Granite residual soil slope evaluation method and system based on machine learning

Abstract

The invention relates to a granite residual soil slope evaluation method and system based on machine learning. According to the method, a standardized data set is constructed by collecting slope multi-source data, and a dynamic graph structure representing a fracture network with nodes and edges is established; simulating seepage-stress feedback for the first time by using a message passing mechanism of a graph neural network, and outputting predicted displacement and a water head; on the basis, fracture expansion behaviors are predicted, hydraulic parameters are updated, secondary feedback simulation is carried out after a graph structure is adjusted, and more accurate node features are obtained; recognizing a seepage front surface according to the secondarily updated water head data, calculating a Gaussian curvature and a clustering coefficient, and comparing the Gaussian curvature and the clustering coefficient with a critical value to realize self-adaptive discrimination of a failure mode; finally, the failure mode and the displacement data are combined to calculate the safety coefficient and trigger early warning, accurate tracking of fracture dynamic evolution is achieved, the numerical value divergence problem of a traditional method is solved, and the accuracy of slope stability evaluation and the early warning reliability are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of geotechnical engineering, and particularly relates to a method and system for evaluating a granite residual soil slope based on machine learning. BACKGROUND

[0002] Granite residual soil slopes are widely distributed in the southeast coastal areas of China, and their stability problems have long plagued engineering construction and safe operation. Such slopes are prone to instability and failure under the action of rainfall infiltration, and the core mechanism lies in the control of the internal developed in-situ fracture network on the seepage path and stress distribution.

[0003] The existing evaluation methods mainly face three technical bottlenecks: first, traditional numerical simulation usually simplifies the fracture as a static geometric entity, ignoring the dynamic evolution characteristics of the actual fracture under dry-wet cycles, especially the non-monotonic evolution of the hydraulic opening caused by the change of the roughness coefficient of the fracture with particle peeling, which makes the simulation results based on fixed parameters deviate seriously from the actual situation. Second, when simulating the coupling of seepage and stress, the change of pore water pressure will cause the change of effective stress, which in turn triggers the expansion of the fracture and changes the permeability, forming a strong positive feedback cycle. The traditional finite element method often fails to converge when solving such strong nonlinear problems due to the ill-conditioned Jacobian matrix. Third, the existing methods for identifying the failure mode of the slope mostly rely on fixed depth thresholds or simplified empirical criteria, which cannot effectively capture the dynamic influence of the connectivity change of the fracture network on the failure mechanism, resulting in large deviations in identifying different failure modes such as shallow fluidization and deep sliding. These limitations make it difficult for the existing technology to accurately predict the whole process from gradual failure to final instability of the slope, posing a serious challenge to the disaster prevention and mitigation of slope engineering.

[0004] Therefore, there is an urgent need to develop a new evaluation method that can fully consider the dynamic characteristics of the fracture, effectively solve the numerical calculation stability, and adaptively identify the failure mode. SUMMARY

[0005] Therefore, it is necessary to provide a method and system for evaluating a granite residual soil slope based on machine learning in view of the above technical problems.

[0006] In a first aspect, the application provides a method for evaluating a granite residual soil slope based on machine learning, comprising:

[0007] S1, collecting fracture geometry data, hydrological data and mechanical data of the slope to obtain slope state data; and performing normalization and interpolation preprocessing on the slope state data to generate a standardized data set;

[0008] S2, represent the fracture network as a dynamic graph structure based on the standardized dataset; wherein in the dynamic graph structure, nodes correspond to fracture intersection points or end points, edges correspond to fracture segments, node features include normal stress, strain energy density and water head, edge features include mechanical opening degree, roughness coefficient and length, and edge weights are calculated based on hydraulic opening degree;

[0009] S3, input the dynamic graph structure into a graph neural network, simulate the bidirectional feedback of seepage and stress through a message passing mechanism, and output the first updated node features; wherein the node features include predicted displacement and predicted water head;

[0010] S4, based on the first updated node features, predict the expansion behavior of the fractures to obtain a prediction result; update the hydraulic opening degree and roughness coefficient of the fractures according to the prediction result to adjust the edge features and edge weights of the dynamic graph structure, and obtain an adjusted graph structure;

[0011] S5, input the adjusted graph structure into the graph neural network, simulate the bidirectional feedback of seepage and stress through the message passing mechanism, and output the second updated node features;

[0012] S6, based on the predicted water head in the second updated node features, identify the seepage front and calculate the Gaussian curvature of the seepage front and the clustering coefficient of the fracture network; according to the comparison of the Gaussian curvature and the clustering coefficient with the critical value, determine the failure mode of the slope;

[0013] S7, based on the failure mode and the predicted displacement in the second updated node features, calculate the safety factor of the slope, and trigger an early warning when the safety factor is lower than a threshold.

[0014] In a second aspect, the application also provides a machine learning-based granite residual soil slope evaluation system for implementing the method of the first aspect, which comprises:

[0015] A slope data standardization module for collecting fracture geometric data, hydrological data and mechanical data of the slope to obtain slope state data; and normalizing and interpolating the slope state data for pretreatment to generate a standardized dataset;

[0016] A fracture network topology modeling module for representing the fracture network as a dynamic graph structure based on the standardized dataset; wherein in the dynamic graph structure, nodes correspond to fracture intersection points or end points, edges correspond to fracture segments, node features include normal stress, strain energy density and water head, edge features include mechanical opening degree, roughness coefficient and length, and edge weights are calculated based on hydraulic opening degree;

[0017] The seepage stress coupling simulation module is configured to input a dynamic graph structure into a graph neural network, simulate bidirectional feedback between seepage and stress through a message passing mechanism, and output first updated node features; wherein the node features include predicted displacements and predicted water heads.

[0018] The graph structure adaptive adjustment module is configured to predict an expansion behavior of the fissure based on the first updated node features, obtain a prediction result, update hydraulic openings and roughness coefficients of the fissure according to the prediction result, adjust edge features and edge weights of the dynamic graph structure, and obtain an adjusted graph structure.

[0019] The multi-round feature iteration optimization module is configured to input the adjusted graph structure into the graph neural network, simulate bidirectional feedback between seepage and stress through the message passing mechanism, and output second updated node features.

[0020] The failure mode intelligent identification module is configured to identify a seepage front based on the predicted water heads in the second updated node features, calculate a Gaussian curvature of the seepage front and a clustering coefficient of the fissure network, and determine a failure mode of the slope according to a comparison between the Gaussian curvature and the clustering coefficient and a critical value.

[0021] The slope safety early warning decision module is configured to calculate a safety coefficient of the slope based on the failure mode and the predicted displacements in the second updated node features, and trigger a warning when the safety coefficient is lower than a threshold.

[0022] In a third aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the method for evaluating a granite residual soil slope based on machine learning according to the first aspect when executing the computer program.

[0023] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method for evaluating a granite residual soil slope based on machine learning according to the first aspect.

[0024] The machine learning-based granite residual soil slope evaluation method and system described above collects the crack geometry, hydrology and mechanics data of the slope and generates a standardized data set after preprocessing, then constructs a dynamic graph structure represented by nodes and edges to depict the crack network, simulates the two-way feedback of seepage and stress using the message passing mechanism of the graph neural network, and first outputs the node features containing the predicted displacement and predicted water head; based on this, the crack propagation behavior is predicted and the hydraulic opening and roughness coefficient are updated to adjust the dynamic graph structure, and then the adjusted graph structure is input into the graph neural network for secondary feedback simulation to output more accurate node features; then the seepage front is identified based on the second updated predicted water head and the Gaussian curvature and crack network clustering coefficient are calculated, and the slope failure mode is adaptively judged by comparing the critical value; finally, the safety factor is calculated in combination with the failure mode and the predicted displacement to trigger the warning, thereby realizing high-precision tracking of crack dynamic evolution, overcoming the numerical divergence problem of traditional numerical simulation, and significantly improving the accuracy and reliability of slope stability evaluation and warning. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0026] Figure 1 A flowchart of a machine learning-based granite residual soil slope evaluation method provided by the present application is shown in the figure.

[0027] Figure 2 A flowchart of constructing a dynamic graph structure in an alternative embodiment of the present application is shown in the figure.

[0028] Figure 3 A structural diagram of a machine learning-based granite residual soil slope evaluation system provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0030] REFERENCE Figure 1 A flowchart of a machine learning-based granite residual soil slope evaluation method provided by the present application is shown in the figure. The method includes the following steps:

[0031] S1, collect the crack geometry data, hydrological data and mechanical data of the slope to obtain the slope state data; and normalize and interpolate the slope state data for pretreatment to generate a standardized data set.

[0032] Specifically, the crack geometry data is used to describe the spatial distribution and morphological characteristics of the internal cracks of the slope, covering key information such as the strike, dip angle and extension length of the cracks. These data are the basis for subsequent construction of a crack network model; the hydrological data focus on the water movement and distribution state of the slope, including the pore water pressure and groundwater level change at different positions, which provide a basis for analyzing seepage action; the mechanical data cover the strength parameters and deformation characteristics of the slope soil, which are key to evaluating the stress state and stability of the slope. After the collection of the three types of data is completed, normalization processing is performed to eliminate the interference of the dimensional differences of different physical quantities on subsequent calculations. The normalization process uses the following formula:

[0033]

[0034] In the formula, denotes the normalized processed data, denotes the original slope state data, denotes the mean value of the original data of this type, denotes the standard deviation of the original data of this type. Through this formula, the original data of different magnitudes can be converted to the same numerical interval, avoiding deviations in subsequent model training or calculations due to differences in data magnitudes. At the same time, considering the possibility of missing data in some areas during actual data collection, an interpolation method is used to complete the missing data. The interpolation processing is based on the spatial distribution law and correlation of the data to ensure that the completed data can truly reflect the continuity and rationality of the slope state. After normalization and interpolation pretreatment, a standardized data set is finally generated. This data set must have integrity, consistency and accuracy to provide a reliable input basis for subsequent steps.

[0035] S2, based on the standardized data set, represent the crack network as a dynamic graph structure; in the dynamic graph structure, the nodes correspond to crack intersection points or endpoints, the edges correspond to crack segments, the node features include normal stress, strain energy density and water head, the edge features include mechanical opening, roughness coefficient and length, and the edge weight is calculated based on the hydraulic opening.

[0036] Specifically, this step abstracts the physical fissure network into a mathematical dynamic graph structure, achieving accurate characterization of the topological relationship and multi-field characteristics of the fissure network. Based on the fissure geometry data in the standardized dataset, first, the nodes and edges of the dynamic graph structure are determined. The node is defined as the intersection or end point in the fissure network. The intersection refers to the position where two or more fissures intersect with each other, and the end point refers to the starting or ending position of a single fissure. The connection relationship of the fissure network can be clearly reflected through the setting of the node. The edge corresponds to the fissure segment between two adjacent nodes. Each edge uniquely corresponds to a specific fissure entity, and the existence of the edge represents the physical connection between the nodes.

[0037] In the dynamic graph structure, both nodes and edges are assigned specific characteristic parameters to reflect their physical properties. The node characteristics include normal stress, strain energy density, and water head. The normal stress refers to the stress value perpendicular to the fissure surface, reflecting the vertical force acting on the fissure surface. The strain energy density refers to the strain energy stored in the soil around the fissure per unit volume, which can be used to evaluate the deformation potential of the soil. The water head refers to the energy possessed by a unit weight of water, reflecting the pressure state of the water in the fissure. The edge characteristics include mechanical opening, roughness coefficient, and length. The mechanical opening refers to the actual distance between the fissure walls, directly affecting the seepage capacity of the fissure. The roughness coefficient is a parameter that describes the roughness of the fissure wall, which has a significant impact on seepage resistance and stress transmission. The length refers to the actual spatial length of the fissure segment corresponding to the edge, which is used to quantify the physical size of the fissure segment.

[0038] The calculation of edge weight is based on the hydraulic opening, which is a key parameter that comprehensively reflects the seepage characteristics of the fissure. The calculation of the hydraulic opening needs to consider the mechanical opening, roughness coefficient, and stress state in the edge characteristics. The specific formula is as follows:

[0039]

[0040] In the formula, represents the hydraulic opening, represents the mechanical opening, represents the stress influence coefficient, represents the normal stress of the node, represents the compressive strength of the granite residual soil, represents the roughness coefficient. After calculating the hydraulic opening through the formula, the edge weight is further determined by combining the hydraulic conductivity coefficient. The formula for calculating the hydraulic conductivity coefficient is as follows:

[0041]

[0042] In the formula, represents the hydraulic conductivity coefficient, represents the gravitational acceleration, represents the dynamic viscosity of water, denotes the length of the edge. The edge weight is the normalized result of the hydraulic conductivity coefficient, which is used to reflect the contribution difference of different fracture segments in the seepage process, and ensures that the dynamic graph structure can accurately reflect the seepage characteristics of the fracture network.

[0043] S3, input the dynamic graph structure into the graph neural network, simulate the bidirectional feedback of seepage and stress through the message passing mechanism, and output the first updated node feature; wherein the node feature includes the predicted displacement and the predicted water head.

[0044] Specifically, the core of this step is to convert the physical coupling relationship between the seepage field and the stress field into the information interaction process of nodes and edges in the graph structure through the message passing mechanism of the graph neural network, and then accurately simulate the bidirectional feedback effect of the two. The physical nature of seepage and stress bidirectional feedback is: on the one hand, the change of water head in the seepage field will change the pore water pressure in the fracture. According to the principle of effective stress, the effective stress is equal to the total stress minus the pore water pressure. The increase of pore water pressure will lead to the decrease of effective stress, and then the normal stress of fracture wall surface decreases and the mechanical opening degree of fracture increases, which ultimately improves the hydraulic conductivity of fracture (feedback of seepage to stress); on the other hand, the change of normal stress in the stress field will directly change the mechanical opening degree of the fracture. The increase or decrease of the opening degree will affect the hydraulic opening degree, and then change the hydraulic conductivity coefficient of the seepage path, resulting in the redistribution of water head (feedback of stress to seepage). The message passing mechanism encodes this physical process into the update logic of node features through mathematical model, rather than simply aggregating information.

[0045] After the dynamic graph structure is input into the graph neural network, the message passing process is divided into three core stages: message generation (MSG function), message interaction (bidirectional feedback encoding), and message aggregation (AGG function). Each stage needs to closely correspond to the physical mechanism of seepage-stress coupling. First, in the message generation stage, the MSG function extracts the seepage-related parameters (hydraulic opening degree, hydraulic conductivity coefficient) and stress-related parameters (mechanical opening degree, normal stress) in the edge features, and combines the water head (seepage field index) and normal stress, strain energy density (stress field index) in the node features to generate coupled messages containing seepage information and stress information. Specifically, for the edge connecting node and neighbor node , the mathematical expression of MSG function is:

[0046]

[0047] In the formula, denotes the initial feature vector of node , is the water head of node , is the normal stress of node , For nodes (strain energy density) For neighboring nodes eigenvectors, Representing an edge eigenvectors ( For the edge hydraulic opening, For the edge The hydraulic conductivity coefficient, For the edge Roughness coefficient); The weight of the percolation message is determined by the edge. hydraulic conductivity Normalization yields ( ), used to quantify the transmission intensity of seepage information, head difference This reflects the seepage direction and potential energy gradient; The stress message weights are determined by the edges. The mechanical opening (negatively correlated with normal stress) is normalized to obtain ( , (Mechanical opening), used to quantify the transmission intensity of stress information, normal stress difference. It reflects the direction and magnitude of stress transmission; The coupling coefficient is determined by the edge. roughness coefficient With hydraulic opening jointly determined ( This is used to characterize the degree of coupling between seepage and stress. The term represents the average hydraulic opening of all edges in the dynamic graph. This reflects the degree to which the hydraulic opening of the current side deviates from the mean, and thus reflects the difference in the contribution of that side to the two-way feedback.

[0048] After message generation, the message interaction phase begins. This phase is crucial for simulating bidirectional feedback and requires the mutual modulation of seepage and stress messages. Specifically, the head difference signal in the seepage message affects the stress message through the following logic: when (Water from nodes) Towards During seepage, nodes The pore water pressure increases, and according to the effective stress principle, its effective normal stress... ( The pore water pressure is determined by the water head. Converted to , The density of water will decrease, at which point the normal stress difference in the stress message will decrease. Make corrections, the correction formula is as follows ,in (Nodes caused by seepage) The pore water pressure increment reflects the feedback of seepage stress; simultaneously, the normal stress difference signal in the stress message affects the seepage message through the following logic: when (Stress from nodes) Towards During transmission, the node Increased normal stress will compress the crack, leading to increased mechanical opening. Reduce, thereby reducing the hydraulic opening. Reduce (according to the hydraulic opening formula in S2) Normal stress When it increases (Reduce), at this time the relevant weights of the hydraulic conductivity coefficient in the seepage message need to be adjusted. Make corrections, the correction formula is as follows ,in The stress influence coefficient (calibrated by indoor tests, reflecting the attenuation rate of normal stress on the hydraulic conductivity coefficient) is used to represent the feedback of stress on seepage. Through this mutual modulation, the message interaction phase transforms the two-way feedback at the physical level into dynamic correction of the message signal, ensuring that the subsequently aggregated messages can truly reflect the coupling effect between the two fields.

[0049] After completing message interaction, the message aggregation phase begins, where the AGG function aggregates the nodes. The corrected messages transmitted by all neighboring nodes are integrated and combined with the node's... The initial feature fusion is used to update node features. Considering that messages from different neighboring nodes have different weights on the current node (e.g., seepage messages transmitted by edges with larger hydraulic conductivity coefficients have higher weights), the AGG function adopts a combination of weighted summation and nonlinear activation. The specific formula is as follows:

[0050]

[0051] In the formula, For the node after the first update Feature vector This is the coupled message after correction during the message interaction phase. For aggregate weights, by edge Edge weights (calculated based on hydraulic opening degree) and stress message weights Jointly determined (e.g.) The weight allocation is optimized through model training to ensure a balanced contribution of seepage and stress information. is an aggregation matrix (dimension [feature dimension x feature dimension]) used for linear transformation of the fusion result of the initial features and the aggregation message, parameters of which are optimized through a training data set (containing historical slope monitoring data and instability cases) to ensure that the transformed features can be mapped to a prediction index with clear physical meaning; is a bias vector used to adjust the feature distribution; is a nonlinear activation function, and ReLU function is selected to introduce nonlinear mapping capability to simulate nonlinear effects in the seepage-stress coupling process (such as the nonlinear relationship between fracture aperture change and stress).

[0052] After the complete process of the above message generation, interaction and aggregation, the first updated node feature output by the graph neural network contains clear physical indicators, i.e. predicted displacement and predicted water head. The calculation of the predicted displacement is by substituting the strain energy density component in the updated node feature vector into the elastic mechanics displacement formula to derive: , wherein is the predicted displacement of the node , is the elastic modulus of the granite residual soil, is the Poisson's ratio, and the formula is derived based on Hooke's law to convert the strain energy density into displacement amplitude, reflecting the deformation of the slope caused by the change of the stress field; the predicted water head directly corresponds to the water head component in the updated node feature vector , which is the water head value obtained after the bidirectional feedback modulation of seepage-stress, reflecting the redistribution result of the seepage field under the influence of stress. At this point, the message passing mechanism combines physical process coding and mathematical model to completely simulate the bidirectional feedback of seepage and stress, and outputs node features with clear physical meaning, laying a foundation for subsequent crack propagation prediction.

[0053] S4, based on the first updated node feature, the propagation behavior of the crack is predicted to obtain a prediction result; the hydraulic aperture and roughness coefficient of the crack are updated according to the prediction result to adjust the edge feature and edge weight of the dynamic graph structure, and an adjusted graph structure is obtained.

[0054] Specifically, based on the first updated node features, the prediction of crack propagation behavior and the adjustment of dynamic graph structure are realized to reflect the dynamic changes of crack network under seepage-stress action. Firstly, the first updated node features include predicted displacement and predicted water head, which can reflect the changes of stress state and seepage state inside the slope, and these changes are the key factors triggering crack propagation: the change of predicted displacement reflects the deformation accumulation of soil, and when the deformation exceeds a certain threshold, it may lead to the initiation and propagation of cracks; the change of predicted water head reflects the change of pore water pressure, and the increase of water pressure will reduce the effective stress of crack surface, thereby promoting the propagation of cracks.

[0055] Based on the above node features, the method combining mechanical analysis and statistical learning is used to predict the crack propagation behavior. By analyzing the predicted displacement gradient and the change rate of predicted water head, the crack segment that may occur expansion is determined; combined with the fracture mechanics parameters of granite residual soil, the direction and length of crack propagation are judged, and finally the prediction result of crack propagation behavior is obtained, which clearly indicates the expanded crack segment, expansion amount and the shape characteristics of the expanded crack.

[0056] According to the prediction result of crack propagation behavior, the hydraulic opening and roughness coefficient of the crack are updated: for the crack segment that occurs expansion, its mechanical opening will increase with the increase of expansion amount, which will further lead to the change of hydraulic opening. The update of hydraulic opening is still based on the hydraulic opening calculation formula in S2 step, but the mechanical opening is replaced by the expanded mechanical opening; at the same time, the wall surface particles may be stripped or worn during the crack propagation process, which will change the roughness coefficient, and the value of roughness coefficient needs to be adjusted according to the expansion degree and the particle characteristics of soil.

[0057] After the hydraulic opening and roughness coefficient are updated, the edge features and edge weights of the dynamic graph structure are further adjusted. The mechanical opening (corresponding to the calculation basis of hydraulic opening) and roughness coefficient in edge features are directly replaced by the updated values; the edge weight is recalculated based on the updated hydraulic opening, and the calculation process still follows the hydraulic conductivity coefficient and edge weight normalization method in S2 step. Through the above adjustment, the adjusted graph structure reflecting the state after crack propagation is obtained. Compared with the initial dynamic graph structure, the edge features and edge weights of the adjusted graph structure are dynamically updated, which is more consistent with the actual evolution state of the slope.

[0058] S5, input the adjusted graph structure into the graph neural network, simulate the two-way feedback of seepage and stress through the message passing mechanism, and output the second updated node features.

[0059] Specifically, based on the adjusted graph structure, the two-way feedback between seepage and stress is simulated again through the graph neural network to obtain node features that are more consistent with the state of the slope after crack propagation. First, the adjusted graph structure obtained in step S4 is input into the same graph neural network as in step S3. The graph structure has updated edge features and edge weights due to crack propagation, and can accurately reflect the topological relationship and physical properties of the current crack network.

[0060] Subsequently, the message passing mechanism in step S3 is used for operation, and the message passing formula and parameter definition remain unchanged. In this message passing process, since the edge features (mechanical opening, roughness coefficient, length) and edge weights of the graph structure have been updated based on the crack propagation results, the messages generated by the message function will contain seepage-stress coupling information after crack propagation, such as the updated hydraulic opening affecting the transmission strength of seepage messages, and the changed roughness coefficient changing the transmission efficiency of stress messages. The aggregation function will fully consider these updated information when integrating neighbor node messages, so that the update of node features can reflect the influence of crack propagation on seepage and stress field.

[0061] After multiple rounds of message passing, the graph neural network outputs the second updated node features, which still contain predicted displacement and predicted water head, but compared with the first updated node features, the values have been adjusted according to the seepage-stress feedback after crack propagation: the predicted displacement will reflect the crack propagation caused by the slope deformation aggravation or distribution change, and the predicted water head will reflect the water energy redistribution caused by the change of seepage path after crack propagation. The second updated node features can more accurately depict the internal state of the slope after crack propagation, and provide reliable basis for subsequent failure mode judgment and safety factor calculation.

[0062] S6, based on the predicted water head in the second updated node features, identify the seepage front and calculate the Gaussian curvature of the seepage front and the clustering coefficient of the crack network; according to the comparison of the Gaussian curvature and the clustering coefficient with the critical value, judge the failure mode of the slope.

[0063] Specifically, this step realizes seepage front identification, key parameter calculation and failure mode judgment by analyzing the second updated node features, and the core is to distinguish different failure modes by using the correlation between seepage field and fracture network topological parameters. First, the seepage front is identified based on the predicted water head in the second updated node features. The seepage front refers to the front interface of water seepage in the fracture network, and its position can be determined by the distribution characteristics of the predicted water head. In the seepage direction, the predicted water head transitions from a high value area to a low value area, and the seepage front corresponds to the position of the predicted water head mutation or a specific threshold. By setting a reasonable water head threshold (determined based on the water holding characteristics and seepage law of the slope soil body), the nodes that meet the threshold condition are screened out, and the face formed by connecting these nodes is the seepage front.

[0064] After identifying the seepage front, the Gaussian curvature of the seepage front is calculated. The Gaussian curvature is a geometric parameter that describes the bending degree of a surface and can reflect the morphological characteristics of the seepage front. The calculation formula is as follows:

[0065]

[0066] In the formula, represents the Gaussian curvature, and represent the principal curvature radii of the seepage front in two mutually perpendicular directions. The calculation of the principal curvature radii is based on the three-dimensional coordinate data of the seepage front, and is obtained by surface fitting and differential geometry operation. The positive and negative values and the size of the Gaussian curvature can reflect whether the seepage front is convex, concave or flat, and the degree of bending.

[0067] At the same time, the clustering coefficient of the fracture network is calculated. The clustering coefficient is a topological parameter that describes the degree of node aggregation in the network and can reflect the connectivity characteristics of the fracture network. The calculation formula is as follows:

[0068]

[0069] In the formula, represents the clustering coefficient, represents the total number of nodes in the fracture network, represents the number of edges actually existing between the neighbor nodes of node , represents the number of neighbor nodes of node . The value range of the clustering coefficient is 0 to 1. The closer the value is to 1, the closer the connection between the neighbor nodes of the node, and the stronger the local connectivity of the fracture network. The closer the value is to 0, the sparser the connection between the neighbor nodes of the node, and the weaker the local connectivity of the fracture network.

[0070] After the calculation of the Gaussian curvature and the clustering coefficient, the two are compared with the preset critical value respectively to determine the failure mode of the slope. If the Gaussian curvature is greater than the critical curvature value and the clustering coefficient is less than the critical clustering value, it indicates that the seepage front surface is severely curved and the local connectivity of the fracture network is weak. At this time, the slope is prone to shallow fluidization failure, because the fracture network with weak local connectivity is difficult to withstand the softening of the soil caused by water infiltration, and the curved seepage front surface is easy to cause uneven stress of the shallow soil and lead to fluidization; if the Gaussian curvature is less than the critical curvature value and the clustering coefficient is greater than the critical clustering value, it indicates that the seepage front surface is gently shaped and the local connectivity of the fracture network is strong. At this time, the slope is prone to deep sliding failure, because the fracture network with strong connectivity will form a through seepage channel, leading to a decrease in the effective stress of the deep soil, and the gently shaped seepage front surface makes the deep soil as a whole, which is easy to cause sliding.

[0071] S7, based on the failure mode and the predicted displacement in the second updated node features, calculating the safety factor of the slope, and triggering a warning when the safety factor is lower than a threshold value.

[0072] Specifically, this step combines the identified slope failure mode with the second updated predicted displacement to calculate the safety factor through a targeted mechanical model, and then starts the warning according to the relationship between the safety factor and the preset threshold value, ensuring accurate judgment and timely response to the risk of slope instability. First, for different failure modes determined in step S6, select the appropriate safety factor calculation model, because the failure mechanisms of shallow fluidization and deep sliding are essentially different, their mechanical control factors and equilibrium conditions are different, and a differentiated model is needed to ensure calculation accuracy.

[0073] For the shallow fluidization failure mode, the essence of the failure is that the shallow soil of the slope softens under the action of seepage, the shear strength of the soil decreases sharply, and finally presents a fluid-like flow state, and the damage range is concentrated in the surface layer of the slope (usually not more than 3 times the fracture development depth). For this mode, a simplified limit equilibrium method considering seepage force and soil softening effect is used to calculate the safety factor, and the specific formula is as follows:

[0074]

[0075] In the formula, safety factor under the shallow fluidization mode, N represents the number of strips obtained by dividing the shallow sliding body along the sliding direction, G represents the self-weight of the th strip (calculated by the volume of the strip and the natural density of the soil), Pw represents the pore water pressure of the bottom surface of the th strip (converted from the second updated predicted water head, where is the density of water, Predicted water head of the first block bottom surface, Width of the first block, Effective internal friction angle of the first block soil body (needs to be modified by the predicted displacement, when the predicted displacement is greater than 0.5 times the allowable deformation value, Linear attenuation coefficient 0.8~0.9 is corrected, and the attenuation coefficient decreases with the increase of displacement), Effective cohesion of the first block soil body (the correction method is consistent with ), Angle between the bottom surface of the first block and the horizontal plane, Additional load on the top surface of the first block (such as the load of engineering facilities on the top of the slope, 0 if there is no additional load), Seepage force on the first block (calculated by seepage velocity and block volume, , Seepage velocity in the first block, derived from the predicted water head gradient combined with Darcy's law). In the calculation process, the second updated predicted displacement is mainly used to modify the soil strength parameters ( and ), because the larger the predicted displacement, the more significant the plastic deformation of the soil has occurred, and the more serious the internal structure damage, so the shear strength parameters decrease, reflecting the correlation between displacement and strength.

[0076] For deep sliding failure mode, the failure feature is that a through sliding surface is formed inside the slope, and the sliding body slides along the surface as a whole, and the failure depth usually exceeds shallow fluidization, and the sliding surface is often connected with the deep fracture network. For this mode, the Morgenstern-Price Method is used to calculate the safety factor, which can consider the geometric shape of the sliding surface and the stress-strain relationship. This method introduces a function relationship between block forces to meet all balance conditions (force balance and moment balance), and has higher calculation accuracy. The specific formula is as follows:

[0077]

[0078] In the formula, Safety factor under deep sliding mode, Total length of the sliding surface, Shear strength of the soil on the sliding surface ( , effective normal stress on the sliding surface, actual shear stress on the sliding surface. In the actual calculation of this formula, the position of the sliding surface is determined in combination with the second updated predicted displacement. First, by analyzing the spatial distribution of the predicted displacement, the displacement discontinuity (such as the area where the displacement gradient is greater than 0.1 mm / m) is determined as the potential sliding surface range. Then, in combination with the geological structure and fracture distribution of the slope, the geometric shape of the sliding surface (such as circular arc type, polyline type) is finally determined. Meanwhile, the predicted displacement is also used to calculate the deformation potential energy of the sliding body. When the predicted displacement exceeds 1 times the allowable deformation value, the damage factor is introduced in the calculation wherein is the predicted displacement, and is the ultimate displacement when the soil fails), and the corrected shear strength is

[0079] After the safety factor calculation is completed, a reasonable safety factor threshold is set. The determination of the threshold takes into account the engineering grade of the slope, the importance of the surrounding environment, and the risk bearing capacity. For a first-class slope (such as a slope next to an important transportation hub, a slope in a densely populated area), the safety factor threshold can be set to 1.3-1.5; for a second-class slope (such as a slope next to a general highway, a slope in a non-densely populated residential area), the threshold can be set to 1.1-1.3; for a third-class slope (such as a slope in a remote area, a slope without important facilities), the threshold can be set to 1.0-1.1.

[0080] Finally, the calculated safety factor is compared with the corresponding threshold. If the safety factor is greater than or equal to the threshold, it indicates that the slope is currently in a stable state and no warning needs to be triggered, but the changes in the predicted displacement and safety factor need to be continuously monitored. If the safety factor is less than the threshold, it indicates that the slope has entered an unstable state and immediate warning needs to be triggered. The warning level is divided into three levels. For example, when the safety factor is 0.8-1.0 times the threshold, a yellow warning is triggered, the monitoring frequency needs to be increased (such as shortening the monitoring interval from 1 hour to 15 minutes), and personnel are organized to investigate abnormal phenomena such as surface cracks and seepage on the slope. When the safety factor is 0.6-0.8 times the threshold, an orange warning is triggered, the dangerous areas around the slope need to be closed to personnel and vehicles, and the preliminary preparations for the slope reinforcement plan (such as preparing reinforcement materials and organizing construction teams) need to be started. When the safety factor is less than 0.6 times the threshold, a red warning is triggered, personnel around the slope need to be immediately evacuated, and emergency reinforcement measures (such as quickly spraying concrete to close the slope surface and setting temporary anti-slide piles) need to be started. At the same time, the local emergency management department needs to be reported to ensure that effective prevention and control measures are taken before the slope fails, thereby reducing disaster losses.

[0081] ​​​The above-mentioned machine learning-based granite residual soil slope evaluation method collects and preprocesses the crack geometry, hydrology and mechanics data of the slope to generate a standardized data set, then constructs a dynamic graph structure represented by nodes and edges to depict the crack network, uses the message passing mechanism of the graph neural network to simulate the two-way feedback of seepage and stress, and first outputs node features containing predicted displacement and predicted water head; based on this, the crack propagation behavior is predicted and the hydraulic opening and roughness coefficient are updated to adjust the dynamic graph structure, and then the adjusted graph structure is input into the graph neural network for secondary feedback simulation to output more accurate node features; then the seepage front is identified based on the second updated predicted water head, and the Gaussian curvature and crack network clustering coefficient are calculated, and the critical value is compared to adaptively judge the slope failure mode; finally, the safety factor is calculated in combination with the failure mode and the predicted displacement to trigger the warning, so as to realize high-precision tracking of crack dynamic evolution, overcome the numerical divergence problem of traditional numerical simulation, and significantly improve the accuracy and reliability of slope stability evaluation and warning.

[0082] Reference Figure 2 In an optional embodiment, based on the standardized data set, the crack network is represented as a dynamic graph structure, including the following steps:

[0083] S11, according to the crack geometry data in the standardized data set, obtaining crack position and intersection information; based on the crack position and intersection information, determining the node set and edge set of the dynamic graph; wherein each node in the node set corresponds to a crack intersection or endpoint, and each edge in the edge set corresponds to a crack segment.

[0084] Specifically, the crack geometry data is called from the standardized data set, which already contains information such as the three-dimensional coordinates of the crack in the slope space, the extension trajectory, etc. Through coordinate analysis and trajectory analysis, the specific position of each crack (represented by a sequence of coordinate points in a spatial rectangular coordinate system) and the intersection information between different cracks are obtained. By comparing the coordinate point sequences of multiple cracks, the points that are coordinate coincident or within a preset error range (the error does not exceed the data acquisition accuracy, ensuring the accuracy of the topological relationship) are selected as the crack intersection points.

[0085] Based on the obtained crack position and intersection information, the node set and edge set of the dynamic graph are further determined: the construction of the node set takes the crack intersection and endpoint as the core. For each crack, the starting point and the ending point of the trajectory are defined as the endpoints, and the points on the trajectory that intersect with other cracks are defined as the intersection points. All the endpoints and intersection points are collected and de-duplicated (to avoid repeated coordinate points as different nodes) to form the node set. Each node corresponds to a unique spatial coordinate and crack association information. The construction of the edge set is based on the crack entity between adjacent nodes. For any two directly connected nodes (i.e., there is no other node between the two nodes, and the corresponding crack segment is continuous and uninterrupted), the continuous crack entity of the segment is taken as the edge. Each edge uniquely corresponds to a specific crack segment, thereby establishing a topological correspondence between nodes and edges and forming the basic topology of the dynamic graph.

[0086] S12, based on the mechanical data in the standardized data set, obtaining stress data and strain data; based on the stress data and strain data, calculating the normal stress and strain energy density of each node in the dynamic graph; based on the hydrological data in the standardized data set, obtaining the pore water pressure, and calculating the water head of each node in the dynamic graph according to the pore water pressure; combining the normal stress, strain energy density and water head into a node feature vector.

[0087] Specifically, the stress data (including total stress tensor information at different positions of the slope) and strain data (including deformation information such as normal strain and shear strain of the soil body) are retrieved from the mechanical data of the standardized data set, and the normal stress and strain energy density of each node are calculated based on these data. The calculation of the normal stress takes the orientation of the crack surface where the node is located as the basis, and the projection operation of the stress tensor in the normal direction of the crack surface is used to obtain the vertical force borne by the crack surface at the node. The strain energy density is calculated by integrating the stress-strain relationship based on the elastic mechanics principle, combined with the stress data and strain data around the node, which reflects the deformation energy state stored in the soil body around the node.

[0088] At the same time, the pore water pressure data of the node corresponding position is retrieved from the hydrological data of the standardized data set, and the calculation of the water head takes the lowest point of the slope toe as the reference surface. The mechanical conversion relationship between the pore water pressure and the water head (the water head is the energy of unit weight water body, which is equal to the pore water pressure divided by the specific weight of water plus the node elevation) is used to convert the pore water pressure into the water head value of the node, which reflects the pressure and position energy state of the water body at the node.

[0089] Finally, the calculated normal stress, strain energy density and water head are combined into a vector form in a fixed order (such as normal stress → strain energy density → water head), and each node corresponds to a three-dimensional node feature vector, which fully represents the core state of the node in terms of mechanics and hydrology, providing feature input for the subsequent information transmission of the graph neural network.

[0090] S13. Based on the crack geometry data in the standardized dataset, obtain the mechanical opening, roughness coefficient, and length; construct the edge feature vector according to the mechanical opening, roughness coefficient, and length.

[0091] Specifically, from the fracture geometry data in the standardized dataset, three core parameters are retrieved for each fracture segment corresponding to each edge. The mechanical opening is the average spacing between the walls, calculated using the spatial coordinates of the fracture segment walls, which directly reflects the physical opening degree of the fracture segment. The roughness coefficient is based on the contour curve characteristics of the fracture segment walls, obtained by comparing with the standard roughness curve or calculating the contour undulation, characterizing the influence of wall roughness on seepage and stress transmission. The length is obtained by calculating the Euclidean distance between the two points using the spatial coordinates of the nodes at both ends of the edge, reflecting the actual spatial scale of the fracture segment.

[0092] The three parameters mentioned above—mechanical opening, roughness coefficient, and length—are combined into a vector form in a preset order (e.g., mechanical opening → roughness coefficient → length). Each edge corresponds to a three-dimensional edge feature vector. This vector can accurately characterize the physical morphology and surface properties of the fracture segment, providing basic parameter support for subsequent hydraulic opening calculation and edge weight determination.

[0093] S14. Based on the fracture geometry data in the standardized dataset, obtain the mechanical opening and roughness coefficient; according to the mechanical opening and roughness coefficient, calculate the hydraulic opening of each edge in the dynamic graph; the formula for calculating the hydraulic opening is:

[0094]

[0095] in, Hydraulic aperture represents the effective aperture of a fracture under hydraulic action; Mechanical aperture, representing the physical aperture of the crack; is the roughness coefficient, representing the roughness of the crack wall surface.

[0096] Specifically, this step calculates the hydraulic aperture using mechanical aperture and roughness coefficient, establishing the correlation between fracture physical morphology and effective seepage aperture. The mechanical aperture corresponding to each edge is retrieved from the fracture geometry data in the standardized dataset. With roughness coefficient (consistent with the parameters obtained in step S13, to ensure data consistency), the two parameters are substituted into the above formula for calculation. Since the hydraulic opening degree is proportional to the third power of the mechanical opening degree, a slight change in the mechanical opening degree will cause a significant change in the hydraulic opening degree, reflecting the key influence of the physical opening degree of the fracture on the seepage capacity; at the same time, the hydraulic opening degree is inversely proportional to the square of the roughness coefficient, and the rough wall surface will reduce the effective seepage opening degree of the fracture by increasing the frictional resistance of the water flow, which conforms to the action law of the wall surface characteristics in the actual seepage process. The hydraulic opening degree calculated by the formula can accurately reflect the actual effective seepage space of the fracture section under hydraulic action, providing core parameters for subsequent edge weight calculation.

[0097] S15, based on the fracture geometry data in the standardized data set, the length of the fracture section is obtained; based on the hydraulic opening degree and the length of the fracture section, the edge weight of each edge in the dynamic graph is calculated; the calculation formula of the edge weight is:

[0098]

[0099] wherein, is the edge weight, indicating the hydraulic conductivity of the fracture section; is the density of water, is the acceleration of gravity, is the dynamic viscosity of water, is the length of the fracture section.

[0100] Specifically, the edge weight is calculated by the hydraulic opening degree and the length of the fracture section, and the seepage contribution ability of each fracture section is quantified. From the fracture geometry data in the standardized data set, the length of the fracture section corresponding to each edge is called , combined with the hydraulic opening degree calculated in step S14, the above formula is calculated. In the formula, is a constant term determined by the physical properties of water, reflecting the influence of the characteristics of water flow on the conductivity; is a variable term, wherein the third power of the hydraulic opening degree reflects the significant influence of the cross-sectional area of the seepage channel on the water flow transmission, and the length of the fracture section reflects the path resistance of the water flow in the fracture. The longer the length, the greater the water flow resistance, and the lower the hydraulic conductivity. The edge weight calculated by the formula can quantitatively represent the seepage importance of each fracture section in the fracture network, and the greater the weight, the stronger the fracture section's ability to transmit water flow, providing a basis for the intensity distribution of message transmission in the subsequent graph neural network.

[0101] S16, combining the node feature vector, the edge feature vector and the edge weight to construct a dynamic graph structure.

[0102] Specifically, this step is the final integration link of dynamic graph structure construction, and the core is to fuse the feature information, weight information and topological structure of the nodes and edges obtained in the previous steps to form a complete dynamic graph. First, based on the node set and edge set determined in S11, the node feature vector constructed in S12 is associated with each node one by one, so that each node not only has topological coordinates, but also has mechanical and hydrological feature information. At the same time, the edge feature vector constructed in S13 is associated with each edge one by one, so that each edge has physical form and surface property information. Finally, the edge weight calculated in S15 is bound to the corresponding edge to determine the weight proportion of each edge in seepage transmission.

[0103] Through the above association and integration, the dynamic graph structure formed contains three core components: one is the topological structure (node set and edge set), which reflects the spatial connection relationship of the fracture network; the second is the feature information (node feature vector and edge feature vector), which reflects the physical state attributes of nodes and edges; the third is the weight information (edge weight), which reflects the seepage transmission capacity of the edge. This dynamic graph structure can completely depict the topological form and multi-field coupling characteristics of the fracture network, and can be directly used as input data for subsequent graph neural networks to support bidirectional feedback simulation of seepage and stress.

[0104] In an optional embodiment, based on the first updated node features, the expansion behavior of the fracture is predicted to obtain a prediction result; and the hydraulic opening and the roughness coefficient of the fracture are updated according to the prediction result to adjust the edge features and the edge weights of the dynamic graph structure to obtain an adjusted graph structure, including the following steps:

[0105] S21, for each fracture segment, based on the normal stress in the first updated node features, calculate the energy release rate; the calculation formula of the energy release rate is:

[0106]

[0107] wherein, is the energy release rate, indicating the energy released when the fracture expands; is the normal stress, indicating the stress perpendicular to the fracture surface; is the elastic modulus, indicating the elastic properties of the soil; is the fracture surface area, indicating the area of the fracture.

[0108] Specifically, this step quantifies the energy basis required for fracture expansion, i.e., the energy release rate, through the normal stress in the first updated node features. First, the energy release rate of each fracture segment corresponding to each edge in the dynamic graph structure is calculated separately to ensure that the expansion potential of each fracture segment can be accurately evaluated. When calculating, the normal stresses of the two end nodes of the corresponding fracture segment are first extracted from the first updated node features, and their average value is taken as the representative normal stress of the fracture segment This is because the stress state of the fractured segment is affected by the stresses at both ends, and the average value can more accurately reflect the overall stress level. Elastic modulus Based on previously collected slope mechanics data, the elastic parameters of the granite residual soil under corresponding stress levels were determined to avoid energy calculation distortions caused by parameter selection biases. (Fracture surface area) The energy release rate is calculated using the geometric parameters of the fracture segment, specifically the product of the fracture segment length and the fracture height (determined from the fracture development depth based on slope borehole data). This ensures the actual stress-bearing area covering the fracture. It directly reflects the energy that a crack can release if it expands under the current stress state, and is the core indicator for judging whether a crack has the conditions to expand.

[0109] S22. Compare the energy release rate with the material fracture energy, when... Time triggers crack propagation, in which Fractal energy is the energy required for a material to resist fracture.

[0110] Specifically, this step is the determination of crack propagation, using energy balance to determine whether crack propagation has begun. Material fracture energy. This is an inherent property of residual granite soil, determined through laboratory tests. Specifically, a three-point bending test was used to conduct fracture tests on residual granite soil samples with different moisture contents and dry densities. The energy absorbed during the fracture process was recorded, and the average value of multiple tests was taken as the final value. The value is used to ensure that it accurately reflects the slope soil's ability to resist crack propagation. The energy release rate of each crack segment calculated in step S21 is then used. , with that Compare them one by one: If This indicates that the energy released by the fracture is insufficient to overcome the fracture resistance of the soil, and the fracture remains stable, requiring no further calculations related to its propagation; if If the energy released by the crack exceeds the soil's resistance, it satisfies the energy condition for expansion, triggering the expansion process of the crack segment and leading to the subsequent calculation of the expansion direction and length.

[0111] S23. For cracks that trigger propagation, determine the propagation direction based on the direction of the maximum principal stress, determine the crack dip angle based on the propagation direction, and calculate the propagation length; the formula for calculating the propagation length is:

[0112]

[0113] in, "Extension length" indicates the length of the crack extension. The expansion resistance coefficient represents the resistance to crack propagation.

[0114] Specifically, this step aims to determine the extension direction and length for the triggered expanding crack, ensuring that the crack extension process conforms to the actual mechanical law. The determination of the extension direction is based on the maximum principal stress direction of the slope. Through the stress tensor data in the first updated node characteristics, the maximum principal stress direction is calculated using the stress principal axis decomposition method. The crack extension direction is always perpendicular to the maximum principal stress direction, because the soil is prone to shear or tensile failure perpendicular to the maximum principal stress direction under the action of the maximum principal stress, which conforms to the basic law of crack extension in rock mechanics. The crack inclination angle is directly determined by the angle between the extension direction and the horizontal plane. For example, if the maximum principal stress direction is vertically upward, the crack extension direction is horizontal, and the corresponding crack inclination angle is 0°. In the calculation of the extension length, is the "excess energy" of crack extension, which is the energy remaining after overcoming the soil fracture resistance. This energy will be used to drive the increase of crack length; the extension resistance coefficient reflects the soil friction, particle engagement and other resistance during crack extension. It is calibrated through indoor direct shear test, and is obtained by fitting the relationship between resistance and extension length under different crack extension speeds, ensuring accurate quantification of the resistance size in the actual extension process. The calculated by this formula is the length that the crack can extend under the current energy condition, which directly determines the geometric shape of the crack after extension.

[0115] S24, based on the extension length and crack inclination angle, update the mechanical opening of the crack to obtain the updated mechanical opening; the update formula of the mechanical opening is:

[0116]

[0117] wherein, is the updated mechanical opening; , is the crack opening change, representing the change of the mechanical opening; is an empirical coefficient, representing the influence factor of the extension length on the opening change; is the crack inclination angle, representing the angle between the crack and the horizontal plane.

[0118] Specifically, this step focuses on the update of the mechanical opening of the crack after extension, and the core is to reflect the influence of the extension length and inclination angle on the physical opening of the crack. The mechanical opening is the actual distance between the crack walls. After the crack extends, not only the length increases, but also the opening changes due to the extrusion or stretching of the soil during the extension process. The calculation of the opening change ; the empirical coefficient needs to be calibrated through field monitoring data. Select the monitoring points of the slope where the crack has expanded, and record the extension length , inclination angle The relationship between the actual opening degree and the actual opening degree was obtained by fitting the data using a linear regression method. The value is used to ensure that it reflects the correlation between crack propagation and aperture variation in local granite residual soil; This is used to correct the effect of the dip angle on the opening change, such as in vertical cracks ( As fractures expand, gravity makes the fracture walls easier to separate, resulting in a greater change in aperture, while horizontal fractures ( The opening is less affected by gravity, and the change in opening is relatively small. The initial mechanical opening... (Crack geometry data from a standardized dataset) and Add them together to get the updated mechanical opening. This parameter directly reflects the change in physical morphology after the fracture expands and is a key basis for subsequent updates to the hydraulic opening.

[0119] S25. Based on the particle spalling state during crack propagation, update the roughness coefficient to obtain the updated roughness coefficient; the update formula for the roughness coefficient is:

[0120]

[0121] in, This refers to the updated roughness coefficient; The degradation coefficient represents the influence factor of particle spalling on roughness.

[0122] Specifically, this step addresses the particle spalling phenomenon during crack propagation by dynamically updating the roughness coefficient to ensure that edge features reflect changes in crack wall morphology. During crack propagation, the soil wall is subjected to shear or tension, resulting in varying degrees of particle spalling. The spalled particles smooth out the originally rough wall surface, leading to a decrease in the roughness coefficient. Degradation coefficient Indoor tests were conducted to simulate crack propagation in granite residual soil samples, recording the changes in crack opening at different degrees. The variation of the roughness coefficient of the lower wall surface was obtained by fitting using the least squares method. This value ensures that it accurately quantifies the degree to which particle spalling reduces roughness. Initial roughness coefficient. Crack geometry data from a standardized dataset, and compared with Subtracting them yields the updated roughness coefficient. , The larger the value, the more significant the change in fissure aperture and the more severe the particle detachment. The greater the reduction, the more consistent it is with the evolution of wall morphology during actual crack propagation.

[0123] S26, based on the updated mechanical opening degree and the updated roughness coefficient, re-calculate the hydraulic opening degree to obtain an updated hydraulic opening degree; update the edge weight in the dynamic graph structure according to the updated hydraulic opening degree to obtain an adjusted graph structure.

[0124] Specifically, this step is the final link of the dynamic graph structure adjustment, and the core is to convert the changes of the mechanical opening degree and the roughness coefficient caused by the fracture expansion into the update of the edge features and the edge weight in the graph structure. First, the calculation of the updated hydraulic opening degree follows the previously determined correlation between the hydraulic opening degree and the mechanical opening degree and the roughness coefficient, and is re-calculated through and to ensure that the hydraulic opening degree can reflect the actual seepage capacity of the expanded fracture. Subsequently, based on the updated hydraulic opening degree, combined with the length of the fracture segment (the sum of the initial length and ), the weight of each edge is re-calculated according to the calculation logic of the edge weight. The greater the hydraulic opening degree, the stronger the seepage capacity of the fracture segment, and the higher the corresponding edge weight, and vice versa. The updated mechanical opening degree and the updated roughness coefficient are substituted into the edge feature vector to replace the original parameters; the re-calculated edge weight is substituted into the weight of the corresponding edge in the dynamic graph structure, and finally the adjusted graph structure is formed. This graph structure completely retains the geometric shape, physical properties and seepage characteristics of the expanded fracture, and provides an accurate input basis for subsequent input into the graph neural network for seepage-stress feedback simulation.

[0125] In an alternative embodiment, the step of calculating the Gaussian curvature of the seepage front face comprises:

[0126] S31, fitting the seepage front face using the moving least squares method to obtain a fitted surface, and calculating the first fundamental form and the second fundamental form coefficients of the fitted surface.

[0127] Specifically, all predicted water head values are extracted from the second updated node features, a water head threshold value is set according to the water holding characteristics and seepage law of the slope soil body (such as the water head value corresponding to the pore water pressure equal to the water head value corresponding to the saturated water content of the soil body), and nodes with predicted water head values within the threshold value range are selected. These nodes are discrete sampling points that constitute the seepage front face, denoted as point set , wherein are the spatial coordinates of the nodes, respectively.

[0128] The moving least squares method is used for surface fitting. This method does not construct a global interpolation function, but constructs an approximate function only through sampling points in a local area, and can effectively handle the problem of irregular distribution of discrete points of the seepage front face. The specific fitting process is as follows: first, define the parametric form of the surface to be fitted, select two-dimensional parameters and spatial coordinates to establish a mapping relationship (such as , ), so that each discrete point corresponds to a point in the parameter space ; then select a weight function, using a Gaussian weight function , where is the Euclidean distance between the point to be fitted and the sampling point , and is the influence domain radius (determined according to the sampling point density, usually 1.5-2 times the average distance between adjacent sampling points), the weight function ensures that only the sampling points near the point to be fitted participate in the fitting, balancing local accuracy and computational efficiency; then select a first-order polynomial basis function , which can balance the smoothness of the fitted surface and the degree of fit to the actual shape of the seepage front surface, avoiding overfitting problems caused by high-order basis functions.

[0129] Based on the above settings, construct the local approximation function , where is the basis function coefficient vector. By minimizing the weighted residual sum of squares , solve , and , take the partial derivative with respect to and set it to zero to obtain the coefficient equation , where is the coefficient matrix, is the weight matrix, is the coordinate vector of the discrete points. Solve the linear equation system to obtain , and substitute the approximation function to obtain the fitted surface of the seepage front surface .

[0130] Subsequently, calculate the first fundamental form coefficients and the second fundamental form coefficients of the fitted surface. The first fundamental form describes the metric properties of the surface, reflecting the geometric quantities such as the arc length and angle of the curve on the surface, and its coefficients are calculated through the first-order partial derivatives of the fitted surface: , , , where , are the first-order partial derivatives of the fitted surface with respect to the parameters , , which can be obtained by differentiating the approximation function ; the second fundamental form describes the curvature properties of the surface, reflecting the concave-convex shape of the surface in space, and its coefficients need to be calculated in combination with the first-order partial derivatives and the second-order partial derivatives: first, calculate the second-order partial derivatives , , , then construct the unit normal vector of the surface , the final coefficient , , The above coefficients completely characterize the geometric features of the fitting surface, and provide key parameters for Gaussian curvature calculation.

[0131] S32, based on the first fundamental form and the second fundamental form coefficients, calculate the Gaussian curvature of the fitting surface corresponding to the seepage front surface; the calculation formula of Gaussian curvature is:

[0132]

[0133] Wherein, is the Gaussian curvature, which represents the bending degree of the fitting surface; is the first order fundamental form coefficient, which represents the first fundamental form of the fitting surface; is the second order fundamental form coefficient, which represents the second fundamental form of the fitting surface.

[0134] Specifically, Gaussian curvature is an intrinsic geometric property of the surface, and the positive and negative values and the size directly reflect the local morphology of the surface. When , the surface at this point is convex (such as the local part of the sphere), indicating that the seepage front surface is convex upward in this area, and the water infiltration direction is concentrated; when , the surface at this point is concave (such as the local part of the saddle), indicating that the seepage front surface is concave downward in this area, and the water infiltration direction is dispersed; when , the surface at this point is flat or cylindrical, indicating that the seepage front surface has no obvious bending in this area, and the water infiltration direction is uniform.

[0135] In the formula, the denominator is the discriminant of the first fundamental form, which is obtained by the first order fundamental form coefficient calculated in step S31, and its value is always positive (due to the regularity of surface parameterization, ensuring that the tangent vectors of any point on the surface are linearly independent), which is the basis for the meaningfulness of Gaussian curvature calculation; the numerator is the discriminant of the second fundamental form, which is obtained by the second order fundamental form coefficient calculated in step S31, and its positive and negative directly determine the positive and negative of Gaussian curvature, and the numerical value reflects the degree of bending of the surface.

[0136] The specific calculation process traverses the key sampling points on the fitting surface (such as the fitting points corresponding to the discrete seepage front surface nodes), and calculates Substitute the formula, get the point of Gaussian curvature value, finally form the seepage front face of Gaussian curvature distribution atlas. The atlas can intuitively reflect the overall bending characteristics of the seepage front face, and provide quantitative basis for subsequent combination of clustering coefficient to judge the failure mode of slope (such as convex region easy to cause shallow fluidization, concave region easy to cause deep sliding).

[0137] In an alternative embodiment, the formula for calculating the clustering coefficient of the fracture network is:

[0138]

[0139] wherein, is the clustering coefficient, indicating the connectivity of the fracture network; is the node set of the adjusted graph structure, indicating the set of all nodes in the dynamic graph; is the total number of nodes in the node set of the adjusted graph structure; is the neighbor set of node , indicating the set of nodes connected to node ; is the total number of nodes in the neighbor set of node ; is the actual edge number between neighbors, indicating the number of edges existing between the neighbors of node .

[0140] Specifically, the core of this step is to quantify the clustering coefficient of the fracture network through the above formula, so as to reflect the aggregation degree of the nodes in the fracture network, and further judge the network connectivity, providing a key topological parameter for subsequent identification of slope failure mode. When calculating, first, the node set in the adjusted graph structure is determined , and the total number of nodes contained is counted , which is the basis for subsequent normalization of summation result, ensuring that the clustering coefficient can reflect the average aggregation level of the entire network, rather than the local node characteristics.

[0141] For each node in the node set (i.e. node j), first determine its neighbor set , which is all nodes directly connected to node j in the adjusted graph structure, and then count the total number of nodes in the set , which reflects the connection density of node j in the network. Subsequently, the actual number of edges between all nodes in the neighbor set is counted , which reflects the connectivity tightness between the neighbor nodes of node j. If the actual number of edges between neighbor nodes is large, it means that these nodes form a dense local connection; if the actual number of edges is small, the local connection is relatively sparse.

[0142] In the formula, the ratio of the numerator to the denominator represents the proportion of the actual number of edges between the neighbor nodes of node j to the theoretically maximum possible number of edges. Theoretically, the maximum number of edges that can be formed between nodes is , so the ratio is essentially the proportion of the aggregation degree of the local network in which node j is located. After summing up the ratio of all nodes j and dividing by the total number of nodes , the clustering coefficient C obtained is in the range of 0 to 1: when C tends to 1, it indicates that there are close connections between the neighbors of the nodes in the entire fracture network, the local connectivity of the network is strong, and the seepage channel is easy to form; when C tends to 0, it indicates that the connection between the neighbors of the nodes is sparse, the local connectivity of the network is weak, and the seepage path is mostly isolated fragments. This feature is directly compared with the critical value to assist in determining whether the slope tends to deep sliding or shallow flow failure.

[0143] The above-mentioned method for evaluating a granite residual soil slope based on machine learning collects and preprocesses the fracture geometry, hydrological and mechanical data of the slope to generate a standardized data set, and then constructs a dynamic graph structure represented by nodes and edges to depict the fracture network. The message passing mechanism of the graph neural network is used to simulate the two-way feedback of seepage and stress, and the first output includes predicted displacement and predicted water head. Based on this, the fracture expansion behavior is predicted and the hydraulic opening and roughness coefficient are updated to adjust the dynamic graph structure, and then the adjusted graph structure is input into the graph neural network for secondary feedback simulation to output more accurate node features. Then, based on the second updated predicted water head, the seepage front is identified and the Gaussian curvature and fracture network clustering coefficient are calculated, and the failure mode of the slope is adaptively determined by comparing the critical value. Finally, the safety factor is calculated in combination with the failure mode and the predicted displacement to trigger the warning, thereby realizing high-precision tracking of the dynamic evolution of the fracture, overcoming the numerical divergence problem of traditional numerical simulation, and significantly improving the accuracy and reliability of slope stability evaluation and warning.

[0144] It should be understood that, although each step in the flowchart involved in each embodiment as described above is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0145] Based on the same inventive concept, the embodiments of the present application also provide a system for implementing the above-mentioned machine learning-based evaluation method for granite residual soil slopes. The implementation scheme for solving problems provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in one or more machine learning-based evaluation system embodiments for granite residual soil slopes provided below can refer to the limitations of the machine learning-based evaluation method for granite residual soil slopes described above, which will not be repeated here.

[0146] In one exemplary embodiment, as shown in Figure 3 A machine learning-based evaluation system 30 for granite residual soil slopes is provided for implementing the method in each of the above method embodiments, which includes:

[0147] A slope data standardization module 31 is configured to collect fissure geometric data, hydrological data and mechanical data of the slope to obtain slope state data, and to normalize and interpolate the slope state data for preprocessing to generate a standardized data set.

[0148] A fissure network topology modeling module 32 is configured to represent the fissure network as a dynamic graph structure based on the standardized data set, wherein in the dynamic graph structure, nodes correspond to fissure intersection points or end points, edges correspond to fissure segments, node features include normal stress, strain energy density and water head, edge features include mechanical opening, roughness coefficient and length, and edge weights are calculated based on hydraulic opening.

[0149] A seepage stress coupling simulation module 33 is configured to input the dynamic graph structure into a graph neural network to simulate the bidirectional feedback of seepage and stress through a message passing mechanism, and output first updated node features, wherein the node features include predicted displacement and predicted water head.

[0150] A graph structure adaptive adjustment module 34 is configured to predict the expansion behavior of the fissures based on the first updated node features to obtain a prediction result, and to update the hydraulic opening and roughness coefficient of the fissures according to the prediction result to adjust the edge features and edge weights of the dynamic graph structure to obtain an adjusted graph structure.

[0151] A multi-round feature iterative optimization module 35 is configured to input the adjusted graph structure into the graph neural network to simulate the bidirectional feedback of seepage and stress through the message passing mechanism, and output second updated node features.

[0152] A failure mode intelligent identification module 36 is configured to identify the seepage front based on the predicted water head in the second updated node features, and to calculate the Gaussian curvature of the seepage front and the clustering coefficient of the fissure network; and to determine the failure mode of the slope according to the comparison of the Gaussian curvature and the clustering coefficient with a critical value.

[0153] The slope safety early warning decision module 37 is configured to calculate a safety factor of the slope based on the failure mode and the predicted displacement in the second updated node features, and trigger an early warning when the safety factor is lower than a threshold.

[0154] Embodiments of the present application further provide a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method embodiments as described above when executing the computer program.

[0155] Embodiments of the present application further provide a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method embodiments as described above.

[0156] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts are described in the part of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present disclosure according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0157] The above embodiments only express several implementation manners of the embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the patent scope of the embodiments of the present application. It should be pointed out that, for those skilled in the art, without departing from the concept of the embodiments of the present application, several modifications and improvements can be made, which are all within the protection scope of the embodiments of the present application.

Claims

1. A method for evaluating a granite residual soil slope based on machine learning, characterized by, The method comprises: S1, collecting crack geometry data, hydrological data and mechanical data of the slope to obtain slope state data; and performing normalization and interpolation preprocessing on the slope state data to generate a standardized data set; S2, based on the standardized data set, representing the crack network as a dynamic graph structure; wherein in the dynamic graph structure, nodes correspond to crack intersection points or end points, edges correspond to crack segments, node features include normal stress, strain energy density and water head, edge features include mechanical opening degree, roughness coefficient and length, and edge weight is calculated based on hydraulic opening degree; S3, inputting the dynamic graph structure into a graph neural network to simulate the bidirectional feedback of seepage and stress through a message passing mechanism, and outputting the first updated node features; wherein the node features include predicted displacement and predicted water head; S4, based on the first updated node features, predicting the crack propagation behavior to obtain a prediction result; updating the hydraulic opening degree and roughness coefficient of the crack according to the prediction result to adjust the edge features and edge weights of the dynamic graph structure, and obtaining an adjusted graph structure; S5, inputting the adjusted graph structure into the graph neural network to simulate the bidirectional feedback of seepage and stress through the message passing mechanism, and outputting the second updated node features; S6, based on the predicted water head in the second updated node features, identifying the seepage front and calculating the Gaussian curvature of the seepage front and the clustering coefficient of the crack network; comparing the Gaussian curvature and the clustering coefficient with a critical value to determine the failure mode of the slope; S7, based on the failure mode and the predicted displacement in the second updated node features, calculating the safety factor of the slope, and triggering an early warning when the safety factor is lower than a threshold.

2. The method of claim 1, wherein, The method comprises: S11, obtaining crack position and intersection information according to crack geometry data in the standardized data set; determining a node set and an edge set of a dynamic graph based on the crack position and the intersection information; wherein each node in the node set corresponds to a crack intersection point or end point, and each edge in the edge set corresponds to a crack segment; S12, obtaining stress data and strain data based on mechanical data in the standardized data set; calculating the normal stress and strain energy density of each node in the dynamic graph based on the stress data and the strain data; obtaining pore water pressure based on hydrological data in the standardized data set, and calculating the water head of each node in the dynamic graph according to the pore water pressure; combining the normal stress, the strain energy density and the water head into a node feature vector; S13, obtaining mechanical opening degree, roughness coefficient and length based on crack geometry data in the standardized data set; constructing an edge feature vector according to the mechanical opening degree, the roughness coefficient and the length; S14, obtain mechanical opening and roughness coefficient based on the fracture geometry data in the standardized data set; calculate the hydraulic opening of each edge in the dynamic graph according to the mechanical opening and the roughness coefficient; the calculation formula of the hydraulic opening is: wherein, is the hydraulic opening, representing the effective opening of the fracture under hydraulic action; is the mechanical opening, representing the physical opening of the fracture; is the roughness coefficient, representing the roughness of the fracture wall. S15, obtain fracture segment length based on the fracture geometry data in the standardized data set; calculate the edge weight of each edge in the dynamic graph based on the hydraulic opening and the fracture segment length; the calculation formula of the edge weight is: wherein, is the edge weight, representing the hydraulic conductivity of the fracture segment; is the density of water, is the acceleration due to gravity, is the dynamic viscosity of water, is the length of the fracture segment; S16, combine the node feature vector, the edge feature vector and the edge weight to construct the dynamic graph structure.

3. The method of claim 1, wherein, The prediction result is obtained by predicting the propagation behavior of the fracture based on the first updated node feature; the hydraulic opening and the roughness coefficient of the fracture are updated according to the prediction result, so as to adjust the edge feature and the edge weight of the dynamic graph structure, and obtain the adjusted graph structure, including: S21, for each fracture segment, calculate the energy release rate based on the normal stress in the first updated node feature; the calculation formula of the energy release rate is: wherein, is the energy release rate, which represents the energy released when the crack propagates; is the normal stress, which represents the stress perpendicular to the crack surface; is the elastic modulus, which represents the elastic property of the soil; is the crack surface area, which represents the area of the crack; S22, compare the energy release rate to a material fracture energy, when a crack extension is triggered, wherein is the material fracture energy, representing the energy required for the material to resist fracture; S23, for the fracture triggering expansion, determine the propagation direction according to the maximum principal stress direction, determine the fracture dip angle according to the propagation direction, and calculate the propagation length; the calculation formula of the propagation length is: wherein, is the extension length, representing the length of the crack extension; is the extension resistance coefficient, representing the resistance of the crack extension; S24, update the mechanical opening of the fracture based on the propagation length and the fracture dip angle, and obtain the updated mechanical opening; the update formula of the mechanical opening is: wherein, is the updated mechanical opening; , is the fracture opening change amount, indicating the change in the mechanical opening; is an empirical coefficient, indicating the influence factor of the extension length on the opening change; is the fracture dip angle, indicating the angle between the fracture and the horizontal plane; S25, update the roughness coefficient based on the particle spalling state in the fracture propagation process, and obtain the updated roughness coefficient; the update formula of the roughness coefficient is: wherein, is the updated roughness coefficient; is a degradation coefficient, representing the influence factor of particle spalling on roughness; S26, recalculate the hydraulic opening based on the updated mechanical opening and the updated roughness coefficient, and obtain the updated hydraulic opening; update the edge weight in the dynamic graph structure according to the updated hydraulic opening, and obtain the adjusted graph structure.

4. The method of claim 1, wherein, The calculation step of the Gaussian curvature of the seepage front face includes: S31, fit the seepage front face using moving least square method to obtain a fitting surface, and calculate the first fundamental form and the second fundamental form coefficient of the fitting surface; S32, calculate the Gaussian curvature of the fitting surface corresponding to the seepage front face based on the first fundamental form and the second fundamental form coefficient; the calculation formula of the Gaussian curvature is: wherein, is the Gaussian curvature of the fitting surface, representing the degree of bending of the fitting surface; is the first fundamental form coefficient of the fitting surface, representing the first fundamental form of the fitting surface; is the second fundamental form coefficient of the fitting surface, representing the second fundamental form of the fitting surface.

5. The method according to any one of claims 1 to 4, characterized in that, The calculation formula of the clustering coefficient of the fracture network is: wherein, is the clustering coefficient, representing the connectivity of the fracture network; is the set of nodes of the adjusted graph structure, representing the set of all nodes in the dynamic graph; represents the total number of nodes in the set of nodes of the adjusted graph structure; is the neighbor set of node , representing the set of nodes connected to node ; represents the total number of nodes in the neighbor set of node ; is the actual number of edges between neighbors, representing the number of edges existing between the neighbors of node .

6. A machine learning based evaluation system for residual soil of granite slopes for implementing the method of any one of claims 1 to 5, characterized in that, The system comprises: A slope data standardization module is configured to collect fracture geometry data, hydrological data and mechanical data of a slope to obtain slope state data, and perform normalization and interpolation preprocessing on the slope state data to generate a standardized data set; A fracture network topology modeling module is configured to represent a fracture network as a dynamic graph structure based on the standardized data set; in the dynamic graph structure, nodes correspond to fracture intersection points or end points, edges correspond to fracture segments, node features include normal stress, strain energy density and water head, edge features include mechanical opening, roughness coefficient and length, and edge weight is calculated based on hydraulic opening. The seepage stress coupling simulation module is configured to input the dynamic graph structure into a graph neural network, simulate bidirectional feedback between seepage and stress through a message passing mechanism, and output first updated node features; wherein the node features include predicted displacement and predicted water head; The graph structure adaptive adjustment module is configured to predict expansion behavior of the fissure based on the first updated node features, obtain a prediction result, update hydraulic opening and roughness coefficient of the fissure according to the prediction result, adjust edge features and edge weights of the dynamic graph structure, and obtain an adjusted graph structure; The multi-round feature iteration optimization module is configured to input the adjusted graph structure into the graph neural network, simulate bidirectional feedback between seepage and stress through the message passing mechanism, and output second updated node features; The failure mode intelligent identification module is configured to identify a seepage front based on the predicted water head in the second updated node features, calculate Gaussian curvature of the seepage front and a clustering coefficient of the fissure network, and determine a failure mode of the slope according to comparison between the Gaussian curvature and the clustering coefficient and a critical value; The slope safety early warning decision module is configured to calculate a safety coefficient of the slope based on the failure mode and the predicted displacement in the second updated node features, and trigger a warning when the safety coefficient is lower than a threshold. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to implement the method in any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 5.

Images