Fractured rock slope seepage channel processing method and system based on deep learning

Abstract

The application discloses a kind of fissure rock slope seepage channel processing method and system based on deep learning, it is related to the application technical field of neural network model, the method includes: based on real-time multi-source data, utilize deep learning model to identify seepage channel and generate fissure network probability set;Introduce reinforcement learning model, according to the fissure network probability set automatically generates optimal drilling strategy;In the drilling phase, fissure observation data are collected in real time, and the fissure network probability set is corrected by Bayes formula, to dynamically adjust drilling strategy;In grouting stage, using physical information neural network, combine real-time grouting response data with fissure network probability set, carry out fissure expansion prediction and grouting parameter optimization, complete the accurate grouting treatment of seepage channel.The application can utilize various neural network models to solve the technical problems that static seepage channel treatment strategy fails due to dynamic changes of fissure network caused by drilling construction and grouting process in the prior art.

Description

Technical Field

[0001] This invention relates to the field of neural network model application technology, and in particular to a method and system for treating seepage channels in fractured rock slopes based on deep learning. Background Technology

[0002] The seepage channels in fractured rock slopes are a key factor affecting slope stability, and their development and connectivity directly relate to the risk of geological disasters such as landslides and debris flows. In engineering fields such as water conservancy and hydropower, transportation infrastructure, and mining, rock slopes often contain complex fracture networks. These fractures not only constitute the dominant seepage paths for groundwater but may also expand further under external disturbances, leading to the deterioration of the slope's rock mass mechanical properties and ultimately causing instability and failure. Therefore, accurately identifying and effectively treating seepage channels in fractured rock slopes is of significant practical importance for ensuring engineering safety and preventing geological disasters.

[0003] Existing methods for treating seepage channels on fractured rock slopes primarily focus on predicting a three-dimensional fracture network through multi-source data fusion before treatment to accurately locate seepage channels and formulate construction strategies based on the initial prediction results. However, these methods often overlook the fact that drilling operations and grouting pressures during actual construction can alter the original morphology of the fracture network, thus affecting the effectiveness of the construction strategy and often resulting in unsatisfactory final treatment outcomes. Specifically, existing methods typically employ an open-loop model of "pre-modeling and post-control," lacking the ability to respond in real-time to the dynamic evolution of fractures under construction disturbances and making it difficult to adaptively adjust construction parameters in uncertain environments, thereby limiting the effectiveness and reliability of seepage channel treatment. Summary of the Invention

[0004] This invention provides a method and system for treating seepage channels on fractured rock slopes based on deep learning. It can solve the technical problem that the static construction strategy fails in existing methods for treating seepage channels on fractured rock slopes because the dynamic changes in the fracture network caused by drilling and grouting processes are ignored.

[0005] In a first aspect, embodiments of the present invention provide a method for treating seepage channels in fractured rock slopes based on deep learning, comprising:

[0006] Based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, seepage channels are identified, and combined with pre-stored fracture attitude statistics, a first fracture network probability set is generated; wherein, the first fracture network probability set includes several fracture networks and the realization probability of each fracture network. Based on the first set of probability of the fracture network, with the goal of maximizing information gain and maximizing governance potential, a drilling strategy is generated through deep reinforcement learning; wherein, the drilling strategy includes drilling sequence, drilling location and drilling parameters; During the drilling process, the drilling equipment acquires fracture observation data in real time. Based on the fracture observation data, the pre-acquired second fracture network probability set is dynamically corrected using a Bayesian update method. The drilling strategy is then adjusted in real time based on the corrected second fracture network probability set until the seepage channel drilling of the target rock slope is completed. The second fracture network probability set is obtained by predicting fracture propagation based on the first fracture network probability set using the currently implemented drilling strategy. During the grouting process, the grouting response data is monitored in real time. Combined with the pre-acquired third fracture network probability set, the fracture propagation is predicted through a pre-trained first physical information neural network to dynamically adjust the grouting parameters until the grouting treatment of the seepage channel of the target rock slope is completed. The third fracture network probability set is the most recently corrected second fracture network probability set.

[0007] This invention constructs an initial cognitive model by generating a first set of fracture network probabilities, quantifying uncertainty in the form of a probability set to provide a correctable cognitive foundation for subsequent dynamic updates. It generates a drilling strategy through deep reinforcement learning, formulating an optimal drilling plan based on the initial cognition, providing initial construction guidance, and reserving space for subsequent adjustments. By acquiring fracture observation data in real time during drilling, it perceives the actual rock mass information revealed by the drilling construction, thereby introducing real-time feedback, breaking the limitations of "static prediction" in existing technologies, and dynamically correcting the first fracture network probabilities through Bayesian updates based on the observation data. The invention employs a second-order fracture network probability set to enable real-time evolution of the model's understanding during construction, ensuring that construction strategies do not fail due to changes in fracture morphology caused by drilling. Real-time monitoring of grouting response data during grouting allows for the perception of rock mass response, providing real-time data for grouting control. Furthermore, a third-order fracture network probability set is used to predict fracture propagation and potential fracture changes induced by grouting pressure, thus considering grouting disturbances and dynamically adjusting grouting parameters to avoid secondary disasters such as hydraulic fracturing. This ensures that construction strategies do not fail due to changes in fracture morphology caused by grouting pressure. Compared to existing technologies, this invention achieves dynamic and precise treatment of seepage channels in fractured rock slopes, effectively solving the technical problem of static treatment strategies failing due to dynamic rock mass evolution caused by drilling and grouting disturbances.

[0008] In some preferred embodiments of the first aspect, based on the fracture observation data, the pre-acquired second fracture network probability set is dynamically corrected using a Bayesian update method, so as to adjust the drilling strategy in real time according to the corrected second fracture network probability set, specifically as follows: Based on the drilling strategy currently implemented by the drilling equipment, the first fracture network probability set is predicted to expand by a pre-trained second physical information neural network model, thereby obtaining the second fracture network probability set. Based on the fracture observation data, the fracture features of the currently observed fracture are extracted, and based on the fracture features, the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture is calculated; wherein, the fracture features include fracture depth, fracture orientation, and fracture aperture. Based on the likelihood, the realization probability of each fracture network in the second fracture network probability set is updated by a preset Bayesian formula to obtain the corrected second fracture network probability set. Based on the modified second fracture network probability set, the expected information gain and governance potential of the un-drilled holes in the drilling strategy are evaluated respectively. The drilling strategy is then adjusted in real time by combining the expected information gain evaluation results and governance potential evaluation results of each un-drilled hole.

[0009] This invention provides a second fracture network by predicting fracture propagation based on the current drilling strategy, ensuring that the model's understanding before updating the drilling strategy already includes borehole perturbations. It extracts currently observed fracture features to provide a quantitative basis for model comparison, and calculates the likelihood of each fracture network relative to the observed fractures to quantitatively assess the fit between each possible model and the actual situation, providing a numerical basis for Bayesian updates. The Bayesian formula update realizes probability, increasing the weight of models with good fit and decreasing the weight of models with poor fit, achieving convergence of understanding to reality. By assessing the information gain and governance potential of undrilled boreholes, the value of each remaining borehole is re-evaluated, providing a decision-making basis for dynamic strategy adjustments.

[0010] In some preferred embodiments of the first aspect, based on the drilling strategy currently implemented by the drilling equipment, a second fracture network probability set is obtained by predicting fracture propagation on the first fracture network probability set using a pre-trained second physical information neural network model, specifically as follows: Based on the drilling strategy currently implemented by the drilling equipment, a drilling disturbance source function is constructed, and the drilling disturbance source function is determined as the boundary condition of the second physical information neural network model; By using a pre-constructed numerical inversion model, stress field inversion is performed on each fracture network in the first fracture network probability set to obtain the first stress field of each fracture network. The boundary conditions, the first stress fields of each fracture network, and the first fracture network probability set are input into a pre-trained second physical information neural network model so that the second physical information neural network model updates the first stress field according to the preset elasticity equilibrium equation and fracture mechanics criterion, thereby obtaining the second stress field of each fracture network in the first fracture network probability set. Based on the second stress field, crack propagation prediction is performed on each crack network in the first crack network probability set to obtain the second crack network probability set.

[0011] This invention, through the following embodiments, constructs a borehole disturbance source function as a boundary condition, enabling the prediction model to receive construction parameters; obtains the first stress field through a numerical inversion model to acquire the initial stress state of the fracture network, providing a benchmark for stress updates; simulates the network using a second physical information neural network embedded with mechanical equations, combining data-driven approaches with physical laws to ensure prediction reliability; and predicts fracture propagation based on the second stress field to obtain an updated network that considers borehole disturbances, providing a basis for subsequent borehole strategy adjustments.

[0012] In some preferred embodiments of the first aspect, based on the fracture observation data, fracture features of the currently observed fracture are extracted, and based on the fracture features, the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture is calculated, specifically as follows: Based on the fracture observation data, the fracture features of the currently observed fracture are extracted to obtain a list of actual features of the currently observed fracture; Based on the drilling strategy currently implemented by the drilling equipment, spatial geometric intersection calculations are performed in each fracture network of the second fracture network probability set to determine the fracture segments in each fracture network that intersect with the current drilling trajectory, and the fracture features of each fracture segment are extracted to obtain a list of predicted features for each fracture network. Based on the actual feature list and the predicted feature list of each fracture network, the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture is calculated.

[0013] This invention extracts the current observed fracture features to obtain an actual feature list, transforming borehole television images into structured fracture data and providing true values ​​for model comparison. Spatial geometric intersection is performed in each fracture network to calculate which fractures each network will "predict and see" under the current borehole parameters, providing predicted values ​​for model comparison. A predicted feature list is obtained by extracting features from each fracture segment, acquiring the predicted fracture depth, orientation, and aperture, providing quantitative predictions for model comparison. Likelihood is calculated based on actual and predicted features to quantitatively evaluate the degree of agreement between each network and the actual observations, thus providing a basis for Bayesian updates.

[0014] In some preferred embodiments of the first aspect, based on the modified second fracture network probability set, the expected information gain and governance potential assessments are performed on the un-drilled boreholes in the borehole strategy, specifically as follows: For each undrilled hole in the drilling strategy, based on the corresponding drilling location and drilling parameters, spatial geometric intersection calculation is performed in each fracture network in the modified fracture network probability set to determine the fracture segment in each fracture network that intersects with the preset trajectory of each undrilled hole, and the fracture features of each fracture segment are extracted to obtain a list of predicted fracture features for each undrilled hole. Based on the predicted fracture feature list and the corrected second fracture network probability set, the expected information gain for each un-drilled hole is calculated using Monte Carlo simulation to obtain the expected information gain evaluation results for each un-drilled hole. Based on the predicted fracture feature list, the realization probability of each fracture network in the modified second fracture network probability set, and the preset graph theory indices, the remediation potential of each un-drilled borehole is evaluated to obtain the remediation potential evaluation result corresponding to each un-drilled borehole; wherein, the graph theory indices include betweenness centrality indices and seepage centrality indices.

[0015] This invention provides foundational data for value assessment by performing spatial geometric intersection on each un-drilled borehole and calculating the fractures that each remaining borehole will traverse in different networks. It also obtains a list of predicted fracture features for each un-drilled borehole to acquire fracture information that each borehole may reveal, providing input for information gain calculation. Monte Carlo simulation is used to calculate the expected information gain, assessing how much cognitive uncertainty the borehole can reduce and quantifying its "exploration value." Furthermore, the invention evaluates the remediation potential based on graph theory indicators (betweenness centrality and flow centrality) to assess the contribution of boreholes in sealing key seepage channels. Finally, by integrating information gain and remediation potential, the strategy is adjusted by reordering, discarding, or optimizing the parameters of the remaining boreholes, achieving dynamic optimization of the strategy.

[0016] In some preferred embodiments of the first aspect, the real-time monitoring of grouting response data, combined with a pre-acquired third fracture network probability set, and the prediction of fracture propagation through a pre-trained first physical information neural network, in order to dynamically adjust the grouting parameters, specifically: The current grouting parameters of the grouting equipment and the probability set of the third fracture network are input into a pre-trained first physical information neural network, so that the first physical information neural network outputs a fracture propagation prediction result; wherein, the grouting response data includes grouting pressure and grouting flow rate; the fracture propagation prediction result includes grout diffusion area, new fracture risk value and theoretical pressure-flow rate curve; An actual pressure-flow response curve is generated based on the grouting response data. By comparing the actual pressure-flow response curve with the theoretical pressure-flow curve, abnormal responses during the grouting process are identified, and abnormal response results are obtained. Grouting parameters are then dynamically adjusted based on the abnormal response results. The abnormal responses include hydraulic fracturing, blockage, and grout leakage.

[0017] This invention, through its embodiments, inputs current grouting parameters and a third fracture network into a first physical information neural network. This network combines real-time construction parameters and the latest cognitive models for prediction, outputting the grout diffusion area, the risk value of newly formed fractures, and the theoretical pressure-flow curve. This predicts the potential impact of grouting and provides a data foundation for subsequent adjustments to grouting parameters. Furthermore, it generates actual pressure-flow curves based on grouting response data to quantify the true feedback from fractures during the grouting process. By comparing the actual curves with the theoretical curves, it identifies abnormal responses and senses grouting disturbances. This ensures that grouting is no longer blindly executed according to a plan, but rather that it can perceive and respond to rock mass reactions in real time, avoiding secondary disasters such as hydraulic fracturing and ensuring effective sealing.

[0018] In some preferred embodiments of the first aspect, an actual pressure-flow response curve is generated based on the grouting response data, and abnormal responses during the grouting process are identified by comparing the actual pressure-flow response curve with the theoretical pressure-flow curve, thereby obtaining abnormal response results, specifically as follows: The grouting pressure and grouting flow rate are filtered and smoothed to obtain the actual pressure-flow rate response curve, and the instantaneous deviation and derivative value between the actual pressure-flow rate response curve and the theoretical pressure-flow rate curve are calculated. Based on the instantaneous deviation value, derivative value, and preset anomaly judgment threshold, combined with preset geomechanical mechanism rules, abnormal responses during the grouting process are identified, and abnormal response results are obtained.

[0019] This invention combines preset geomechanical mechanism rules for comprehensive identification, and associates data features with physical mechanisms to accurately identify hydraulic fracturing, blockage, leakage, etc., thereby improving the precision of control.

[0020] In some preferred embodiments of the first aspect, the step of identifying seepage channels based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, and generating a first fracture network probability set in conjunction with pre-stored fracture attitude statistics, specifically involves: The surface image data and the slope topography data are aligned to obtain a multimodal image set. The multimodal image set is then input into a pre-trained multi-scale attention segmentation network so that the multi-scale attention segmentation network can generate a surface fissure vector map with seepage channel location markers by identifying surface fissures and seepage traces in the multimodal image set. Based on the slope topography data, the elevation data corresponding to each fracture point in the surface fracture vector map is obtained, and a three-dimensional surface fracture network is generated based on the elevation data and the elevation difference. Based on pre-stored fracture attitude statistics, the three-dimensional surface fracture network is subjected to subsurface simulation to generate several subsurface fracture networks with different realization probabilities; wherein, the subsurface simulation includes fracture dip angle simulation, fracture depth simulation, and realization probability simulation. The three-dimensional surface fracture network is combined with each of the underground fracture networks to construct several fracture networks with heterogeneous graph structures. The realization probability of the underground fracture network is determined as the realization probability of the corresponding fracture network, so as to integrate all fracture networks and form a first fracture network probability set.

[0021] This invention identifies surface fissures and seepage traces using a multi-scale attention segmentation network, automatically extracts fissure information from images using deep learning, achieves efficient and accurate surface mapping, and generates a surface fissure vector map with seepage channel markings, providing a foundation for 3D extrapolation. A 3D surface fissure network is generated based on elevation data, mapping 2D fissures onto the real terrain to obtain a 3D structure with real geographic coordinates. Subsurface extrapolation is performed on the 3D network based on fissure attitude statistics, inferring possible subsurface fissure extensions from the surface, constructing a 3D cognitive model, and using probability sets to characterize the uncertainty of the subsurface structure, providing a correctable cognitive basis for subsequent dynamic updates. By combining surface and subsurface data to construct a heterogeneous graph structure, a comprehensive data foundation is provided for subsequent graph theory analysis and neural network processing.

[0022] In some preferred embodiments of the first aspect, the step of performing subsurface extrapolation on the three-dimensional surface fracture network based on pre-stored fracture attitude statistics to generate several subsurface fracture networks with different realization probabilities specifically involves: Extract the node features of each surface node in the three-dimensional surface fracture network, and combine them with the pre-stored fracture attitude statistics and the node features to infer the probability distribution of the first attitude of underground fractures at each surface node; wherein, the node features include the orientation features and seepage marker features of the connected fractures; The Monte Carlo method is used to sample multiple times from the first attitude probability distribution to generate several sets of three-dimensional coordinates of underground nodes. A pre-trained graph neural network is then connected to a predictor to infer the connection relationship of each set of underground nodes based on the node features and the three-dimensional coordinates, thereby generating underground fracture edges connecting each set of underground nodes to construct several initial underground fracture networks. The initial subsurface fracture network is statistically analyzed to obtain the second attitude probability distribution of each initial subsurface fracture network, and pre-stored regional attitude statistics are obtained to calculate the similarity between each second attitude probability distribution and the regional attitude statistics. The realization probability of each initial underground fracture network is determined based on the similarity, resulting in several final underground fracture networks.

[0023] This invention provides comprehensive and structured surface information by extracting surface node features (trend, seepage markers), and inferring the probability distribution of the first attitude by combining fracture attitude statistics, thereby using statistical laws to guide the extension direction of underground fractures. Multiple sets of underground node coordinates are generated through Monte Carlo sampling, and multiple possible underground structures are generated through random sampling to quantify uncertainty. A graph neural network prediction algorithm is used to infer node connectivity relationships and predict which underground nodes are connected by fractures, thus constructing a complete underground network topology. By statistically analyzing the attitude distribution of the initial network, the overall attitude characteristics of each realization are calculated, providing a basis for probability weighting. The degree of conformity between each realization and geological laws is evaluated by calculating the similarity with regional attitude statistics, quantifying the rationality of the realization. The realization probability is determined based on similarity, allowing realizations that better conform to geological laws to receive higher weights.

[0024] Secondly, embodiments of the present invention provide a deep learning-based system for treating seepage channels in fractured rock slopes, comprising a fracture network generation module, a borehole strategy generation module, a borehole treatment module, and a grouting treatment module, wherein... The fracture network generation module is used to identify seepage channels based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, and generate a first fracture network probability set by combining pre-stored fracture attitude statistics data; wherein, the first fracture network probability set includes several fracture networks and the realization probability of each fracture network. The drilling strategy generation module is used to generate a drilling strategy based on the first fracture network probability set, with the goal of maximizing information gain and maximizing governance potential, through deep reinforcement learning; wherein, the drilling strategy includes drilling sequence, drilling location and drilling parameters; The drilling processing module is used to acquire fracture observation data in real time during the drilling process of the drilling equipment executing the drilling strategy, and dynamically correct the pre-acquired second fracture network probability set according to the fracture observation data using a Bayesian update method. Based on the corrected second fracture network probability set, the drilling strategy is adjusted in real time until the seepage channel drilling processing of the target rock slope is completed. The second fracture network probability set is obtained by predicting fracture propagation from the first fracture network probability set based on the currently implemented drilling strategy. The grouting module is used to monitor grouting response data in real time during the grouting process of the grouting equipment, and combine it with the pre-acquired third fracture network probability set to predict fracture propagation through a pre-trained first physical information neural network, so as to dynamically adjust the grouting parameters until the grouting treatment of the seepage channel of the target rock slope is completed; wherein, the third fracture network probability set is the most recently corrected second fracture network probability set.

[0025] This invention employs a fracture network generation module to generate a first fracture network probability set, constructing an initial cognitive model and quantifying uncertainty using probability sets to provide a correctable cognitive foundation for subsequent dynamic updates. A drilling strategy generation module utilizes deep reinforcement learning to generate drilling strategies, formulating optimal drilling plans based on the initial cognition, providing initial construction guidance, and reserving space for subsequent adjustments. A drilling processing module acquires fracture observation data in real time during drilling to perceive the actual rock mass information revealed by the drilling operation, thereby introducing real-time feedback, breaking the limitations of "static prediction" in existing technologies, and based on the observation data... The Bayesian update dynamically corrects the second fracture network probability set, enabling the model's understanding to evolve in real time with the construction process. This ensures that the construction strategy will not fail due to changes in fracture morphology caused by drilling. Through the grouting processing module, grouting response data is monitored in real time during the grouting process, thereby perceiving the rock mass response during grouting and providing real-time basis for grouting control. Combined with the third fracture network probability set, fracture propagation prediction is performed to predict fracture changes that may be induced by grouting pressure. This takes into account grouting disturbances and dynamically adjusts grouting parameters to avoid secondary disasters such as hydraulic fracturing, ensuring that the construction strategy will not fail due to changes in fracture morphology caused by grouting pressure.

[0026] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0027] Figure 1 A schematic diagram of a deep learning-based method for treating seepage channels in fractured rock slopes, provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a deep learning-based seepage channel treatment system for fractured rock slopes, provided as an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1: like Figure 1 As shown in the figure, a method for treating seepage channels in fractured rock slopes based on deep learning is provided by an embodiment of the present invention, including the following steps: S101, based on the real-time acquired surface image data of the target rock slope and the pre-stored slope topography data, identify seepage channels, and combine them with the pre-stored fracture attitude statistics to generate a first fracture network probability set; wherein, the first fracture network probability set includes several fracture networks and the realization probability of each fracture network. In this embodiment, the step of identifying seepage channels based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, and generating a first fracture network probability set by combining pre-stored fracture attitude statistics, specifically involves: aligning the surface image data with the slope topographic data to obtain a multimodal image set; inputting the multimodal image set into a pre-trained multi-scale attention segmentation network; enabling the multi-scale attention segmentation network to generate a surface fracture vector map with seepage channel location markers by identifying surface fractures and seepage traces in the multimodal image set; and obtaining the surface fracture vector map based on the slope topographic data. The elevation data corresponding to each fracture point in the figure are used to generate a three-dimensional surface fracture network based on the elevation difference. Subsurface simulation is performed on the three-dimensional surface fracture network based on pre-stored fracture attitude statistics to generate several subsurface fracture networks with different realization probabilities. The subsurface simulation includes fracture dip angle simulation, fracture depth simulation, and realization probability simulation. The three-dimensional surface fracture network is combined with each of the subsurface fracture networks to construct several fracture networks with heterogeneous graph structures. The realization probability of each subsurface fracture network is determined as the realization probability of the corresponding fracture network, thus integrating all fracture networks to form a first fracture network probability set.

[0030] In one specific embodiment, multi-source data preprocessing and alignment are required: Geometric and radiometric corrections are performed on surface image data (visible light, thermal infrared, and multispectral) to eliminate lens distortion and illumination differences; using DSM topographic data as the reference coordinate system, all images are unified to the same spatial reference system through feature point matching or geographic coordinate registration to generate a multimodal aligned image set; temperature calibration is performed on the thermal infrared images, converting the original DN values ​​into actual temperature values ​​(°C); the Normalized Difference Vegetation Index (NDVI) is calculated for the multispectral images to identify vegetation cover areas. Pre-stored fracture attitude statistics are digitized into probability distribution parameters for subsequent subsurface simulation, ultimately resulting in a multimodal aligned image set (each pixel contains multi-channel information such as visible light RGB, temperature value, and NDVI), a topographic elevation raster (consistent with image resolution, with each pixel accompanied by an elevation value), and fracture attitude statistical probability distribution parameters.

[0031] In one specific embodiment, the process of identifying surface fissures and seepage traces is as follows: a multimodal aligned image set is input into a pre-trained multi-scale attention segmentation network. The network extracts multi-scale features through an encoder-decoder structure and uses an attention mechanism to enhance its sensitivity to small fissures and seepage anomalies, so that the network outputs two probability maps with the same resolution as the input image—a fissure probability map (used to identify the probability of each pixel belonging to a fissure) and a seepage probability map (the probability of each pixel belonging to an active seepage area); the fissure probability map is thresholded (e.g., threshold 0.5) to obtain a binary fissure map; Morphological operations are performed to remove isolated noise points and connect fractures. The binary image is then extracted and vectorized to generate surface fracture vector segments. Each segment contains: a point sequence (two-dimensional pixel coordinates), fracture length, average width (estimated based on a probability map), and fracture orientation (fitted from the point sequence). Each fracture segment is overlaid with a seepage probability map. If the average seepage probability of pixels on the segment exceeds a threshold (e.g., 0.5), the fracture is marked as a seepage channel, and the seepage intensity (average probability value) is recorded. Finally, a surface fracture vector map (two-dimensional, with seepage attributes) with seepage channel location markers is output.

[0032] In one specific embodiment, the process of generating the three-dimensional surface fracture network is as follows: converting the coordinates of each two-dimensional pixel in the surface fracture vector map into geographic coordinates, obtaining the ground elevation of each fracture point through terrain elevation raster (DSM) interpolation, and generating three-dimensional point coordinates; for fractures crossing steep slopes or terrain, correcting the elevation connection logic using slope abrupt change information in the terrain data to ensure that the fracture line is continuous along the terrain surface; connecting the points of the same fracture in the original order to generate a three-dimensional surface fracture network, each fracture containing a node sequence (three-dimensional coordinates), geometric attributes (fitted from adjacent points), and seepage attributes (whether it is a seepage channel, seepage intensity); finally outputting a three-dimensional surface fracture network (nodes are fracture points, edges are fracture segments).

[0033] In one specific embodiment, the subsurface fracture network derivation process is as follows: Feature vectors of each surface node in the three-dimensional surface fracture network are extracted, including node coordinates, node type (endpoint / intersection), orientation statistics of connecting fractures (mean orientation, orientation variance), and seepage markers (if there are seepage channels in the connected fractures); the node features are input into a pre-trained conditional variational autoencoder, which has been trained using regional fracture attitude statistics and measured data of similar slopes, outputting the first attitude probability distribution of subsurface fractures at each surface node (mean and variance of dip and dip angle); the number of Monte Carlo samplings is set (N=100), and for each surface node, a set of attitude values ​​(dip α, dip angle β) is randomly sampled from its attitude probability distribution, and extended downwards along this direction to generate a series of subsurface node coordinates (x+d). sinβ sinα,y+d sinβ cosα,z d (cosβ), where d is the step size (e.g., 1m), and the extension depth is determined by engineering needs (e.g., below the potential slip surface). For each set of sampled subsurface nodes, a pre-trained graph neural network is used to connect the predictor. The input information includes the spatial distance between node pairs, consistency of attitude, and whether they belong to the same parent fracture. The output is the probability that there is a fracture connection between nodes. If the probability exceeds a threshold (e.g., 0.5), a subsurface fracture edge is added. After the above operations, N initial subsurface fracture networks are generated, each network containing subsurface nodes and edges.

[0034] In one specific embodiment, the realization probability determination process is as follows: For each initial subsurface fracture network, the attitude distribution (histograms of dip and dip angles) of all fracture segments is statistically analyzed to obtain the second attitude probability distribution of the network; pre-stored regional attitude statistics (as geological priors) are obtained, and the similarity (such as KL divergence or cosine similarity) between the second attitude probability distribution of each initial subsurface fracture network and the regional attitude statistics distribution is calculated; the similarity is normalized to obtain the realization probability of each initial subsurface fracture network. (satisfy The higher the similarity (i.e., the more it conforms to the regional geological patterns), the greater the probability of realization; based on the realization probability, all networks are selected or retained to form several final underground fracture networks, each with a probability weight.

[0035] In one specific embodiment, the process of integrating the three-dimensional surface fracture network with each final subsurface fracture network is as follows: surface nodes and subsurface nodes are merged to form a complete node set; surface edges and subsurface edges are retained respectively, and connections are established at the intersections (if the spatial positions of subsurface nodes and surface nodes coincide or are close); a fracture network with a heterogeneous graph structure is constructed, and finally, all integrated fracture networks and their realization probabilities are combined to form the first fracture network probability set. .

[0036] Optionally, the node attributes of the fracture network include three-dimensional coordinates, node type (surface / subsurface / intersection), seepage marker, and implementation ID.

[0037] Optionally, the edge attributes of the fracture network include the connection of node pairs, length, aperture, orientation, permeability, whether it is a seepage channel, and the implementation ID to which it belongs.

[0038] In this embodiment, the step of performing subsurface extrapolation on the three-dimensional surface fracture network based on pre-stored fracture attitude statistics to generate several subsurface fracture networks with different realization probabilities specifically involves: extracting node features of each surface node in the three-dimensional surface fracture network, and combining the pre-stored fracture attitude statistics and the node features to infer the first attitude probability distribution of subsurface fractures at each surface node; wherein, the node features include the orientation features and seepage marker features of the connected fractures; and generating several sets of three-dimensional coordinates of subsurface nodes by multiple samplings from the first attitude probability distribution using the Monte Carlo method, and then using pre-stored fracture attitude statistics to generate several sets of subsurface node three-dimensional coordinates. The trained graph neural network connection predictor infers the connection relationships of each group of underground nodes based on the node features and the three-dimensional coordinates, and generates underground fracture edges connecting each group of underground nodes to construct several initial underground fracture networks. The initial underground fracture networks are then statistically analyzed for their attitude distributions to obtain a second attitude probability distribution for each initial underground fracture network. Pre-stored regional attitude statistics are also obtained to calculate the similarity between each second attitude probability distribution and the regional attitude statistics. Based on the similarity, the realization probability of each initial underground fracture network is determined, resulting in several final underground fracture networks.

[0039] In one specific embodiment, the surface node feature extraction process is as follows: traversing all nodes in the three-dimensional surface fracture network and constructing a feature vector for each node; wherein, the feature vector includes: Coordinate characteristics: Normalized 3D coordinates (coordinates are scaled to a uniform range to eliminate scale effects); Connecting fracture orientation statistics: Statistically analyze the orientation of all surface fractures connected to the node, and calculate the average orientation and orientation variance (if the node is an isolated endpoint, take the orientation of the connected fractures; if it is an intersection point, take the mean and variance of the orientation of all connected fractures). Seepage marker: If any crack connected to this node is a seepage channel, the seepage marker is 1; otherwise, it is 0 (one-hot encoding can be used). Node type: Endpoint / intersection category code (e.g., endpoint=0, intersection=1); Organize the feature vectors of all nodes into a matrix to obtain the surface node feature matrix, with each node accompanied by a complete feature vector.

[0040] In one specific embodiment, the inference process for the probability distribution of subsurface fracture attitude is as follows: the feature matrix of the surface node is input into the encoder part of the pre-trained conditional variational autoencoder (CVAE); the CVAE encoder outputs the latent variable distribution parameters corresponding to each node, including the mean and log-variance, and the dimension of the latent variable is usually 2 (corresponding to dip and dip angle) or higher; latent variables are sampled from the latent variable distribution through reparameterization techniques, and then input into the decoder; the decoder outputs the first probability distribution parameters of the subsurface fracture attitude at the node, including the mean and standard deviation of the dip distribution and the mean and standard deviation of the dip angle distribution, which reflects the possible range of the direction of extension of the subsurface fracture starting from the surface node.

[0041] In one specific embodiment, the process of generating underground nodes through Monte Carlo sampling is as follows: The number of sampling times is set to N (e.g., N=100), i.e., N possible underground fracture networks are generated; for the k-th sampling (k=1... N) Perform the following operations: Initialize the set of underground nodes for this sampling; Traverse each surface node and randomly sample a set of attitude values—dip and dip angle—from the first attitude probability distribution of that node, ensuring that the dip is in the range of [0, 360] and the dip angle is in the range of [0, 90]. Starting from the surface node coordinates, extend downward along the sampled attitude direction and generate a series of underground node coordinates with a fixed step size (e.g., 1m). Add all the generated underground nodes and record the parent node (corresponding surface node) and depth layer of each node.

[0042] After the kth sampling is completed, the set of underground nodes for that sampling is obtained, which includes all underground nodes extending from the surface.

[0043] In one specific embodiment, the graph neural network connection inference process is as follows: traversing all node pairs, if the spatial distance between two points is less than a preset maximum connection distance (e.g., 10m, set according to the statistical length of regional fissures), then it is considered a candidate pair; extracting feature vectors for each candidate node pair, including: Spatial distance and orientation consistency: If two points both originate from the extension of the same surface parent fissure, or have similar orientation directions, then the consistency is high (which can be measured by the dot product of direction vectors). Homogeneity: Do the two points originate from the same surface node? Are there any other nodes in the middle: Are there any other nodes blocking the direction of the connection? Depth difference: The difference in depth between two points; Furthermore, the feature vector is input into the pre-trained graph neural network connection predictor, which outputs the probability that there is a fracture connection between the node pairs. If the probability is greater than a preset threshold (e.g., 0.5), an underground fracture edge is added between the nodes, and the edge is recorded as belonging to the underground fracture edge set of the k-th sampling. After traversing all candidate node pairs, the underground fracture map of the k-th sampling is obtained.

[0044] In one specific embodiment, the similarity between each network attitude distribution and the regional attitude distribution can be calculated using one of the following methods: the negative value of the Kullback-Leibler divergence, cosine similarity (calculating the cosine distance after vectorizing the distribution), or cross-entropy.

[0045] S102, based on the first gap network probability set, with the goal of maximizing information gain and maximizing governance potential, a drilling strategy is generated through deep reinforcement learning; wherein, the drilling strategy includes drilling sequence, drilling location and drilling parameters; In one specific embodiment, before generating the drilling strategy using deep reinforcement learning, the first set of fracture network probabilities needs to be transformed into a current cognitive state encoding. This process specifically involves: using a graph neural network to extract features from each fracture network implementation, generating an embedding vector for each node. This vector integrates the node's geometric location, seepage properties, and connectivity with surrounding fractures. The node embeddings from all implementations are weighted and averaged according to their respective probability weights to obtain the average embedding for each node, representing the best current estimate for that node. Simultaneously, the degree of embedding difference for each node across implementations is calculated as a measure of the cognitive uncertainty of that node. A greater difference indicates greater uncertainty in the geological conditions at that node. The average embedding and uncertainty measure of all nodes are integrated into a global state vector. This vector comprehensively describes the current overall understanding of the slope fracture network, including the most likely structure and uncertainty distribution. This state vector will serve as the input to the reinforcement learning model.

[0046] In one specific embodiment, before generating drilling strategies through deep reinforcement learning, a candidate drilling action space needs to be constructed. This process involves: discretizing the reachable area of ​​the drilling rig into a grid of candidate drilling points. Each grid point is combined with terrain data to obtain ground elevation, forming three-dimensional coordinates. For each candidate point, feasible combinations of drilling parameters are defined based on construction constraints, including dip angle (selecting several typical values ​​within the allowable range), depth (selecting several typical values ​​within the maximum depth range), and azimuth angle (which can be a pre-selected fixed value based on the dominant direction of local fractures, or discretized into several main directions). Each candidate point and its parameter combination constitute a candidate action, and all candidate actions form the action space. The action space is pre-calculated and stored before inference.

[0047] In one specific embodiment, the borehole sequence decision-making process based on reinforcement learning is as follows: An initial state vector is input into a pre-trained deep reinforcement learning model. This model evaluates the value of each candidate action in the current state (i.e., the expected cumulative benefit of the action); the model outputs a score for each candidate action, which comprehensively considers the expected information gain and governance potential of the borehole, minus the borehole cost. Information gain is measured by the degree of uncertainty reduction in the area surrounding the borehole; governance potential is evaluated by the contribution of the fracture segment traversed by the borehole to the seepage network (e.g., whether it is located at a critical node); cost is related to borehole depth; the candidate action with the highest score is selected as the first borehole and recorded in the borehole list. Furthermore, to determine the next borehole, it is necessary to estimate the change in cognitive state after executing the first borehole. This is achieved through a pre-trained state transition network: the network takes the current state and the selected action as input and outputs a rough new state (i.e., the updated cognitive state) after executing the borehole. The new state reflects the expectation of reduced uncertainty due to the information revealed by the borehole.

[0048] Furthermore, the new state is used as the current state, and the above process is repeated, selecting subsequent boreholes in sequence until the budget limit (such as the maximum number of boreholes) is reached or the remaining action score is lower than a preset threshold. This results in an ordered list of boreholes, each containing its sequence, location, and parameters.

[0049] S103, during the drilling process, the drilling equipment acquires fracture observation data in real time. Based on the fracture observation data, the pre-acquired second fracture network probability set is dynamically corrected using a Bayesian update method. The drilling strategy is then adjusted in real time based on the corrected second fracture network probability set until the seepage channel drilling of the target rock slope is completed. The second fracture network probability set is obtained by predicting fracture propagation based on the first fracture network probability set using the currently implemented drilling strategy. In this embodiment, based on the fracture observation data, the pre-acquired second fracture network probability set is dynamically corrected using a Bayesian update method. This allows for real-time adjustment of the drilling strategy based on the corrected second fracture network probability set. Specifically, based on the drilling strategy currently implemented by the drilling equipment, a pre-trained second physical information neural network model is used to predict fracture propagation in the first fracture network probability set to obtain the second fracture network probability set. Furthermore, based on the fracture observation data, fracture features of the currently observed fractures are extracted, and the second fracture network probability is calculated based on these features. The likelihood of each fracture network in the set relative to the currently observed fracture is calculated; wherein the fracture features include fracture depth, fracture orientation, and fracture aperture; based on the likelihood, the realization probability of each fracture network in the second fracture network probability set is updated using a preset Bayesian formula to obtain a corrected second fracture network probability set; based on the corrected second fracture network probability set, the expected information gain assessment and governance potential assessment are performed on the un-drilled boreholes in the drilling strategy, so as to comprehensively consider the expected information gain assessment results and governance potential assessment results of each un-drilled borehole and adjust the drilling strategy in real time.

[0050] In this embodiment, based on the drilling strategy currently implemented by the drilling equipment, a second fracture network probability set is obtained by predicting fracture propagation using a pre-trained second physical information neural network model. Specifically, a drilling disturbance source function is constructed according to the drilling strategy currently implemented by the drilling equipment, and the drilling disturbance source function is determined as the boundary condition of the second physical information neural network model. A stress field inversion is performed on each fracture network in the first fracture network probability set using a pre-constructed numerical inversion model to obtain the first stress field of each fracture network. The boundary condition, the first stress field of each fracture network, and the first fracture network probability set are input into the pre-trained second physical information neural network model, so that the second physical information neural network model updates the first stress field according to a preset elasticity equilibrium equation and fracture mechanics criterion, obtaining the second stress field of each fracture network in the first fracture network probability set. Based on the second stress field, fracture propagation prediction is performed on each fracture network in the first fracture network probability set to obtain the second fracture network probability set.

[0051] In one specific embodiment, the process of constructing the borehole disturbance source function is as follows: Based on the currently implemented borehole parameters, the trajectory line of the borehole in three-dimensional space is determined; combining the borehole diameter and typical drilling techniques (such as stress release generated by rotary drilling), a disturbance source function describing the mechanical effects of the borehole on the surrounding rock mass is constructed. This function uses the borehole wall as the boundary of action, characterizing the stress unloading caused by borehole hollowing and the additional dynamics generated by drilling impact; the disturbance source function is transformed into a boundary condition format acceptable to a physical information neural network, such as applying a normal stress release amount (estimated based on the initial geostress) or a displacement boundary on the borehole wall.

[0052] In one specific embodiment, the process of obtaining the first stress field of each fracture network is as follows: for each fracture network realization in the first fracture network probability set, its geometric structure (node ​​coordinates, fracture orientation) is input into a pre-constructed numerical inversion model; the numerical inversion model uses the regional tectonic stress field as the background, combined with the pre-stored regional geostress measured data, and solves the elasticity equilibrium equation to invert the initial stress tensor (including three normal stress components and three shear stress components) at each node; the inverted first stress field is attached as an attribute to the node of the corresponding realization to form a "fracture network realization with initial stress".

[0053] In one specific embodiment, the stress field update process of the physical information neural network is as follows: each fracture network implementation (including node coordinates and mechanical properties of edge connections), the first stress field of the implementation, and the pre-constructed borehole disturbance boundary conditions are input into a pre-trained second physical information neural network; the second physical information neural network embeds elasticity equilibrium equations (such as stress equilibrium differential equations) and fracture mechanics criteria (such as maximum circumferential stress criteria); the network simulates the stress redistribution of the entire rock mass region under borehole disturbance by solving partial differential equations; the network outputs the updated stress field, i.e., the second stress field at each node (stress tensor after borehole disturbance), and may also output intermediate results such as stress concentration regions.

[0054] In one specific embodiment, the process of predicting crack propagation based on the second stress field is as follows: For each crack network implementation, based on the updated second stress field, the stress state of each crack segment is analyzed one by one, and the normal stress and shear stress acting on the crack surface are calculated; if the normal stress decreases relative to the initial state (i.e., tensile stress increases or compressive stress decreases), it is determined that the crack may open, the increase in opening is calculated according to the normal stiffness of the crack, and the opening attribute of the edge is updated; the stress intensity factor at the crack tip is calculated, and if it exceeds the fracture toughness of the material, it is determined that the crack may propagate. The propagation direction and length of the new crack are determined according to the direction of the maximum principal stress or fracture mechanics criteria, and new nodes and edges are added at the corresponding positions and assigned initial attributes (such as small opening and permeability); all existing crack segments and new crack segments together constitute the updated crack network implementation, and the probability weights of all implementations remain unchanged (the propagation prediction itself does not change the probability, only the network structure).

[0055] In this embodiment, based on the fracture observation data, fracture features of the currently observed fracture are extracted, and based on the fracture features, the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture is calculated. Specifically, based on the fracture observation data, fracture features of the currently observed fracture are extracted to obtain an actual feature list of the currently observed fracture; based on the drilling strategy currently implemented by the drilling equipment, spatial geometric intersection calculations are performed in each fracture network of the second fracture network probability set to determine the fracture segments in each fracture network that intersect with the current drilling trajectory, and fracture features of each fracture segment are extracted to obtain a predicted feature list of each fracture network; based on the actual feature list and the predicted feature list of each fracture network, the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture is calculated.

[0056] In one specific embodiment, the real-time observation data preprocessing and fracture feature extraction process involves: real-time processing of the borehole television image stream, and using a lightweight semantic segmentation network (such as MobileNet-UNet) to identify fracture pixels in each frame of the image. This network has been pre-trained on a large number of borehole images and is capable of accurately segmenting fracture regions; Based on the geometric relationships of the image, pixel coordinates are converted into angular coordinates on the hole wall, and combined with depth information, the cracks identified in consecutive frames are associated to form independent crack objects; for each identified crack, the following features are extracted: Depth location: The depth of the fracture center point along the borehole axis, determined by the image depth marker; Attitude: dip and dip angle, obtained by fitting the traces of the fracture on the borehole wall (using ellipse fitting or sine wave fitting algorithm). Opening: The average width of the crack on the image, converted to millimeters using the image scale; Confidence score: The probability value output by the segmentation network, reflecting the reliability of the crack identification; Furthermore, all extracted fractures are sorted by depth to generate a fracture observation list for the current borehole. Each observation includes depth, attitude (dip, dip angle), aperture, and confidence level.

[0057] In one specific embodiment, the process of generating a predicted fracture list for each fracture network is as follows: Based on the current borehole parameters, construct the spatial linear equation of the borehole trajectory; determine the starting point coordinates and direction vector, as well as the depth range (from the borehole opening to the current depth); traverse each implementation in the second fracture network probability set... For each fracture segment (edge), treat it as a spatial line segment and calculate the nearest spatial distance between that line segment and the straight line of the borehole trajectory. If the nearest distance is less than a preset threshold (e.g., 0.1 meters, considering borehole diameter and positioning error), it is considered that the borehole has passed through the fracture segment. Calculate the coordinates of the intersection point and the corresponding borehole depth (distance from the borehole opening to the intersection point along the trajectory); record the predicted features of the fracture segment: predicted depth, predicted attitude (obtained from edge attributes), predicted aperture (obtained from edge attributes), and the fracture segment identifier; sort all traversed fracture segments by predicted depth to generate a predicted fracture list for this implementation. ,Include The characteristics of the predicted fracture.

[0058] In one specific embodiment, the process of matching observations and predictions and calculating likelihood is as follows: For each implementation, the observation list and prediction list are associated and matched. The nearest neighbor association method is used, matching one by one in depth order: Observations are taken out sequentially from the observation list. The prediction list is searched for the prediction with the closest depth and the depth difference is less than the threshold (e.g., 3 times the depth standard deviation). If a match is found and is not occupied, a matching pair is established. After one round of matching, unmatched observations are recorded as "new observations" (i.e. cracks not predicted in the implementation) and unmatched predictions are recorded as "missed predictions" (i.e. cracks that exist in the implementation but have not been observed).

[0059] Furthermore, the probability of each matching pair is calculated based on the observation noise model: Depth deviation, attitude deviation, and aperture deviation each follow a preset normal or log-normal distribution. The joint probability density of each deviation is used as the matching probability of the matching pair. For unmatched observations (new observations), a small fixed probability (e.g., 0.01) is assigned, representing the possibility that the fracture actually exists and is observed, even though it was not predicted in the implementation. For unmatched predictions (missed detections), a small fixed probability (e.g., 0.01) is also assigned, representing the possibility that the fracture exists in the implementation but is not recognized by the image. Assuming that the observations are independent, the likelihood of the implementation generating the current observation list is the product of the probabilities of all matching pairs, the probability of new observations, and the probability of missed detection predictions. For computational stability, the log-likelihood is usually taken.

[0060] Repeat the above calculation for all implementations to obtain the likelihood of each implementation.

[0061] In this embodiment, based on the modified second fracture network probability set, the expected information gain and governance potential assessments are performed on the unexecuted boreholes in the borehole strategy. Specifically, for each unexecuted borehole in the borehole strategy, spatial geometric intersection calculations are performed in each fracture network in the modified fracture network probability set according to the corresponding borehole location and borehole parameters. The fracture segments in each fracture network that intersect with the preset trajectories of each unexecuted borehole are determined, and the fracture features of each fracture segment are extracted to obtain a predicted fracture feature list for each unexecuted borehole. The predicted fracture feature list and the modified second fracture network probability set are used to calculate the expected information gain for each un-drilled borehole using Monte Carlo simulation, thus obtaining the expected information gain evaluation result for each un-drilled borehole. Based on the realization probability of each fracture network in the predicted fracture feature list and the modified second fracture network probability set, and the preset graph theory indices, the remediation potential of each un-drilled borehole is evaluated, thus obtaining the remediation potential evaluation result for each un-drilled borehole. The graph theory indices include betweenness centrality indices and seepage centrality indices.

[0062] In one specific embodiment, the process of obtaining information about fracture segments traversed by performing spatial geometric intersection for each un-drilled borehole is as follows: For each borehole, a spatial linear equation for the borehole trajectory is constructed based on its borehole parameters (location, dip angle, azimuth angle, depth); each fracture network implementation in the modified second fracture network probability set is traversed: spatial geometric intersection calculation is performed in the implementation (only the network structure is read, without changing the network), and the spatial nearest distance between the borehole trajectory and each fracture segment (edge) is determined; if the nearest distance is less than a preset threshold (e.g., 0.1 meters), it is considered that the borehole has traversed the fracture segment, and the identifier, attributes (attitude, aperture, seepage mark) and predicted depth at the intersection point of the fracture segment are recorded; the information of all fracture segments traversed by the borehole in the implementation is summarized to obtain the set of fracture segments traversed by the borehole in the implementation. In one specific embodiment, the process of evaluating the expected information gain (based on Monte Carlo simulation) is as follows: First, the overall uncertainty of the current probability set is calculated. This can be approximated by the sum of the entropies of the existence probabilities of all fracture segments, or by the entropy of the existence probabilities of key nodes. For each un-drilled borehole, a Monte Carlo simulation is performed: the number of simulations is set; for each simulation, a "real network" is randomly sampled from the current probability set according to probability weights as the real subsurface structure for this simulation; based on this real network, the fracture data that the borehole might observe (i.e., the fracture segments traversed by the borehole) is simulated, noise is added (random errors in depth, attitude, and aperture are generated according to the observation noise model), generating a set of simulated observation data; this set of simulated observation data is applied to the current probability set, a Bayesian update is performed, the updated probability set is obtained, and its uncertainty measure is calculated. The information gain of this simulation is recorded as the current uncertainty minus the posterior uncertainty; the average information gain of this simulation is calculated as the expected information gain of the borehole.

[0063] Preferably, if computational resources are limited, the sum of the probability variances of the nodes in the area surrounding the borehole can be used as an approximation of the information gain, without the need for a full Monte Carlo simulation.

[0064] In one specific embodiment, the governance potential assessment (based on graph theory metrics) process is as follows: For each un-drilled borehole, calculate the betweenness centrality (measuring the importance of a node in the shortest path) and seepage centrality (considering the weighted importance of seepage channels) of all nodes (or key nodes) in each fracture network implementation; for each fracture segment traversed by the borehole, obtain the centrality values ​​of the nodes at its two ends, and take the average as the governance contribution of that fracture segment; sum the governance contributions of all segments traversed (aperture weighting can be considered) to obtain the governance potential value of the borehole in the implementation; and take the weighted average of the governance potential values ​​of each implementation according to probability weights to obtain the overall governance potential of each un-drilled borehole.

[0065] S104, during the grouting process, the grouting response data is monitored in real time, and combined with the pre-acquired third fracture network probability set, the fracture propagation is predicted through the pre-trained first physical information neural network to dynamically adjust the grouting parameters until the grouting treatment of the seepage channel of the target rock slope is completed; wherein, the third fracture network probability set is the most recently corrected second fracture network probability set.

[0066] In this embodiment, the real-time monitoring of grouting response data, combined with a pre-acquired third fracture network probability set, and the prediction of fracture propagation through a pre-trained first physical information neural network to dynamically adjust grouting parameters, specifically involves: inputting the current grouting parameters of the grouting equipment and the third fracture network probability set into the pre-trained first physical information neural network, so that the first physical information neural network outputs fracture propagation prediction results; wherein, the grouting response data includes grouting pressure and grouting flow rate; the fracture propagation prediction results include grout diffusion area, new fracture risk value, and theoretical pressure-flow rate curve; generating an actual pressure-flow rate response curve based on the grouting response data, and identifying abnormal responses during the grouting process by comparing the actual pressure-flow rate response curve with the theoretical pressure-flow rate curve, obtaining abnormal response results, and dynamically adjusting grouting parameters based on the abnormal response results; wherein, the abnormal responses include hydraulic fracturing, blockage, and grout leakage.

[0067] In this embodiment, an actual pressure-flow response curve is generated based on the grouting response data. Abnormal responses during the grouting process are identified by comparing the actual pressure-flow response curve with the theoretical pressure-flow curve, and abnormal response results are obtained. Specifically, the grouting pressure and grouting flow rate are filtered and smoothed to obtain the actual pressure-flow response curve, and the instantaneous deviation and derivative values ​​between the actual pressure-flow response curve and the theoretical pressure-flow curve are calculated. Based on the instantaneous deviation value, derivative value, and a preset anomaly judgment threshold, combined with preset geomechanical mechanism rules, abnormal responses during the grouting process are identified, and abnormal response results are obtained.

[0068] In one specific embodiment, the real-time data preprocessing and smoothing filtering process is as follows: the real-time collected grouting pressure and grouting flow data are subjected to sliding window filtering (such as moving average or low-pass filtering) to remove sensor noise and instantaneous interference, and a smoothed pressure curve and a smoothed flow curve are obtained; the smoothed data are organized in chronological order to form the actual pressure-flow response curve, and aligned with the theoretical pressure-flow curve on the same time axis.

[0069] In one specific embodiment, the process of calculating instantaneous deviation and trend of change is as follows: at each time point, the difference between the actual pressure value and the theoretical pressure value is calculated to obtain the instantaneous pressure deviation; the difference between the actual flow rate value and the theoretical flow rate value is calculated to obtain the instantaneous flow deviation; the rate of change of pressure (the amount of pressure change per unit time) is calculated to obtain the pressure derivative; the rate of change of flow rate (the amount of flow change per unit time) is calculated to obtain the flow derivative; and the instantaneous deviation and derivative are used as the characteristic values ​​at the current moment.

[0070] In one specific embodiment, the anomaly type discrimination process based on thresholds and rules is as follows: The instantaneous deviation at the current moment is compared with a preset deviation tolerance. If the deviation is within the tolerance range, it is initially judged as a normal response; if it exceeds the tolerance, the anomaly discrimination process begins; the pressure derivative and flow derivative are compared with preset thresholds for sudden drop, rise, and increase to determine whether events such as sudden pressure drop, sudden pressure rise, or sudden flow increase have occurred; and the current characteristics are comprehensively matched in conjunction with geomechanical mechanism rules. If a sudden drop in pressure and a sudden increase in flow are detected, and the pressure has exceeded the critical fracturing pressure of the region (which can be obtained from the probability set of the third fracture network), then it is determined to be a hydraulic fracturing anomaly. If a sudden increase in pressure and a sharp drop in flow rate are detected (the negative derivative of the flow rate exceeds the threshold), it is determined to be an abnormal blockage. If the pressure remains below the theoretical value (negative deviation and persistent) and the flow rate remains above the theoretical value (positive deviation and persistent), and the cumulative grouting volume far exceeds expectations, then it is determined to be an abnormal grout leakage. If the pressure is stable but the flow rate is higher than the theoretical value, it is judged as an abnormal crack opening. If the pressure increases and the flow rate approaches zero, the filling is considered complete. For ambiguous situations, multiple features can be combined to calculate the confidence level of each anomaly type, and the type with the highest confidence level can be selected as the identification result.

[0071] Optionally, the identified abnormal response results can be recorded in a structured manner, including the abnormality type, occurrence time, current pressure, flow rate, pressure deviation, flow rate deviation, cumulative grouting volume, and confidence level (optional).

[0072] Optionally, if it is hydraulic fracturing, the possible location of occurrence (such as the stress concentration zone near the grouting section) can be inferred by combining the probability set of the third fracture network and added to the record.

[0073] This invention constructs an initial cognitive model by generating a first set of fracture network probabilities, quantifying uncertainty in the form of a probability set to provide a correctable cognitive foundation for subsequent dynamic updates. It generates a drilling strategy through deep reinforcement learning, formulating an optimal drilling plan based on the initial cognition, providing initial construction guidance, and reserving space for subsequent adjustments. By acquiring fracture observation data in real time during drilling, it perceives the actual rock mass information revealed by the drilling construction, thereby introducing real-time feedback, breaking the limitations of "static prediction" in existing technologies, and dynamically correcting the first fracture network probabilities through Bayesian updates based on the observation data. The invention employs a second-order fracture network probability set to enable real-time evolution of the model's understanding during construction, ensuring that construction strategies do not fail due to changes in fracture morphology caused by drilling. Real-time monitoring of grouting response data during grouting allows for the perception of rock mass response, providing real-time data for grouting control. Furthermore, a third-order fracture network probability set is used to predict fracture propagation and potential fracture changes induced by grouting pressure, thus considering grouting disturbances and dynamically adjusting grouting parameters to avoid secondary disasters such as hydraulic fracturing. This ensures that construction strategies do not fail due to changes in fracture morphology caused by grouting pressure. Compared to existing technologies, this invention achieves dynamic and precise treatment of seepage channels in fractured rock slopes, effectively solving the technical problem of static treatment strategies failing due to dynamic rock mass evolution caused by drilling and grouting disturbances.

[0074] Example 2: like Figure 2 As shown, this embodiment provides a deep learning-based seepage channel treatment system for fractured rock slopes, including a fracture network generation module 201, a borehole strategy generation module 202, a borehole processing module 203, and a grouting processing module 204. The fracture network generation module 201 is used to identify seepage channels based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, and generate a first fracture network probability set in combination with pre-stored fracture attitude statistics; wherein, the first fracture network probability set includes several fracture networks and the realization probability of each fracture network. In this embodiment, the fracture network generation module 201 identifies seepage channels based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, and generates a first fracture network probability set by combining it with pre-stored fracture attitude statistics. Specifically, the fracture network generation module 201 aligns the surface image data with the slope topographic data to obtain a multimodal image set, and inputs the multimodal image set into a pre-trained multi-scale attention segmentation network, so that the multi-scale attention segmentation network can generate a surface fracture vector map with seepage channel location markers by identifying surface fractures and seepage traces in the multimodal image set; based on the slope topographic data... According to the method, the elevation data corresponding to each fracture point in the surface fracture vector map is obtained, and a three-dimensional surface fracture network is generated based on the elevation difference of the elevation data. According to the pre-stored fracture attitude statistics, the three-dimensional surface fracture network is subjected to underground simulation to generate several underground fracture networks with different realization probabilities. The underground simulation includes fracture dip angle simulation, fracture depth simulation, and realization probability simulation. The three-dimensional surface fracture network is combined with each of the underground fracture networks to construct several fracture networks with heterogeneous map structures, and the realization probability of the underground fracture network is determined as the realization probability of the corresponding fracture network to integrate all fracture networks and form a first fracture network probability set.

[0075] The drilling strategy generation module 202 is used to generate a drilling strategy based on the first fracture network probability set, with the goal of maximizing information gain and maximizing governance potential, through deep reinforcement learning; wherein, the drilling strategy includes drilling sequence, drilling location and drilling parameters; The drilling processing module 203 is used to acquire fracture observation data in real time during the drilling process of the drilling equipment executing the drilling strategy, and dynamically correct the pre-acquired second fracture network probability set according to the fracture observation data using a Bayesian update method, and adjust the drilling strategy in real time according to the corrected second fracture network probability set until the seepage channel drilling processing of the target rock slope is completed; wherein, the second fracture network probability set is obtained by predicting fracture propagation of the first fracture network probability set according to the currently implemented drilling strategy; In this embodiment, the drilling processing module 203 dynamically corrects the pre-acquired second fracture network probability set based on the fracture observation data using a Bayesian update method, so as to adjust the drilling strategy in real time according to the corrected second fracture network probability set. Specifically, the drilling processing module 203, based on the drilling strategy currently implemented by the drilling equipment, performs fracture propagation prediction on the first fracture network probability set using a pre-trained second physical information neural network model to obtain the second fracture network probability set; based on the fracture observation data, it extracts the fracture features of the currently observed fractures, and calculates the fracture characteristics according to the fracture features. Calculate the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture; wherein the fracture features include fracture depth, fracture orientation, and fracture aperture; based on the likelihood, update the realization probability of each fracture network in the second fracture network probability set using a preset Bayesian formula to obtain a corrected second fracture network probability set; based on the corrected second fracture network probability set, perform expected information gain assessment and governance potential assessment on the unexecuted boreholes in the borehole strategy, so as to comprehensively adjust the borehole strategy in real time by combining the expected information gain assessment results and governance potential assessment results of each unexecuted borehole.

[0076] In this embodiment, the drilling processing module 203, based on the drilling strategy currently implemented by the drilling equipment, performs crack propagation prediction on the first crack network probability set using a pre-trained second physical information neural network model to obtain a second crack network probability set. Specifically, the drilling processing module 203 constructs a drilling disturbance source function according to the drilling strategy currently implemented by the drilling equipment, and determines the drilling disturbance source function as the boundary condition of the second physical information neural network model; through a pre-constructed numerical inversion model, it performs stress field inversion on each crack network in the first crack network probability set to obtain the first stress field of each crack network; the boundary condition, the first stress field of each crack network, and the first crack network probability set are input into the pre-trained second physical information neural network model, so that the second physical information neural network model updates the first stress field according to the preset elasticity equilibrium equation and fracture mechanics criterion to obtain the second stress field of each crack network in the first crack network probability set; based on the second stress field, it performs crack propagation prediction on each crack network in the first crack network probability set to obtain the second crack network probability set.

[0077] In this embodiment, the drilling processing module 203 extracts the fracture features of the currently observed fracture based on the fracture observation data, and calculates the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture based on the fracture features. Specifically, the drilling processing module 203 extracts the fracture features of the currently observed fracture based on the fracture observation data to obtain an actual feature list of the currently observed fracture; according to the drilling strategy currently implemented by the drilling equipment, it performs spatial geometric intersection calculations in each fracture network of the second fracture network probability set to determine the fracture segments in each fracture network that intersect with the current drilling trajectory, and extracts the fracture features of each fracture segment to obtain a predicted feature list of each fracture network; based on the actual feature list and the predicted feature list of each fracture network, it calculates the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture.

[0078] In this embodiment, the borehole processing module 203 performs expected information gain assessment and governance potential assessment on the unexecuted boreholes in the borehole strategy based on the modified second fracture network probability set. Specifically, for each unexecuted borehole in the borehole strategy, the borehole processing module 203 performs spatial geometric intersection calculations in each fracture network in the modified fracture network probability set according to the corresponding borehole location and borehole parameters. This determines the fracture segments in each fracture network that intersect with the preset trajectories of each unexecuted borehole, and extracts the fracture features of each fracture segment to obtain the information gain of each unexecuted borehole. A list of predicted fracture features is generated. Based on the predicted fracture feature list and the modified second fracture network probability set, the expected information gain for each un-drilled borehole is calculated using Monte Carlo simulation to obtain the expected information gain evaluation result for each un-drilled borehole. Based on the realization probability of each fracture network in the predicted fracture feature list and the modified second fracture network probability set, and a preset graph theory index, the remediation potential of each un-drilled borehole is evaluated to obtain the remediation potential evaluation result for each un-drilled borehole. The graph theory index includes betweenness centrality index and seepage centrality index.

[0079] The grouting module 204 is used to monitor the grouting response data in real time during the grouting process of the grouting equipment, and combine it with the pre-acquired third fracture network probability set to predict fracture propagation through a pre-trained first physical information neural network, so as to dynamically adjust the grouting parameters until the grouting treatment of the seepage channel of the target rock slope is completed; wherein, the third fracture network probability set is the most recently corrected second fracture network probability set.

[0080] In this embodiment, the grouting processing module 204 monitors the grouting response data in real time and, in conjunction with the pre-acquired third fracture network probability set, performs fracture propagation prediction through a pre-trained first physical information neural network to dynamically adjust the grouting parameters. Specifically, the grouting processing module 204 inputs the current grouting parameters of the grouting equipment and the third fracture network probability set into the pre-trained first physical information neural network, so that the first physical information neural network outputs fracture propagation prediction results. The grouting response data includes grouting pressure and grouting flow rate; the fracture propagation prediction results include grout diffusion area, new fracture risk value, and theoretical pressure-flow rate curve; an actual pressure-flow rate response curve is generated based on the grouting response data, and abnormal responses during the grouting process are identified by comparing the actual pressure-flow rate response curve with the theoretical pressure-flow rate curve, obtaining abnormal response results, and the grouting parameters are dynamically adjusted based on the abnormal response results; the abnormal responses include hydraulic fracturing, blockage, and grout leakage.

[0081] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that 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 for those skilled in the art.

Claims

1. A method for treating seepage channels in fractured rock slopes based on deep learning, characterized in that, include: Based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, seepage channels are identified, and combined with pre-stored fracture attitude statistics, a first fracture network probability set is generated; wherein, the first fracture network probability set includes several fracture networks and the realization probability of each fracture network. Based on the first set of probability of the fracture network, with the goal of maximizing information gain and maximizing governance potential, a drilling strategy is generated through deep reinforcement learning; wherein, the drilling strategy includes drilling sequence, drilling location and drilling parameters; During the drilling process, the drilling equipment acquires fracture observation data in real time. Based on the fracture observation data, the pre-acquired second fracture network probability set is dynamically corrected using a Bayesian update method. The drilling strategy is then adjusted in real time based on the corrected second fracture network probability set until the seepage channel drilling of the target rock slope is completed. The second fracture network probability set is obtained by predicting fracture propagation based on the first fracture network probability set using the currently implemented drilling strategy. During the grouting process, the grouting response data is monitored in real time. Combined with the pre-acquired third fracture network probability set, the fracture propagation is predicted through a pre-trained first physical information neural network to dynamically adjust the grouting parameters until the grouting treatment of the seepage channel of the target rock slope is completed. The third fracture network probability set is the most recently corrected second fracture network probability set.

2. The method for treating seepage channels in fractured rock slopes based on deep learning as described in claim 1, characterized in that, Based on the fracture observation data, the pre-acquired second fracture network probability set is dynamically corrected using a Bayesian update method. This allows for real-time adjustment of the drilling strategy based on the corrected second fracture network probability set. Specifically: Based on the drilling strategy currently implemented by the drilling equipment, the first fracture network probability set is predicted to expand by a pre-trained second physical information neural network model, thereby obtaining the second fracture network probability set. Based on the fracture observation data, the fracture features of the currently observed fracture are extracted, and based on the fracture features, the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture is calculated; wherein, the fracture features include fracture depth, fracture orientation, and fracture aperture. Based on the likelihood, the realization probability of each fracture network in the second fracture network probability set is updated by a preset Bayesian formula to obtain the corrected second fracture network probability set. Based on the modified second fracture network probability set, the expected information gain and governance potential of the un-drilled holes in the drilling strategy are evaluated respectively. The drilling strategy is then adjusted in real time by combining the expected information gain evaluation results and governance potential evaluation results of each un-drilled hole.

3. The method for treating seepage channels in fractured rock slopes based on deep learning as described in claim 2, characterized in that, Based on the drilling strategy currently implemented by the drilling equipment, a second fracture network probability set is obtained by predicting fracture propagation on the first fracture network probability set using a pre-trained second physical information neural network model, specifically as follows: Based on the drilling strategy currently implemented by the drilling equipment, a drilling disturbance source function is constructed, and the drilling disturbance source function is determined as the boundary condition of the second physical information neural network model; By using a pre-constructed numerical inversion model, stress field inversion is performed on each fracture network in the first fracture network probability set to obtain the first stress field of each fracture network. The boundary conditions, the first stress fields of each fracture network, and the first fracture network probability set are input into a pre-trained second physical information neural network model so that the second physical information neural network model updates the first stress field according to the preset elasticity equilibrium equation and fracture mechanics criterion, thereby obtaining the second stress field of each fracture network in the first fracture network probability set. Based on the second stress field, crack propagation prediction is performed on each crack network in the first crack network probability set to obtain the second crack network probability set.

4. The method for treating seepage channels in fractured rock slopes based on deep learning as described in claim 2, characterized in that, Based on the fracture observation data, the fracture features of the currently observed fracture are extracted, and based on the fracture features, the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture is calculated, specifically as follows: Based on the fracture observation data, the fracture features of the currently observed fracture are extracted to obtain a list of actual features of the currently observed fracture; Based on the drilling strategy currently implemented by the drilling equipment, spatial geometric intersection calculations are performed in each fracture network of the second fracture network probability set to determine the fracture segments in each fracture network that intersect with the current drilling trajectory, and the fracture features of each fracture segment are extracted to obtain a list of predicted features for each fracture network. Based on the actual feature list and the predicted feature list of each fracture network, the likelihood of each fracture network in the second fracture network probability set relative to the currently observed fracture is calculated.

5. The method for treating seepage channels in fractured rock slopes based on deep learning as described in claim 2, characterized in that, Based on the modified second fracture network probability set, the expected information gain and governance potential of the un-drilled holes in the drilling strategy are evaluated, specifically as follows: For each undrilled hole in the drilling strategy, based on the corresponding drilling location and drilling parameters, spatial geometric intersection calculation is performed in each fracture network in the modified fracture network probability set to determine the fracture segment in each fracture network that intersects with the preset trajectory of each undrilled hole, and the fracture features of each fracture segment are extracted to obtain a list of predicted fracture features for each undrilled hole. Based on the predicted fracture feature list and the corrected second fracture network probability set, the expected information gain for each un-drilled hole is calculated using Monte Carlo simulation to obtain the expected information gain evaluation results for each un-drilled hole. Based on the predicted fracture feature list, the realization probability of each fracture network in the modified second fracture network probability set, and the preset graph theory indices, the remediation potential of each un-drilled borehole is evaluated to obtain the remediation potential evaluation result corresponding to each un-drilled borehole; wherein, the graph theory indices include betweenness centrality indices and seepage centrality indices.

6. The method for treating seepage channels in fractured rock slopes based on deep learning as described in claim 1, characterized in that, The real-time monitoring of grouting response data, combined with the pre-acquired third fracture network probability set, is used to predict fracture propagation through a pre-trained first physical information neural network, so as to dynamically adjust the grouting parameters. Specifically: The current grouting parameters of the grouting equipment and the probability set of the third fracture network are input into a pre-trained first physical information neural network, so that the first physical information neural network outputs a fracture propagation prediction result; wherein, the grouting response data includes grouting pressure and grouting flow rate; the fracture propagation prediction result includes grout diffusion area, new fracture risk value and theoretical pressure-flow rate curve; An actual pressure-flow response curve is generated based on the grouting response data. By comparing the actual pressure-flow response curve with the theoretical pressure-flow curve, abnormal responses during the grouting process are identified, and abnormal response results are obtained. Grouting parameters are then dynamically adjusted based on the abnormal response results. The abnormal responses include hydraulic fracturing, blockage, and grout leakage.

7. The method for treating seepage channels in fractured rock slopes based on deep learning as described in claim 6, characterized in that, An actual pressure-flow response curve is generated based on the grouting response data. By comparing the actual pressure-flow response curve with the theoretical pressure-flow curve, abnormal responses during the grouting process are identified, and abnormal response results are obtained. Specifically: The grouting pressure and grouting flow rate are filtered and smoothed to obtain the actual pressure-flow rate response curve, and the instantaneous deviation and derivative value between the actual pressure-flow rate response curve and the theoretical pressure-flow rate curve are calculated. Based on the instantaneous deviation value, derivative value, and preset anomaly judgment threshold, combined with preset geomechanical mechanism rules, abnormal responses during the grouting process are identified, and abnormal response results are obtained.

8. The method for treating seepage channels in fractured rock slopes based on deep learning as described in claim 1, characterized in that, The process involves identifying seepage channels based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, and then generating a first fracture network probability set by combining it with pre-stored fracture attitude statistics. Specifically: The surface image data and the slope topography data are aligned to obtain a multimodal image set. The multimodal image set is then input into a pre-trained multi-scale attention segmentation network so that the multi-scale attention segmentation network can generate a surface fissure vector map with seepage channel location markers by identifying surface fissures and seepage traces in the multimodal image set. Based on the slope topography data, the elevation data corresponding to each fracture point in the surface fracture vector map is obtained, and a three-dimensional surface fracture network is generated based on the elevation data and the elevation difference. Based on pre-stored fracture attitude statistics, the three-dimensional surface fracture network is subjected to subsurface simulation to generate several subsurface fracture networks with different realization probabilities; wherein, the subsurface simulation includes fracture dip angle simulation, fracture depth simulation, and realization probability simulation. The three-dimensional surface fracture network is combined with each of the underground fracture networks to construct several fracture networks with heterogeneous graph structures. The realization probability of the underground fracture network is determined as the realization probability of the corresponding fracture network, so as to integrate all fracture networks and form a first fracture network probability set.

9. The method for treating seepage channels in fractured rock slopes based on deep learning as described in claim 8, characterized in that, The process involves performing subsurface extrapolation on the three-dimensional surface fracture network based on pre-stored fracture attitude statistics to generate several subsurface fracture networks with different realization probabilities. Specifically: Extract the node features of each surface node in the three-dimensional surface fracture network, and combine them with the pre-stored fracture attitude statistics and the node features to infer the probability distribution of the first attitude of underground fractures at each surface node; wherein, the node features include the orientation features and seepage marker features of the connected fractures; The Monte Carlo method is used to sample multiple times from the first attitude probability distribution to generate several sets of three-dimensional coordinates of underground nodes. A pre-trained graph neural network is then connected to a predictor to infer the connection relationship of each set of underground nodes based on the node features and the three-dimensional coordinates, thereby generating underground fracture edges connecting each set of underground nodes to construct several initial underground fracture networks. The initial subsurface fracture network is statistically analyzed to obtain the second attitude probability distribution of each initial subsurface fracture network, and pre-stored regional attitude statistics are obtained to calculate the similarity between each second attitude probability distribution and the regional attitude statistics. The realization probability of each initial underground fracture network is determined based on the similarity, resulting in several final underground fracture networks.

10. A deep learning-based system for treating seepage channels in fractured rock slopes, characterized in that, It includes a fracture network generation module, a borehole strategy generation module, a borehole processing module, and a grouting processing module, among which, The fracture network generation module is used to identify seepage channels based on real-time acquired surface image data of the target rock slope and pre-stored slope topographic data, and generate a first fracture network probability set by combining pre-stored fracture attitude statistics data; wherein, the first fracture network probability set includes several fracture networks and the realization probability of each fracture network. The drilling strategy generation module is used to generate a drilling strategy based on the first fracture network probability set, with the goal of maximizing information gain and maximizing governance potential, through deep reinforcement learning; wherein, the drilling strategy includes drilling sequence, drilling location and drilling parameters; The drilling processing module is used to acquire fracture observation data in real time during the drilling process of the drilling equipment executing the drilling strategy, and dynamically correct the pre-acquired second fracture network probability set according to the fracture observation data using a Bayesian update method. Based on the corrected second fracture network probability set, the drilling strategy is adjusted in real time until the seepage channel drilling processing of the target rock slope is completed. The second fracture network probability set is obtained by predicting fracture propagation from the first fracture network probability set based on the currently implemented drilling strategy. The grouting module is used to monitor grouting response data in real time during the grouting process of the grouting equipment, and combine it with the pre-acquired third fracture network probability set to predict fracture propagation through a pre-trained first physical information neural network, so as to dynamically adjust the grouting parameters until the grouting treatment of the seepage channel of the target rock slope is completed; wherein, the third fracture network probability set is the most recently corrected second fracture network probability set.

Images