibrate the shallow features from the encoder in the spatial dimension, generating an attention weight matrix. This adaptively enhances the response of the crack region and suppresses interference from non-target background noise such as rock texture and underwater suspended matter. Finally, the decoder progressively upsamples the recalibrated features to the original image resolution and outputs a two-dimensional binary mask image representing the probability of the crack region through an activation function.

[0033] As one implementation, the discrete wavelet transform can be orthogonally decomposed using a Haar wavelet basis, outputting a low-frequency approximate component LL, a horizontal high-frequency component LH, a vertical high-frequency component HL, and a diagonal high-frequency component HH, and expressed by the formula... Perform full-scale cascading splicing.

[0034] This technique, by organically combining frequency domain decomposition and spatial attention mechanisms, significantly improves the ability and segmentation accuracy of deep neural networks to identify micro-cracks in complex geological images. Its beneficial effects are as follows: on the one hand, frequency domain decomposition preserves the high-frequency edge details of cracks that are easily lost in the traditional downsampling process, ensuring the topological continuity of the segmentation results; on the other hand, the spatial attention mechanism effectively suppresses the interference of complex background noise, improves the purity and accuracy of the segmentation mask, and provides a high-quality data foundation for subsequent refined feature extraction and causal classification of cracks.

[0035] S3, the two-dimensional binary mask image is skeletonized to construct a topological central axis network, the topological key nodes are extracted based on the connection degree of the skeleton pixels, and spatially independent crack samples are separated.

[0036] Specifically, the mask image undergoes skeletonization processing to transform the binarized crack region representation into a single-pixel-width topological central axis network. This eliminates width redundancy while preserving the spatial connectivity and geometric orientation of the crack, providing a foundation for subsequent topological structure analysis and independent sample separation. This process employs an iterative thinning algorithm to peel away the edge pixels of the crack region layer by layer until a single-pixel-width skeleton located at the crack's centerline is obtained. This skeleton network completely maps the topological connectivity and spatial extension path of the original crack region.

[0037] Building upon this foundation, topological attribute analysis is performed on each pixel in the skeleton network using the calculation of the connectivity degree of neighboring pixels. By counting the number of connected neighbors for each skeleton pixel within a preset neighborhood window, nodes with specific topological significance can be identified, including endpoints, bifurcation points, and intersection points. These nodes constitute the key elements describing the topology of the crack network. Subsequently, based on the connectivity between the extracted key topological nodes and the skeleton network, a connected component analysis algorithm is used to segment the entire skeleton network into multiple spatially independent and unconnected crack sample objects. Each sample object corresponds to an independent crack or an isolated crack segment.

[0038] As one implementation method, the connection degree of the skeleton pixels can be calculated using an eight-neighbor convolution kernel, and the node type can be determined according to the degree threshold. For example, a pixel with a degree of 1 is marked as an endpoint, a pixel with a degree of 3 is marked as a bifurcation point, and a pixel with a degree of 4 is marked as a cross-cutting point. Then, a breadth-first search algorithm is used to traverse the skeleton network and divide the network into independent connected components according to the connected paths between nodes, thereby separating each crack sample.

[0039] This step transforms the binarized mask into a topological skeleton network and extracts key nodes, thereby achieving an abstract and structured representation of the spatial morphology of cracks. This provides an accurate input basis for the subsequent construction of multidimensional feature vectors based on topological and geometric features, while ensuring the accurate separation of independent crack samples and avoiding misjudgments caused by crack adhesion or intersection, thus improving the robustness and accuracy of the classification method.

[0040] S4. For each fracture sample, calculate its global topological connectivity index, dimensionless aperture variation coefficient, strike angle, tortuosity, and cutting priority to construct a multidimensional feature space vector. Input the multidimensional feature space vector into the expert decision tree and determine the geological genesis type and evolution stage of each fracture sample according to the preset condition threshold. Output the visualized classification results.

[0041] Specifically, for each independent fracture sample obtained after skeletalization and spatial separation, this step aims to construct a multidimensional feature spatial vector that can comprehensively characterize its geomechanical and morphological properties, and to automatically determine the fracture origin type and evolution stage based on this vector.

[0042] Specifically, this multidimensional feature space vector consists of several quantitative indicators with clear physical meanings. These indicators characterize crack samples from multiple dimensions, including topological connectivity, non-uniformity of aperture spatial distribution, spatial distribution direction, trajectory curvature, and local intersection / cutting relationships. Specifically, the global topological connectivity index measures the distribution density and connection complexity of nodes (including endpoints, bifurcation points, and intersection / cutting points) within the crack network; the dimensionless aperture variation coefficient analyzes the statistical dispersion of aperture at each point on the crack skeleton line to reflect the roughness of the crack wall and the drastic change in aperture along the trajectory; the orientation angle characterizes the relative relationship between the main extension direction of the crack and a preset reference direction (such as regional bedding or principal stress direction); tortuosity quantifies the degree to which the crack trajectory deviates from an ideal straight line, thus reflecting the path deflection characteristics during its expansion; and cutting priority is based on local geometric morphology analysis at crack intersection nodes, inferring the cutting and truncation temporal relationship between cracks of different phases by evaluating the linear continuity of the skeleton in the intersection region. The aforementioned indicators collectively constitute a high-dimensional feature space vector, which can quantitatively describe the geometry and topology of fracture samples from multiple complementary perspectives. This multidimensional feature space vector is then input into a pre-constructed expert decision tree. This decision tree contains a series of conditional decision branches based on geomechanical principles and statistical experience. By comparing and logically judging each dimension component of the input vector with corresponding preset thresholds at each level, it automatically classifies each fracture sample into predefined geological genesis types and evolutionary stages, and finally outputs the classification results in a visual form.

[0043] As one implementation method, the global topological connectivity index can be calculated based on the number of endpoints, bifurcation points, and intersection cutting points in the skeleton network; the dimensionless aperture variation coefficient can be obtained by the ratio of the standard deviation to the mean of the aperture distribution sequence of the crack skeleton line; and the cutting priority can be determined by analyzing the principal component analysis results of the local skeleton of the intersection node.

[0044] This technical step involves constructing a multidimensional feature space vector that includes topological, geometric, and temporal relationships, and introducing an expert decision tree based on conditional rules for automated classification. This enables a leap from measuring a single geometric parameter to comprehensively characterizing multidimensional physical and mechanical properties of crack samples, significantly improving the objectivity and interpretability of determining the cause and evolution stage of cracks, and providing a more refined and reliable quantitative basis for geological engineering evaluation.

[0045] Furthermore, in step S2, the deep neural network model is based on an encoder-decoder architecture. In the feature downsampling path of the encoder, a discrete wavelet transform (DWT) module is used to decompose the feature map into low-frequency approximate components and multiple high-frequency detail components through a preset orthogonal filter. The decomposed components are then concatenated and stitched together in the channel dimension to achieve lossless transmission of high-frequency edge texture features.

[0046] Furthermore, a spatial attention gating module is embedded in the skip connection path between the encoder and decoder. The spatial attention matrix generated by the deep features is used to recalibrate the shallow features in order to adaptively suppress background noise.

[0047] Further, in step S4, the method for extracting the dimensionless aperture variation coefficient is as follows: perform distance transformation on the mask image to obtain the aperture distribution sequence on the crack skeleton line, calculate the ratio of the standard deviation to the mean of the sequence, and use it as a dimensionless index to measure the spatial heterogeneity of the crack.

[0048] Further, in step S4, the method for determining the cutting priority is as follows: when there are intersecting cutting points in the crack sample object, the local skeleton pixel coordinates around the intersecting cutting points are extracted for principal component analysis, and the ratio of the largest eigenvalue to the second largest eigenvalue of its local covariance matrix is ​​calculated as the local interleaving linearity. If the linearity is greater than a preset linearity threshold, it is determined that the current crack has not been interrupted in the intersection area and is set to the late generation level. If the linearity is less than or equal to the preset linearity threshold, it is determined that the current crack has been interrupted in the intersection area and is set to the early generation level.

[0049] Further, in step S4, the expert decision classification tree includes at least the following judgment rules: Artificial artifact filtering rule: when the tortuosity and dimensionless aperture coefficient of variation of the sample object are both lower than the first preset threshold, and the global topological connectivity index is zero, it is judged as a non-geological artificial artifact and removed. Shear fracture surface judgment rule: when the cutting priority is late generation level, and the dimensionless aperture coefficient of variation is lower than the second preset threshold, it is judged as a late-stage dominant high-penetration shear fracture surface. Multi-stage tension fracture judgment rule: when the cutting priority is early generation level, and the dimensionless aperture coefficient of variation is higher than the third preset threshold, it is judged as an early-evolutionary multi-stage staggered stress tension fracture. It also includes: multi-stage superimposed high-connectivity networks under the judgment of strong heterogeneous composite fracture networks; highly efficient expansion tectonic tension fractures belonging to high-angle development and fluid overpressure-driven hydraulic expansion; primary bedding-parallel detachment fractures of the earliest developed syn-sedimentary sequence, strictly limited by weak bedding planes; and diagenetic local contraction microfractures with extremely small scale and free and non-connected ends.

[0050] Example 2 like Figure 2 As shown, the present invention proposes a tracked and composite power amphibious in-situ crack acquisition device, comprising: a power module (1): a tracked walking chassis (11), wherein the chassis adopts a symmetrical double tracked structure, and four load wheels (12) are configured below each track mechanism on each side, for providing obstacle crossing and walking power on irregular land surfaces or in narrow holes.

[0051] The retractable dynamic positioning assembly is symmetrically embedded in the middle of the chassis and includes two sets of retractable propellers (13). The propellers (13) are stored inside the chassis in the land walking mode, and the retraction actuator (14) operates in the underwater operation mode, opening outward to provide underwater multi-degree-of-freedom positioning and waterway propulsion power.

[0052] Image in-situ acquisition and processing module (2): Equipped with a high-pixel lens device (21) and a ring light illumination device (22), it is integrated at the front end of the first-layer cabin (31) to provide real sample data under different water and land conditions. The returned data is transmitted to the second layer where the data transmission module (23) is concentrated for processing. The results will then be transmitted back to the terminal by the intelligent control module (24) to display the target crack situation.

[0053] The composite cabin (3) is fixedly installed above the chassis, wherein: The first-layer cabin (31) is a pressure-resistant and waterproof streamlined cabin, with an image in-situ acquisition and processing module (2) at its front end, used to acquire digital images of the surface of the land environment and water environment containing microcracks.

[0054] The second-layer cabin (32) is a communication transmission module cabin, which is fixedly arranged above the first-layer cabin. It is equipped with a communication relay unit for transmitting the digital images to an external control terminal in real time.

[0055] The device includes a propeller thruster (13) and a retraction / deployment mechanism (14). The retraction / deployment mechanism (14) is housed in a tracked chassis (11), with its tail end supported and fixed to the track near the hull. The first movable part from the inside out consists of a cube and a spherical balloon. The inner wall of the cube is coated with a special lubricant for protection and lubrication. The device contains a spherical balloon that can achieve nearly 360° position changes. The second movable part from the inside out consists of a column and movable support arms at both ends, forming an articulated extension arm with a range of motion between 180° and 230°. This provides a wide range of operability and addresses the angle issues encountered during sampling from multiple angles.

[0056] It is understood that the device of this invention adopts a tank-like dual-track mechanism with four road wheels as the chassis walking mechanism, and embeds two sets of retractable propellers in the middle. Combined with a double-layer composite cabin consisting of a lower streamlined pressure-resistant and waterproof cabin and an upper independent communication cabin, it improves the ability to overcome obstacles in irregular land terrain and maintain posture in complex wading flow fields. The external control terminal uses a deep network including discrete wavelet transform and spatial attention mechanism to perform edge-preserving denoising and segmentation on the images acquired and transmitted back by the front-end image module; by constructing a topological skeleton network, it extracts multi-dimensional feature vectors including global topological connectivity index, dimensionless openness variation coefficient, and local interpenetration linearity; and then uses an expert decision classification tree to realize the automated determination of the fracture formation period and the visualization color overlay output. This invention optimizes the quality of fracture data acquisition under complex amphibious geological conditions and is suitable for soil and rock mass stability analysis and quantitative evaluation of complex reservoirs.

[0057] Example 3 like Figures 2 to 5 As shown, the amphibious pore fissure in-situ acquisition device in this embodiment adopts a composite structure combining a mobile chassis and a double-layered upper cabin.

[0058] The walking mechanism adopts a tank-like dual-track structure, which is controlled by an intelligent control module (24). Four road wheels (12) are linearly arranged below each track mechanism. This structure provides high adhesion by increasing the ground contact area, which is used to achieve propulsion and obstacle crossing on soft gravel surfaces or in narrow vertical holes. The power positioning assembly is embedded on both sides of the counterweight axis in the middle of the chassis. This assembly includes two sets of four symmetrically arranged propellers (13) and a retraction and extension mechanism (14). In the land walking mode, the two sets of propellers (13) are horizontally stored in the embedded groove in the middle of the chassis, not exceeding the outer edge of the track. After entering the water wading working mode, the retraction and extension mechanism (14) extends the propellers (13) to both sides of the chassis.

[0059] A double-layer cabin supported by a lightweight, high-strength titanium alloy frame is erected above the chassis. The first layer cabin (31) is a sealed, pressure-resistant, and waterproof streamlined cabin with a low-resistance streamlined design. A high-pixel lens device (21) and a ring light illumination device (22) are embedded in the front end, forming an in-situ image acquisition and processing module (2), which is responsible for performing forward path shooting and image acquisition for underwater and land surfaces. The second layer cabin (32) is independently set on top of the first layer cabin (31). The interior is specially designed to house the image transmission repeater and high-frequency communication device integrated into the data transmission module (23). The intelligent control module (24) covers the motion control design of the device, ensuring the stability of the device when sampling in different environments. It is connected to the data processing unit of the first layer cabin (31) through a data bus to realize functional zoning and signal anti-interference.

[0060] like Figures 6 to 9The core design of the retractable propeller is shown. When the operating location is underwater, the four road wheels (12) are controlled to stop rotating and the control is transmitted to the tracked chassis (11) to stop, and then transforms into an auxiliary module to resist water flow disturbance. At this time, the retraction actuator (14) near the hull serves as the main power source. The retractable actuator (14) is retracted into the tracked chassis (11), and its tail end is supported and fixed to the side of the track near the hull. The first movable part from the inside out is composed of a cube and a balloon. The inner wall of the cube is coated with a special lubricant to protect and lubricate it. The device is embedded with a balloon that can achieve a near 360° position change. The second movable part from the inside out is composed of a column and movable support arms at both ends. The two form an articulated extension and movement arm with a range of motion between 180° and 230°, which can meet all angle problems encountered by the device when sampling. The outermost part is the propeller thruster (13). The support arms and connecting rods of the retractable propeller thruster are equipped with relevant electrical control lines. The base near the propeller thruster (13) is embedded with an electronically controlled servo motor to provide power. When working in the underwater area, the intelligent control module (24) uses the built-in control system to make the column joint type extendable mobile arm on the outer side act first, moving outward at a certain angle until the propeller thruster (13) is fully extended at the outer end of the tracked walking chassis (11). At this time, the balloon motion device close to the inner side can adjust the speed and yaw angle of the two propellers in real time according to the underwater conditions through the data of flow velocity detection and inertial navigation feedback built into the intelligent control module (24), and can manually control the movement at any angle, providing dynamic thrust to resist the disturbance of the shear flow field, and maintaining the stability of the device's posture when encountering the sampling target.

[0061] The device's dual-track walking mechanism enhances its land obstacle-crossing and adhesion performance; its retractable propeller, combined with a multi-degree-of-freedom adjustable structure, resists water flow interference and ensures stable underwater sampling posture. The double-layered streamlined cabin provides waterproofing and drag reduction capabilities, while its functional partition design minimizes electromagnetic interference and ensures normal image data transmission. The device is highly adaptable to both land and water conditions, and can stably complete in-situ data acquisition from geological fissures in various complex sites.

[0062] Example 4 like Figure 10This is a flowchart illustrating the overall process of data acquisition and classification interpretation provided in this embodiment of the invention. Specifically, the diagram shows the complete workflow. The entire device comprises three modules: a power module (1), an in-situ image acquisition and processing module (2), and a composite cabin (3), suitable for amphibious environment sampling. Both on land and underwater, the intelligent control module (24) controls the power drive. On land, it moves along tank-like tracks, while underwater, it is driven by a propeller (13) and a retraction / deployment mechanism (14). Under the intelligent control module (24), it can move and hover in the water for sampling. A high-pixel lens device (21) and a ring light illumination device (22) are integrated at the front of the first-layer cabin (31) for image sampling; this is the first stage, amphibious in-situ dynamic data sampling. The acquired photo samples are transmitted back to the external terminal for processing via the image transmission repeater and high-frequency communication device in the data transmission module (23). The external terminal acquires sample data, and after overlapping window cropping, the photo samples are cropped to the corresponding (512×512) pixel image format and enhanced using elastic deformation, while preserving as much of the original properties of the compared photo samples as possible. Subsequently, through inference using the DWT-Attn-UNet neural network, leveraging its differenceless concatenation and attention gate denoising properties, the system fuses the direct output characteristics of the recognition probability map and distance-weighted reconstruction to achieve skeletonized output of the two-dimensional binary mask image. After correction by neighborhood pixels, individual crack samples can be separated, and a topologically and geometrically joint feature vector is constructed. The two are then processed to obtain a decision classification; for each crack, the feature vector is processed and compared sequentially. Check if the current crack geometry parameters meet the tortuosity requirement. Coefficient of variation of opening degree And the density of intersection points If the conditions are met, the trajectory is determined to be a straight line segment without width distortion and spatial isolation, and the output label is "non-geological artificial artifact"; the priority of crack intersection is determined. and Is this condition met? If met, it indicates that the fracture is located at the latest geometrical dating point, and the fracture wall is straight and uniform due to strong shearing. The output label is "Late-stage dominant high-penetration shear fracture surface"; Judgment. and Is this condition met? If met, it indicates that the crack is topologically truncated by a later-stage crack and undergoes strong heterogeneity in aperture due to later-stage heterogeneous mineral cementation. The output label is "early-evolutionary multi-stage interlaced stress-tension crack." (Judgment) 0.35 or the number of boundary intersections or trace Check if the condition of "greater than half the image width" holds true. If either condition is true, trigger the determination of a strongly heterogeneous composite seam network, and output the label as a multi-phase superimposed highly connected network; determine the direction angle. °、 and Is it valid? If valid, it indicates that it is a high-angle development, fluid overpressure-driven hydraulic expansion fracture, and the output label is high-efficiency expansion structure tension fracture; [Judgment] °、 and Is this condition valid? If valid, it indicates that the fracture is strictly confined by weak bedding planes and is the earliest developed fracture, with the output label being a primary bedding-parallel detachment fracture in a syn-sedimentary sedimentary sequence; determine the trace length. Pixels and Is this condition met? If met, it indicates that the fracture possesses extremely small-scale and double-ended, unconnected topological features, and the output label is "locally contracted microfracture from the diagenetic period." At this point, the classification and identification of different fractures are complete. Finally, through the overlay of a high-contrast color layer, a visual representation and detailed classification analysis are achieved, such as... Figure 14 As described in Table 1.

[0063] Table 1

[0064] This embodiment preserves the original image features completely through image cropping and data augmentation preprocessing. Relying on a neural network integrating discrete wavelet transform and attention mechanisms, it effectively filters background noise, preserves crack edge details, accurately generates crack masks, and extracts independent crack samples. The system constructs feature vectors by integrating topological, geometric, and temporal multi-dimensional features, and, combined with hierarchical judgment rules, can accurately distinguish between artificial artifacts and various geological cracks of different causes and evolutionary stages. The system offers comprehensive recognition dimensions and rigorous judgment logic. Finally, the results are visualized through color overlay. The entire process is highly automated, with strong recognition accuracy and classification reliability, meeting the practical application needs of refined crack exploration and source tracing analysis in complex geological scenarios.

[0065] Example 5 After processing the image received by the terminal, step S2 inputs the original digital image into a preset deep neural network model. Frequency domain decomposition is used to preserve high-frequency edge features, and a spatial attention mechanism is combined to suppress background noise, outputting a two-dimensional binarized mask image of the crack. Specifically, an encoder-decoder network architecture incorporating Discrete Wavelet Transform (DWT) and spatial attention gating is employed, with the following steps: Specifically, in the feature downsampling stage of the encoder, the input feature map is orthogonally decomposed using a Haar wavelet basis, outputting a low-frequency approximation component (LL) and three high-frequency detail components (including horizontal high frequency LH, vertical high frequency HL, and diagonal high frequency HH). These four components are then concatenated and fused along the channel dimension and passed to the next layer. This process reduces the spatial resolution of the feature map while explicitly preserving high-frequency detail information representing the edges of minute cracks. Spatial attention gating is introduced at the skip connections of the decoder, using deep semantic information to generate a weight matrix and recalibrate shallow features. This adaptively reduces the weight of non-crack background noise such as rock bedding and underwater suspended matter, improving the continuity of crack binarization segmentation.

[0066] like Figure 11 The diagram shown is a pre-set deep neural network architecture topology diagram in this embodiment of the invention. Specifically, it involves encoding and decoding in image processing. After in-situ sampling by the high-pixel lens device (21) and the ring light illumination device (22) in the in-situ image acquisition and processing module (2), the image is transmitted to the terminal by the data transmission module (23). The input image is processed into three-channel feature dimension data with a pixel size of 512×512×3. It is input from the upper left corner and undergoes four cascaded convolution extractions and wavelet downsampling. The spatial resolution is halved each time, and the number of channels is doubled. The original color image is input. After the first Enc_block double convolution, the spatial resolution is preserved. The number of channels remains unchanged, but it is mapped to 64 channels through multi-kernel upscaling. The output feature map is now named the shallow feature map. It contains the richest high-frequency geometric details, such as crack edges, but also carries the most noise. Shallow features Two output directions are generated: one is pulled out laterally and sent as the source signal for a jump connection to the AG1 gate on the right; the other enters the DWT downsampling module; and so on, resulting in four shallow features. Through two-dimensional Haar wavelet decomposition and lossless concatenation, the spatial resolution is halved to... The number of channels remains at 64. The data stream flows sequentially through each level of double convolution and DWT downsampling. After each downsampling, the information becomes more spatially compact and richer in terms of channels. Level 2 output features: Level 3 output characteristics: Level 4 output characteristics: Characteristics before reaching the bottom: The encoding process is now complete, and we move on to the decoder operation. The decoder feature map below undergoes transposed convolution upsampling, increasing its spatial resolution from... Enlarged The number of channels is adjusted accordingly. This tensor, which has just been amplified and contains deep semantic information, is extracted as an upsampling gating signal. Shallow features from the left encoder It contains abundant crack edges, but also contains a significant amount of mechanical scratches and noise. The AG1 module simultaneously receives shallow features. and gating signals A two-dimensional attention weight matrix is ​​generated through internal matrix dot product and activation. This matrix is ​​then multiplied by... The weights of noisy regions are forcibly reduced to 0, while the weights of genuine crack regions are kept at 1, resulting in the output recalibrated features. The purified shallow features after AG1 filtration (512×512×64), and the decoded feature tensor rising vertically from the right, are concatenated along the channel axis. This series of convolutions in this invention allows the network to simultaneously possess a filtered, finely detailed fracture skeleton. The decoder is constructed sequentially, and the concatenated high-dimensional dense tensor undergoes channel fusion and feature smoothing in the final group, ultimately... Convolution and the Sigmoid activation function compress 64 high-dimensional channels into 1 channel, outputting a 512×512×1 pixel-level binary semantic segmentation probability mask.

[0067] like Figure 12 This is a schematic diagram of the feature decomposition and lossless cascading principle of the Discrete Wavelet Transform (DWT) module in this embodiment of the invention. Specifically, after in-situ sampling by the high-pixel lens device (21) and the ring light illumination device (22) in the in-situ image acquisition and processing module (2), the image is transmitted to the terminal by the data transmission module (23), and the image is encoded and input into the three-channel feature map. First, an orthogonal filtering matrix based on a two-dimensional Haar wavelet basis is used. The core formula of the Haar wavelet is based on adding (low-pass) and subtracting (high-pass) adjacent pixels. The orthogonal filtering matrix performs cross product operations of low-pass and high-pass basis functions in the horizontal and vertical directions, transforming the input spatial domain features into four independent frequency domain subbands with half the spatial resolution and C channels. , , A complete image is divided into four images with dimensions of... The small image shows that no information is lost during this process. The formula is as follows:

[0068] in, These are the coordinates of two adjacent pixels. Each small image corresponds to a sub-band with a different frequency, but the dimensions are the same.

[0069] Subsequently, to avoid the high-frequency edge energy dissipation caused by the traditional max-pooling operator discarding non-maximum responses, this invention introduces a lossless cascading mechanism. This involves full stacking and using the Concat concatenation operation to combine high-frequency components containing micro-crack edges, sharp corner directions, etc. , , Low-frequency components representing the macroscopic matrix background Perform full splicing along the channel axis to obtain the cascaded feature tensor. The formula is as follows:

[0070] Finally, a 1×1 convolutional layer is used for cross-channel information interaction and dynamic weighting to compress the number of channels to a specified value. Dimension, Output Thus, scale reduction was achieved without losing any low-contrast microcrack topological continuity features.

[0071] This embodiment employs an encoder-decoder network that integrates discrete wavelet transform and spatial attention gating to process images. Haar wavelets are used to perform feature frequency domain decomposition, fully preserving high-frequency details such as crack edges and avoiding information loss caused by traditional downsampling. Channel stitching and convolution optimization achieve lossless feature fusion, ensuring the continuity and integrity of micro-crack features. Simultaneously, an attention mechanism is used to recalibrate shallow features, effectively suppressing background interference such as rock textures and underwater suspended matter. The entire network can accurately perform pixel-level segmentation and output high-quality crack binarization masks, significantly improving the accuracy and segmentation effect of micro-crack recognition in complex water and land scenes.

[0072] Example 6 In step S4, for each crack sample, its global topological connectivity index, dimensionless aperture variation coefficient, strike angle, tortuosity, and cutting priority are calculated to construct a multidimensional feature space vector. The feature space vector... The extraction and mathematical calculation process is as follows: Global topological connectivity index ( ): Endpoints calculated based on the pixel degree of a 3×3 neighborhood in the skeleton network. (Degree is 1), bifurcation point (degree of 3), intersection cutting point The quantity (of degree 4) is calculated and obtained as follows:

[0073] Dimensionless aperture variation coefficient (CV): The true aperture distribution sequence along the crack skeleton line is obtained through Euclidean distance transformation, and its standard deviation is calculated. with the mean .definition (in To prevent the use of a small correction constant with a denominator of zero, a dimensionless measure of the non-uniformity of the opening space is achieved.

[0074] Cutting priority When a crack intersects at a cutting point X, principal component analysis (PCA) is performed on the pixel coordinates of a local skeleton window centered on that node. The ratio of the first eigenvalue to the second eigenvalue of its covariance matrix is ​​calculated as the local interpenetration linearity. If this linearity is greater than a preset linearity threshold, it indicates that the skeleton is continuous and unbroken in that region, and a threshold is set. (Indicates late generation); otherwise, it is set. (Indicates an early generation that was cut off by a later crack). If there are no intersecting nodes, the default value is 0.

[0075] Heading towards the included angle ( ) and tortuosity ( PCA is used to extract the main orientation of individual cracks and calculate the angle between them and the reference bedding; the length of the straight trace is evaluated using the longer side of the minimum bounding rectangle, and the tortuosity is calculated as the ratio of the actual skeleton pixel length to the length L of the straight trace. .

[0076] This embodiment extracts fracture features from multiple dimensions, including topological morphology, aperture distribution, spatial orientation, tortuosity, and cutting sequence, constructing a multi-dimensional feature vector. Utilizing pixel neighborhood statistics, distance transformation, and principal component analysis, it quantifies indicators such as connectivity characteristics, aperture variation, strike angle, tortuosity, and cutting priority, achieving a digital representation of fracture morphology and development sequence. These features complement each other, comprehensively reflecting the geometric morphology, spatial distribution, and evolutionary relationships of fractures, avoiding the limitations of single-feature judgment. The feature calculation method is stable and reliable, accurately distinguishing different types and formation stages of geological fractures, providing sufficient and effective data support for subsequent intelligent classification and judgment.

[0077] Example 7 In step S4, the multidimensional feature space vector is input into the expert decision tree. Based on preset threshold conditions, the geological genesis and evolutionary stage of each fracture sample are determined, and a visual classification result is output. The constructed classification decision tree executes the judgment logic sequentially. To facilitate understanding of the implementation process of this technical solution, an empirical threshold decision-making process under a specific experimental scenario is used as an example. Table 2 shows the optimal implementation values ​​for various feature thresholds after statistical optimization of multiple unconventional oil and gas reservoir samples. Table 2

[0078] First, the system executes manual artifact filtering rules. If the tortuosity of the identified object... Dimensionless aperture variation coefficient And the global topological connectivity index If the object is determined to be an artificial artifact of non-geological origin (such as mechanical scratches caused by track movement or sampling section knife marks), it will be removed.

[0079] Then we proceeded to compare the causes and periods: When the condition is met and When the fracture has good penetration and low spatial non-uniformity, it is determined to be a late-stage dominant high-penetration shear fracture surface. When the condition is met and At that time, it was determined to be an early-stage, multi-stage, interlaced stress tension crack; When approaching the angle and At that time, it was determined to be a high-efficiency expansion tectonic tension crack.

[0080] Based on the above classification results, the processing terminal maps a preset high-contrast color layer onto the original image background according to the corresponding geological category, and performs visualization overlay and annotation output of the result image.

[0081] Furthermore, such as Figure 13 The flowchart for constructing a multidimensional topological-geometric feature-based decision classification tree in this embodiment of the invention is shown below. Specifically, the calculation of each feature vector is as follows: Global topological connectivity index ( ): Endpoints calculated based on the pixel degree of a 3×3 neighborhood in the skeleton network. (Degree is 1), bifurcation point (degree of 3), intersection cutting point The quantity (of degree 4) is calculated and obtained as follows:

[0082] Dimensionless aperture variation coefficient (CV): The true aperture distribution sequence along the crack skeleton line is obtained through Euclidean distance transformation, and its standard deviation is calculated. with the mean .definition (in To prevent the use of a small correction constant with a denominator of zero, a dimensionless measure of the non-uniformity of the opening space is achieved.

[0083] Cutting priority When a crack intersects at a cutting point X, principal component analysis (PCA) is performed on the pixel coordinates of a local skeleton window centered on that node. The ratio of the first eigenvalue to the second eigenvalue of its covariance matrix is ​​calculated as the local interpenetration linearity. If this linearity is greater than a preset linearity threshold, it indicates that the skeleton is continuous and unbroken in that region, and a threshold is set. (Indicates late generation); otherwise, it is set. (Indicates an early generation that was cut off by a later crack). If there are no intersecting nodes, the default value is 0.

[0084] Heading towards the included angle ( ) and tortuosity ( PCA is used to extract the main orientation of individual cracks and calculate the angle between them and the reference bedding; the length of the straight trace is evaluated using the longer side of the minimum bounding rectangle, and the tortuosity is calculated as the ratio of the actual skeleton pixel length to the length L of the straight trace. .

[0085] Steps S2 and S3 output a two-dimensional binary mask image representing the crack region and independent crack samples after skeletonization. The skeletonized mask image can calculate various feature data of the crack in terms of multi-dimensional topology and geometry. Each calculated mask image will be used as raw data in this feature judgment for calculation. Proceed to the judgment flowchart: Step 1 (S1 node): Check whether the current crack geometry parameters meet the tortuosity requirement. Coefficient of variation of opening degree And the density of intersection points If the conditions are met, the trajectory is determined to be a straight line segment without width distortion and spatial isolation, and the output label is "[Non-geological artificial artifact]" and it is removed; if the conditions are not met, the next level of geological origin screening is performed.

[0086] Step 2 (S2 node): Determine the crack intersection priority and Is this condition met? If met, it indicates that the fracture is located in the latest geometrical dating, and the fracture wall is straight and uniform due to strong shearing. The output label is "[Late-stage dominant high-penetration shear fracture surface]"; otherwise, proceed to the next level.

[0087] Step 3 (S3 node): Determination and Is this condition met? If met, it indicates that the crack is topologically truncated by a later-stage crack and its aperture is highly heterogeneous due to later heterogeneous mineral cementation. The output label is "[Early Evolution Multi-Stage Interlaced Stress Tension Crack]". Otherwise, proceed to the next level.

[0088] Step 4 (S4 node): Determine 0.35 or the number of boundary intersections or trace Check if the condition of being greater than half the image width holds true. If either condition is true, trigger the strong non-homogeneous composite seam network determination and output the label "[Multi-stage superimposed highly connected network]"; otherwise, proceed to the next level.

[0089] Step 5 (S5 node): Determine the direction angle °、 and Is it valid? If valid, it indicates that it is a high-angle development, fluid overpressure-driven hydraulic expansion fracture, and the output label is "High-efficiency expansion tectonic tension fracture"; otherwise, proceed to the next level.

[0090] Step Six (S6 Node): Determination °、 and Is this condition met? If it is, it indicates that the fracture is strictly confined by weak bedding planes and developed earliest, and the output label is "[Primary bedding-parallel detachment fracture in syn-sedimentary sequence]"; otherwise, proceed to the next level.

[0091] Step 7 (S7 node): Determine the trace length Pixels and Is it true? If true, it indicates that it has a topological feature of extremely small scale and two-way free disconnection, and the output label is "[Diagenetic local shrinkage microfracture]"; otherwise, the output label is "[Undefined transitional geological microstructure]".

[0092] This embodiment of the decision tree relies on multi-dimensional features to set hierarchical judgment rules. It first accurately filters out artificial artifacts such as mechanical scratches, and then sequentially distinguishes various geological fractures of different causes and evolutionary stages. The rule logic hierarchy is clear, and the judgment criteria are well-defined, effectively identifying multiple types such as shear fracture surfaces, tension fractures, primary fractures, and microfractures, while also distinguishing multi-stage superimposed fracture networks. Finally, color-coding is used to visualize the results. The classification accuracy is high, and the degree of automation is strong, enabling refined fracture interpretation and intuitive display, meeting the needs of geological exploration and analysis.

[0093] 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.

[0094] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0095] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. An intelligent classification method based on an amphibious fissure in-situ acquisition device, characterized in that, include: S1, acquire the original digital image containing the geological structure surface transmitted back by the amphibious fissure in-situ acquisition device; S2, input the original digital image into a preset deep neural network model, use frequency domain decomposition to retain high-frequency edge features and combine spatial attention mechanism to suppress background noise, and output a two-dimensional binarized mask image of the crack. S3, skeletonize the two-dimensional binary mask image to construct a topological central axis network, extract key topological nodes based on the connectivity degree of the skeleton pixels, and separate spatially independent crack samples. S4. For each fracture sample, calculate its global topological connectivity index, dimensionless aperture variation coefficient, strike angle, tortuosity, and cutting priority to construct a multidimensional feature space vector. Input the multidimensional feature space vector into the expert decision tree and determine the geological genesis type and evolution stage of each fracture sample according to the preset condition threshold. Output the visual classification results. The S2 includes: In the feature downsampling path of the encoder, the discrete wavelet transform module is used to decompose the input feature map into a low-frequency approximation component and three high-frequency detail components through a preset orthogonal filter. The low-frequency approximation component and the three high-frequency detail components are concatenated and stitched together in the channel dimension to achieve lossless transmission of high-frequency edge texture features. In the skip connection path between the encoder and decoder, a spatial attention gating module is embedded. The spatial attention weight matrix generated by the deep semantic features is used to recalibrate the shallow features in order to adaptively suppress background noise and output a two-dimensional binarized mask image. The expert decision tree includes the following judgment rules: Artificial artifact filtering rules: When the tortuosity and dimensionless aperture coefficient of variation of the sample object are both lower than the first preset threshold, and the global topological connectivity index is zero, it is judged as a non-geological artificial artifact and removed. Shear fracture surface determination rule: When the cutting priority is late generation level and the dimensionless aperture variation coefficient is lower than the second preset threshold, it is determined to be a late master-controlled high penetration shear fracture surface. Multi-stage tension crack determination rules: When the cutting priority is at the early generation level and the dimensionless aperture variation coefficient is higher than the third preset threshold, it is determined to be an early-evolution multi-stage staggered stress tension crack; it also includes: Multi-stage superimposed highly connected networks under the strong heterogeneous composite fracture network criteria; highly efficient expansion tectonic tension fractures belonging to high-angle development and fluid overpressure-driven hydraulic expansion; primary bedding-parallel detachment fractures of the earliest developed syn-sedimentary sequence, strictly limited by weak bedding planes; and diagenetic local contraction microfractures with extremely small scale and free, non-connected ends.

2. The method according to claim 1, characterized in that, The discrete wavelet transform module decomposes the input feature map into one low-frequency approximation component and three high-frequency detail components through a preset orthogonal filter, including: The orthogonal filtering matrix of the Haar wavelet basis is used to perform cross product operations of low-pass and high-pass basis functions in the horizontal and vertical directions, which transforms the input spatial domain feature map into four frequency domain independent sub-bands with half the spatial resolution. The four frequency domain independent sub-bands include low-frequency approximate components, horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components.

3. The method according to claim 1, characterized in that, The step of concatenating and splicing the low-frequency approximation component with the three high-frequency detail components in the channel dimension includes: The low-frequency approximation component and the three high-frequency detail components are fully concatenated along the channel axis by the Concat concatenation operation to obtain the concatenated feature tensor. A convolutional layer with a kernel size of 1×1 is used to perform cross-channel information interaction and dynamic weighting on the cascaded feature tensor, compressing the number of channels to a specified 2C dimension, and outputting the processed feature data.

4. The method according to claim 1, characterized in that, The S3 includes: A skeletonization algorithm is performed on the two-dimensional binary mask image to obtain a topological skeleton network with a single pixel width; The connection degree of each skeleton pixel in the topological skeleton network is calculated using 3×3 neighborhood convolution, and the endpoints with a degree of 1, the bifurcation points with a degree of 3, and the cross-cutting points with a degree of 4 are extracted as topological key nodes. Spatially independent crack samples were separated based on the key nodes of the topology using a connected component algorithm.

5. The method according to claim 1, characterized in that, The global topology connectivity index is calculated as follows: the total number of crack bifurcation points and intersection points is counted, and then the total number of all crack endpoints, bifurcation points and intersection points is combined to calculate the global topology connectivity index. The dimensionless aperture variation coefficient is extracted by using Euclidean distance transformation technology to extract the aperture data sequence of the entire crack skeleton, calculating the fluctuation range and average value of the aperture data, and thus obtaining the corresponding coefficient; at the same time, a small correction value is added to avoid the situation where the denominator is empty during the calculation process.

6. The method according to claim 1, characterized in that, The method for determining the cutting priority is as follows: When there are intersecting points in the crack sample, the pixel coordinates of the local skeleton window centered on the intersecting points are extracted for principal component analysis, and the ratio of the first eigenvalue to the second eigenvalue of its covariance matrix is ​​calculated as the local interpenetration linearity. If the local interleaving linearity is greater than a preset linearity threshold, then a cutting priority is set. , indicating late-stage generation; If the local interleaving linearity is less than or equal to a preset linearity threshold, then a cutting priority is set. This indicates that the node was generated early and cut off by a later crack; if there are no intersecting nodes, a cutting priority is set. .

7. The method according to claim 1, characterized in that, The method for extracting the orientation angle is as follows: principal component analysis is used to extract the main orientation of the individual fracture, and the angle between the main orientation and the reference bedding is calculated. The method for extracting the tortuosity is as follows: the length of the straight trace is evaluated using the long side of the smallest bounding rectangle, and the tortuosity is the ratio of the actual skeleton pixel length to the length of the straight trace.

8. An amphibious in-situ fissure sampling device, using the method described in claim 1, characterized in that, include: The system comprises a power module, an in-situ image acquisition and processing module, and a composite cabin, among which: Power module: Tracked chassis with symmetrical track mechanisms on both sides of the chassis, and four road wheels under each track to provide power for walking on irregular geological surfaces or in narrow holes. The retractable dynamic positioning assembly is symmetrically embedded in the central area of ​​the chassis and includes two sets of retractable propellers. The retraction mechanism controls the opening or closing of the propellers, corresponding to the underwater working mode and the land working mode, respectively. Image in-situ acquisition and processing module: Equipped with a high-pixel lens device and a ring light illumination device, it is integrated at the front of the first-layer cabin to provide real sample data under different water and land conditions. The returned data is temporarily stored in the data transmission module in the second layer, and then the image data is transmitted to the terminal for visualization and analysis. The on-board intelligent control module controls the movement and sampling of the control device under different water and land conditions. Composite Cabin: A double-layer composite cabin, fixedly mounted on top of the chassis. The first layer is a pressure-resistant and waterproof streamlined cabin, with an in-situ image acquisition and processing module facing the direction of travel at its front end. The second layer is stacked on top of the first layer and has pressure-resistant and waterproof functions. It integrates a data transmission module and an intelligent control module.

Images