Prediction method and system for slope catastrophe

Abstract

The invention discloses a side slope catastrophe prediction method and system, and relates to the technical field of geological disaster prevention and control. The method comprises the steps that the slope geometry of a target slope is obtained, and the change range of rock-soil mechanical parameters of the target slope is determined; carrying out numerical simulation of a landslide dynamic process through an improved particle finite element method based on the variation range of side slope geometry and rock-soil mechanical parameters so as to construct a training database; the improved particle finite element method adopts a node integration technology to eliminate variable mapping errors in grid updating, adopts a non-slip complementary contact theory to process a boundary contact problem, and solves the problem through a convex optimization solver; the training database comprises input parameters and corresponding landslide results; training the composite artificial neural network model based on the training database to obtain a trained prediction model; working condition parameters of a target slope are obtained and input into the prediction model, and a landslide catastrophe prediction result is obtained. According to the method, efficient and rapid prediction of slope catastrophe can be realized on the premise of ensuring high precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological disaster prevention, in particular to a prediction method and system for slope disaster. BACKGROUND

[0002] In recent years, the prediction of large deformation and failure process of slope instability has attracted increasing attention, which is extremely important for disaster prevention and mitigation. With the development of numerical simulation technology, dynamic simulation of the whole process of soil movement after slope instability has become a research hotspot, and its accuracy directly affects the reliability of engineering safety design.

[0003] Currently, a variety of numerical methods have been applied to slope large deformation analysis, such as material point method (MPM), smoothed particle hydrodynamics (SPH), and particle finite element method (PFEM). These methods can reproduce the dynamic process of landslide to some extent when simulating the large deformation behavior of geotechnical materials. However, these existing numerical methods face significant limitations in practical application, especially in terms of computational efficiency, which cannot meet the engineering requirements. Specifically, existing methods often require several hours or even days of computing time when simulating a single working condition. This makes it impractical to conduct systematic sensitivity analysis or uncertainty assessment on multiple parameter combinations, and the huge demand for computing resources seriously restricts the rapid judgment and early warning response to slope landslide risk under engineering construction and environmental changes.

[0004] Therefore, the current technical problem of low computational efficiency in the field of slope large deformation disaster prediction needs to be solved to break through the bottleneck of time cost in existing simulation methods and achieve rapid and efficient dynamic prediction, providing a feasible path for engineering design optimization and risk precise control. SUMMARY

[0005] To solve the problems in the prior art, the present application provides a prediction method and system for slope disaster, which can achieve efficient and rapid prediction of slope disaster while ensuring high precision.

[0006] To achieve the above-mentioned purpose, the present application provides a prediction method for slope disaster, comprising: obtaining the slope geometry of the target slope and determining the variation range of the rock-soil mechanics parameters of the target slope; based on the slope geometry and the variation range of the rock-soil mechanics parameters, performing numerical simulation of the dynamic process of landslide by an improved particle finite element method to construct a training database; the improved particle finite element method adopts node integration technology to eliminate variable mapping error in grid updating, adopts non-sliding complementary contact theory to handle boundary contact problems, and solves through a convex optimization solver; the training database includes input parameters and corresponding landslide results; training a composite artificial neural network model based on the training database to obtain a trained prediction model; obtaining working condition parameters of the target slope and inputting the working condition parameters to the prediction model to obtain a landslide disaster prediction result.

[0007] Optionally, the numerical simulation of the landslide dynamic process is performed by using the improved particle finite element method, including: discretizing the slope geometry into a group of particle clouds; determining a plurality of parameter combinations based on the variation range of the geotechnical mechanics parameters; for each parameter combination, performing the following simulation process: assigning material mechanics properties to the particle clouds according to the current parameter combination; in each time step starting from an initial time step, the following steps are repeatedly performed until a preset termination condition is met: reconstructing the calculation boundary based on the particle clouds of the current time step; generating a background mesh with the calculation boundary as a constraint; creating a smoothing field based on the nodes of the background mesh, and calculating strain by using a node integration technique; based on the calculated strain and the material mechanics properties, constructing an elastoplastic control equation and a contact constraint into a second-order cone programming problem and solving the problem; updating the state and position of the particle clouds according to the solving result.

[0008] Optionally, the determining of the plurality of parameter combinations based on the variation range of the geotechnical mechanics parameters includes: performing full-factorial experimental design on the variation range of the geotechnical mechanics parameters to generate parameter combinations covering all possible cases.

[0009] Optionally, the calculating of strain by using the node integration technique includes: based on the smoothing field, calculating the strain at each mesh node as an area-weighted average value of the strains of all adjacent elements; the smoothing field is defined by a polygon boundary formed by connecting the midpoints of the sides of the triangular elements in the background mesh to the centers of the elements.

[0010] Optionally, the constructing of an elastoplastic control equation and a contact constraint into a second-order cone programming problem and the solving of the problem based on the calculated strain and the material mechanics properties include: determining a stress state based on the calculated strain and the material mechanics properties; using the Hellinger-Reissner variational principle and an implicit time integration method, discretizing the elastoplastic dynamic control equation containing the stress state into a variable-based minimum-maximum value optimization problem; introducing contact constraints as non-slip complementary contact conditions of the optimization problem; reconfiguring the optimization problem containing non-slip complementary contact conditions into a second-order cone programming problem in a standard form through mathematical transformation by introducing auxiliary variables; calling a convex optimization solver to solve the second-order cone programming problem.

[0011] Optionally, the composite artificial neural network model comprises a slope stability prediction sub-network and a landslide damage consequence prediction sub-network; the slope stability prediction sub-network is configured to predict a safety factor of the slope according to the input parameters; and the landslide damage consequence prediction sub-network is configured to predict a dimensionless landslide disaster consequence index according to the safety factor and the input parameters.

[0012] Optionally, the input parameters comprise a slope depth, a slope gradient and a dimensionless stability parameter; and the prediction result comprises one or more of a safety factor of the slope, a sliding distance, a sliding depth, a slope top influence distance and a sliding soil volume.

[0013] Optionally, the slope geometry comprises a slope height, a slope gradient and a slope bottom depth; and the geotechnical mechanics parameters comprise a cohesion, an internal friction angle and a unit weight.

[0014] The application further provides a slope disaster prediction system, comprising: a parameter acquisition unit configured to acquire a slope geometry of a target slope and determine a variation range of geotechnical mechanics parameters of the target slope; a database construction unit configured to perform numerical simulation of a landslide dynamic process based on the slope geometry and the variation range of the geotechnical mechanics parameters by using an improved particle finite element method to construct a training database; the improved particle finite element method adopts a node integration technique to eliminate variable mapping error in grid updating, adopts a non-slip complementary contact theory to process a boundary contact problem, and is solved by a convex optimization solver; and the training database comprises input parameters and corresponding landslide results; a model training unit configured to train a composite artificial neural network model based on the training database to obtain a trained prediction model; a prediction execution unit configured to acquire working condition parameters of the target slope and input the working condition parameters into the prediction model to obtain a landslide disaster prediction result.

[0015] Optionally, in the numerical simulation of the landslide dynamic process by using the improved particle finite element method, the database construction unit is specifically configured to: discretize the slope geometry into a group of particle clouds; determine a plurality of parameter combinations based on the variation range of the geotechnical mechanics parameters; for each parameter combination, perform the following simulation process: impose material mechanical properties on the particle cloud according to the current parameter combination; In each time step starting from an initial time step, the following steps are cyclically executed until a preset termination condition is met: A boundary is calculated based on the particle cloud of the current time step; A background mesh is generated with the boundary as a constraint; A smoothing field is created based on the nodes of the background mesh, and a strain is calculated using a node integration technique; Based on the calculated strain and the material mechanical properties, an elastoplastic control equation and a contact constraint are jointly constructed into a second-order cone programming problem and solved; The state and position of the particle cloud are updated according to the solution result.

[0016] According to the specific embodiments provided by the present application, the following technical effects are disclosed: The prediction method of slope disaster provided by the present application effectively solves the core technical problems of low calculation efficiency and difficulty in meeting the real-time needs of engineering in the prior art by constructing a high-quality training database using an improved particle finite element method and training a composite artificial neural network model. The method places the high-precision but time-consuming physical numerical simulation process in the database construction stage. The improved particle finite element method used in this stage eliminates the variable mapping error caused by frequent mesh updating through a node integration technique, and accurately handles complex boundary contact problems through second-order convex optimization theory, thereby ensuring the accuracy and reliability of the constructed database, laying a solid physical foundation for subsequent model training.

[0017] In the prediction stage, the trained composite artificial neural network model is used to replace the traditional numerical solver. Thanks to the high-speed forward reasoning ability of the neural network model, the disaster prediction result of the new target slope working condition can be obtained in a short time. This realizes the leapfrog efficiency improvement from "several hours or even several days" of traditional simulation to "minutes", making rapid evaluation and prediction of slope stability and damage consequences a reality. The present application breaks through the bottleneck of time cost in the prior art, making large-scale and systematic parameter sensitivity analysis and uncertainty quantification evaluation feasible, providing a powerful technical tool for precise and rapid control of engineering design optimization and risk. BRIEF DESCRIPTION OF DRAWINGS

[0018] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings, in which like reference characters refer to like parts throughout the several views, and in which the exemplary embodiments of the present application are shown.

[0019] Figure 1A method flowchart of the method for predicting slope disaster according to an embodiment of the present application is shown in FIG. 1. Figure 2 A technical roadmap according to an embodiment of the present application is shown in FIG. 2. Figure 3 A schematic diagram of the nodal integration technique according to an embodiment of the present application is shown in FIG. 3. Figure 4 A schematic diagram of the boundary treatment method according to an embodiment of the present application is shown in FIG. 4. Figure 5 A schematic diagram of the composite neural network architecture according to an embodiment of the present application is shown in FIG. 5. Figure 6 A schematic diagram of the GUI interface according to an embodiment of the present application is shown in FIG. 6. Figure 7 A schematic diagram of a slope case and its calculation grid according to an embodiment of the present application is shown in FIG. 7. Figure 8 Schematic diagrams of the slope landslide process and its movement speed at simulation times of 1.01s, 1.50s, 3.04s, 5.77s and 8.01s according to an embodiment of the present application are shown in FIG. 8. Figure 9 A module structure schematic diagram of the prediction system for slope disaster according to an embodiment of the present application is shown in FIG. 9. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0021] Please refer to Figure 1 and Figure 2 , Figure 1 A method flowchart of the method for predicting slope disaster, Figure 2 A technical roadmap.

[0022] The method for predicting slope disaster comprises: S101: obtaining the slope geometry of a target slope and determining the variation range of the rock-soil mechanics parameters of the target slope.

[0023] Illustratively, the slope geometry comprises the slope height, slope degree and slope bottom depth; and the rock-soil mechanics parameters comprise the cohesion, internal friction angle and specific gravity.

[0024] When performing slope disaster prediction, the first step is to accurately define the analysis object. This involves obtaining the slope geometry of the target slope and determining the range of variation of its geotechnical parameters, laying the foundation for subsequent high-precision numerical simulations. Specifically, for a slope problem with a specific geometric shape, it is necessary to determine its slope geometry and boundary conditions. This slope geometry forms the basis for constructing the numerical model and typically includes key parameters such as slope height, slope gradient, and slope base depth, which together define the macroscopic morphology of the slope. For example, for a slope with a height of 5 meters and a slope gradient of 45 degrees, it can be modeled using standard finite element method software. The geometric domain is discretized into a mesh, and the mesh nodes constitute an initial particle cloud, thus providing a geometric model for subsequent particle finite element method calculations.

[0025] After determining the slope geometry, it is necessary to further determine the mechanical properties of the soil and rock materials constituting the slope. These geotechnical mechanical parameters are intrinsic factors affecting slope stability, and determining their reasonable range of values ​​directly affects the reliability of the simulation results. For example, cohesion can vary between 18 kPa and 22 kPa, the internal friction angle between 10° and 16°, and the unit weight is determined according to the typical range based on the soil type. This process can refer to authoritative national standards, relevant geological survey data, and engineering experience. Key geotechnical mechanical parameters to be analyzed may also include cohesion. internal friction angle Elastic modulus E Poisson's ratio And strain softening mechanical parameters, etc.

[0026] S102: Based on the variation range of slope geometry and geotechnical parameters, numerical simulation of landslide dynamics is carried out using an improved particle finite element method to construct a training database.

[0027] Among them, the improved particle finite element method uses nodal integration technology to eliminate variable mapping errors in mesh updates, adopts non-slip complementary contact theory to handle boundary contact problems, and solves the problem through a convex optimization solver; the training database includes input parameters and corresponding landslide results.

[0028] In this application, after obtaining the range of variation of the slope geometry and geotechnical parameters of the target slope, numerical simulations of the landslide dynamic process are performed to construct a high-quality training database. The aim is to systematically change the geotechnical parameters of a slope with a specific geometry and run simulations using an improved high-precision particle finite element method, thereby constructing a dedicated database for subsequent training of predictive models capable of predicting landslide catastrophic consequences. The construction of this database is a crucial foundation for the rapid and efficient implementation of the entire prediction method; its data quality directly determines the accuracy and reliability of the final prediction model.

[0029] This training database contains clearly defined inputs and outputs. Input parameters are systematically varying combinations of parameters, covering different slope geometric parameters such as slope height, slope gradient, and slope base depth, as well as key geotechnical parameters such as cohesion, internal friction angle, and unit weight. Outputs are the landslide results obtained through high-precision simulation calculations for each set of input parameters. These results specifically include sliding distance, sliding depth, influence distance at the top of the slope, and volume of the sliding soil mass. By storing all input parameters and their corresponding simulation results, a comprehensive training database suitable for machine learning models is constructed.

[0030] To ensure the physical authenticity and computational accuracy of the generated data, this invention employs an improved particle finite element method. This high-precision simulation method possesses two core advantages. See also Figure 3 and Figure 4 First, it employs nodal integration, concentrating all variables and operations on the grid nodes. Existing methods, during large deformation simulations, require continuous regeneration of the computational grid. This necessitates mapping state variables from the old grid integration points to the new grid. This "variable mapping" process not only introduces numerical computation errors but also significantly reduces overall computational efficiency. In contrast, this invention, through nodal integration, ensures that state variables are attached to the nodes and have no direct dependency on the grid. During large deformation calculations, no variable mapping operation is required after grid reconstruction, fundamentally eliminating the resulting errors and improving computational accuracy and reliability.

[0031] Secondly, this invention employs non-slip complementary contact theory to handle boundary contact problems and utilizes a convex optimization solver for efficient and accurate solutions. Traditional particle finite element methods (PEMs) are complex and difficult to apply when dealing with contact problems between rigid boundaries such as soil and rock masses and structures. To address this issue, this invention establishes a clear and efficient contact handling method within a second-order cone programming framework. It uniformly and directly expresses the non-embedding conditions between deformable bodies and rigid boundaries, as well as the strength criteria at the interface, as standard second-order cone constraints. This approach avoids complex contact algorithms and can directly and accurately solve soil-structure interaction problems, significantly improving the model's practicality and robustness in real-world engineering applications.

[0032] S103: Train a composite artificial neural network model based on the training database to obtain a trained prediction model.

[0033] The composite artificial neural network model includes a slope stability prediction subnetwork and a landslide failure consequence prediction subnetwork. The slope stability prediction subnetwork is used to predict the safety factor of the slope based on the input parameters. The landslide failure consequence prediction subnetwork is used to predict the dimensionless landslide disaster consequence index based on the safety factor and the input parameters.

[0034] For example, the input parameters include dimensionless slope depth, slope and stability parameters; the prediction results include one or more of the following: slope safety factor, sliding distance, sliding depth, slope crest influence distance and sliding soil volume.

[0035] The composite artificial neural network model consists of two functionally interconnected sub-networks, serving as a highly efficient surrogate model to replace computationally intensive traditional numerical simulations of landslide motion processes. The first sub-network is a stability prediction model, which predicts the slope's safety factor based on input parameters. This model uses dimensionless slope depth, gradient, and stability parameters as its main input features. The second sub-network is a failure consequence prediction model. If the safety factor is less than 1, indicating slope instability, the model uses the safety factor of the instable slope predicted by the first sub-network, combined with inputs such as slope and stability parameters, to predict a set of dimensionless landslide catastrophic consequences indicators, such as normalized sliding distance and sliding depth.

[0036] During the training phase, the constructed complete database needs to be divided into training and testing sets according to a predetermined ratio, such as 8:2. The training process uses the backpropagation algorithm and a specific optimizer, such as the Levenberg-Marquardt optimizer, to iteratively optimize the network weights. The goal of training is to minimize the error between the neural network's predicted values ​​and the true values ​​obtained from high-precision numerical simulations in the database, until the model reaches a preset accuracy requirement on the testing set, such as a deterministic coefficient R² greater than 0.99. Only then is the training considered complete and the model parameters saved.

[0037] Predictive models such as Figure 5 As shown, the trained prediction model can inherit the physical realism of high-precision numerical simulations while achieving an order-of-magnitude improvement in computational efficiency. It reduces the time required for a single-condition dynamic simulation, which previously took hours or even days, to minutes or even milliseconds. This makes timely assessment and rapid prediction of slope hazards possible, providing an efficient tool for parameter sensitivity analysis and uncertainty assessment in engineering projects.

[0038] S104: Obtain the working parameters of the target slope and input them into the prediction model to obtain the landslide disaster prediction results.

[0039] After training the composite artificial neural network model, the final application and prediction phase begins. The trained prediction model is then deployed to practical engineering applications to achieve rapid and accurate assessment of the catastrophic consequences of target slopes under specific working conditions. For example... Figure 6As shown, to facilitate user operation, the trained composite artificial neural network model can be deployed and integrated into a graphical user interface (GUI). This interface integrates the trained prediction model, allowing users to perform professional risk assessments without requiring extensive numerical simulation knowledge.

[0040] In practical applications, users only need to input the working parameters of the target slope in the graphical user interface. These parameters typically include the slope's geometric parameters, such as slope height, slope gradient, and slope base depth, as well as the material parameters of the soil and rock mass, such as cohesion, internal friction angle, and unit weight. The types and formats of these parameters are consistent with the input parameters used when building the training database. Upon receiving the input working parameters, the system immediately performs calculations using its built-in prediction model. Thanks to the efficiency of the prediction model, the entire calculation process is completed quickly, ultimately outputting the landslide disaster prediction result. See also... Figure 5 The prediction model includes a slope stability prediction subnetwork ANN-I and a landslide failure consequence prediction subnetwork ANN-II, as shown in the figure. Representing the v The first hidden layer w ;like Figure 5 As shown, slope depth ,slope and stability parameters As input to ANN-I, the safety factor of the slope is predicted. Then the safety factor of the slope ,slope and stability parameters As input to ANN-II, the indicators of large deformation failure consequences are predicted. This refers to indicators of landslide disaster consequences; specific landslide disaster prediction results may include the slope safety factor. And a series of specific indicators of landslide disaster consequences, such as sliding distance. Sliding depth The distance affected by the top of the slope and the volume of sliding soil .

[0041] In this way, the present invention can achieve timely assessment and prediction of slope disasters. When analyzing the risk of landslides under certain conditions, the method can quickly predict possible consequences based on existing working parameters and prediction models, providing an effective reference for disaster prevention and mitigation and disaster prevention scheme design.

[0042] In one embodiment, the numerical simulation of landslide dynamics using the improved particle finite element method includes: The slope geometry is discretized into a set of particle clouds; Based on the range of variation of geotechnical parameters, multiple sets of parameter combinations are determined; For each set of parameters, perform the following simulation procedure: Assign material mechanical properties to the particle cloud based on the current parameter combination; Within each time step starting from the initial time step, the following steps are executed repeatedly until a preset termination condition is met: Reconstructing the computational boundary of the particle cloud based on the current time step; The background mesh is generated using the computational boundary as a constraint; A smooth domain is created based on the nodes of the background mesh, and strain is calculated using nodal integration techniques. Based on the calculated strain and material mechanical properties, the elastoplastic control equations and contact constraints are combined to construct a second-order cone programming problem and then solved. Update the state and position of the particle cloud based on the solution results.

[0043] The preset termination condition is that the calculation reaches the preset termination time and / or the absolute value of the velocity field is less than the preset threshold.

[0044] Based on the variation range of geotechnical parameters, the above-mentioned multiple parameter combinations are determined, including: Full factorial experimental design is performed within the range of variations of geotechnical mechanics parameters to generate parameter combinations that cover all possible scenarios.

[0045] For example, the above-mentioned calculation of strain using nodal integration techniques includes: Based on the smoothing domain, the strain at each grid node is calculated as the area-weighted average of the strains of all adjacent elements; the smoothing domain is defined by the polygonal boundary formed by connecting the midpoints of each side of the triangular elements in the background grid with the centroid of the element.

[0046] For example, based on the calculated strain and material mechanical properties, the above-mentioned elastoplastic governing equations and contact constraints are jointly constructed into a second-order cone programming problem and solved, including: Based on the calculated strain and material mechanical properties, the stress state is determined; Using the Hellinger-Reissner variational principle and implicit time integration method, the elastoplastic dynamic control equations involving stress states are discretized into a variable-based minimum-maximum optimization problem. Contact constraints are introduced as non-slip complementary contact conditions for the optimization problem. By introducing auxiliary variables and performing mathematical transformations, the optimization problem involving non-slip complementary contact conditions is reconstructed into a standard form second-order cone programming problem. The convex optimization solver is invoked to solve the second-order cone programming problem.

[0047] In application, four to five values ​​can be uniformly selected for each geotechnical parameter within its range, where the first value is the minimum and the last value is the maximum. Full factorial experiments are designed for all possible parameter combinations. For each set of parameters, a complete dynamic simulation is performed using the improved particle finite element method (PFEM) of this invention.

[0048] The simulation steps for each landslide process are as follows: 1) Computation initialization: In the first time step after the computation begins At that time, the slope geometry is discretized into a set of initial particle clouds. The particles are arranged with precise particle spacing. The coordinates and velocity information of all particles are input into the program and initialized. Then, the identifier of the computational problem domain, model parameters, and the density and external force information of all particles are input.

[0049] 2) Determining the problem domain boundary: Obtaining the boundary at the current time step The following represents a set of particle clouds representing the computational object. Then, the particle cloud is geometrically analyzed using the α-shape algorithm, through a preset... α Using the value as the criterion, edges or triangles that meet the geometric constraints are selected from the Delaunay triangulation formed by the particle cloud, and these edges or triangles are combined into a continuous geometric figure. Finally, the outer contour of the geometric figure is extracted and defined as the computational boundary of the problem domain, so as to accurately reconstruct the macroscopic geometric shape of the computational object at the current moment from the discrete particles.

[0050] 3) Generation of the computational mesh: Using the problem domain boundary determined in step 2) as a mandatory constraint, the computational mesh is generated for the current time step. All particles below The constrained Delaunay triangulation algorithm is executed on the mesh nodes to generate a set of non-overlapping triangular elements that completely cover the entire computational domain. This process ensures that all constraint boundary segments exist as edges in the generated mesh, resulting in a geometrically accurate, topologically valid background mesh that can be directly used for subsequent finite element numerical calculations. .

[0051] 4) Creation of the nodal smoothing domain: based on the finite element mesh generated in step 3). A smooth domain with "no overlap" and "no gap" is generated, centered on each grid node. The smooth domain is formed by multiple straight boundary segments that connect the midpoints of the sides of the triangular cells to the centroids of the cells.

[0052] Therefore, the first The strain (or smooth strain) in each integration region is all the strain related to the nodes. The weighted average strain of adjacent third-order elements is expressed by the formula: (1) in, The strain in this integration region; k It is the integration region number; i This is the number of the finite element; the triangular elements connected to the nodes total [number missing]. indivual; It is the nodal integration region k area It is a finite element i area It is a unit i strain, It is the nodal integration region k The strain-displacement matrix; This represents the nodal displacements of the relevant finite element elements within the integration region. Therefore, after employing the nodal integration technique, the strain can be expressed as: (2) in, ε For overall response, For global node displacement, The global strain-displacement matrix is... A set of.

[0053] 5) Variational Form of the Governing Equations: Based on the generalized Hellinger-Reissner (HR) variational principle and implicit time integration method, this invention expresses the elastoplastic dynamics governing equations as an equivalent minimum-maximum optimization problem. Subsequently, the triangular element mesh generated in the previous steps is used... The minimum-maximum optimization problem is spatially discretized using the corresponding smooth domain. During this process, physical fields such as displacement, strain, stress, and dynamics are approximated through nodal variables, ultimately forming a discretized min-max optimization problem: (3) in: (4) In the formula, the subscript The index represents the state variable at the next time step. The table represents the state variables at the current time step. The superscript T represents the transpose of the matrix, and Δ represents the variable increment. Represents the computational domain. Represents the boundary of the computational domain; Represents the global node displacement matrix for the next time step; A matrix representing the global node displacement increments at the next time step; Represents the global strain-stress parameter matrix; It is the shape function matrix of stress; This represents the global stress component vector at the next time step. This represents the global stress component vector at the current time step. This represents the global stress component increment vector for the next time step; Represents the global compliance matrix; It is the unit compliance matrix; t Indicates time; It is the time step; Represents the global dynamic component vector; Represents the global dynamics coefficient matrix; It is the shape function matrix of the displacement; It is the density of the material; Indicates equivalent density; This is the global node displacement increment vector; Represents the global equivalent shape function integral matrix; Represents the global equivalent external load vector; Represents the global equivalent volume force matrix. ,in It is per unit weight. yes The global velocity vector at time t; It is the standard The first time integration parameter of the method, It is the standard The second time integration parameter of the method, and The values ​​range from [0, 1]. Represents the global boundary force matrix. ,in It is a boundary force; It is the yield equation. The underscore in the text indicates integration over the region using a nodal integration scheme; The underscore in the text indicates that integration is performed on the region using a nodal integration scheme.

[0054] 6) Introduction of Rigid Body Boundary Contact Conditions: To handle the interaction between deformable bodies and rigid boundaries in the computational model, this invention introduces non-embedding conditions. These conditions use a set of mathematical constraints to ensure that the nodes of the deformable body do not penetrate the rigid surface. (5) in, It is a boundary node I displacement increment, n I It is a boundary nodeI The outer normal vector, p I It is a boundary node I Contact force, It is a boundary node I The initial gap distance, It is a boundary node I The current gap distance, with the superscript T representing the transpose of the matrix. Mathematical constraints link the displacement increment of the nodes, the boundary normal vector, the initial gap, and the normal contact force, and apply corresponding complementary conditions. After considering the complementary relationship formula (5) between the mechanical model and the boundary or structural contact problem, the calculation formula can be extended to: (6) in, N b It is the total number of all nodes involved in boundary contact, including both already contacted and potentially contacted nodes; the boundary normal vector is collected in... n In the middle, the tangential vector is collected in In the middle; the contact forces in the normal and tangential directions form vectors respectively. p and q ; It is a strength criterion on the boundary surface.

[0055] 7) Reconstruction of the Optimization Problem into a Second-Order Cone Programming Problem: The optimization problem containing contact conditions is mathematically transformed into an equivalent minimization problem, and then further reconstructed into a standard form of second-order cone programming (SOCP) problem. The key to this reconstruction process is the introduction of auxiliary variables, transforming the quadratic terms in the original objective function into rotating quadratic cone constraints in SOCP. The final result is a standard SOCP problem that can be directly solved using the interior point method. (7) in, This represents the second-order cone-of-revolution constraint condition for the first cone of revolution. This represents the second-order rotating cone constraint condition for the second rotating cone. Let be the first independent variable of the first rotating cone. Let the second independent variable of the first rotating cone be placed in the objective function. The third independent variable of the first rotating cone is fixed at 1 here; Let be the first independent variable of the second rotating cone. Let the second independent variable of the second rotating cone be placed in the objective function. The third independent variable of the second rotating cone is fixed at 1 here; express m +2-dimensional real space, where m It is the dimension of the vector.

[0056] 8) Particle State Update and Mesh Reset: The SOCP problem is solved using the convex optimization solver MOSEK to obtain the stress increment. The dual variables corresponding to the equilibrium equation constraints are extracted from the solution results as the displacement increment. The spatial position of each particle is updated to obtain the next time step. Particle cloud coordinates Simultaneously, determine the information changes of each particle, such as velocity and strain. After the particle state update is complete, clear the background mesh temporarily generated for this calculation. and its smoothing region Only the updated particle information is retained for the next calculation cycle.

[0057] 9) Time-step iteration: Repeat all calculation steps 2) to 8) above. This loop continues until the calculation reaches the preset termination time and the absolute value of the velocity field is less than the threshold, thus completing the dynamic or quasi-static process simulation of the entire soil flow and large deformation problem.

[0058] 10) Parametric Systematic Simulation and Training Database Construction: Execute the calculation process from steps 1) to 9) to perform systematic numerical simulations of working conditions with various parameter combinations. The input parameter combinations cover different slope geometric parameters (slope height). ,slope and slope bottom depth ) and soil and rock mechanical parameters (cohesion) internal friction angle and severe The simulation results include the slope safety factor. F s Indicators of large deformation and failure consequences of landslides (landslide distance) Sliding depth The distance affected by the top of the slope and the volume of sliding soil The input parameters and their corresponding simulation results are stored to construct a comprehensive training database. To improve the model's generalization ability, the database data is dimensionless. (Slope depth) Defined as the depth of the slope base With slope height The ratio: (8) Stability parameters The calculation formula is: (9) Dimensionless large deformation failure consequence index The correspondence between these correspondences and the specific physical consequences of damage is as follows: (10) In the formula, This is an index variable used to distinguish different types of indicators of the consequences of large deformation damage in landslides. This is the sliding distance; The sliding depth; The distance affected by the top of the slope; Let be the volume of the sliding soil mass. The dimensionless index can be predicted by the surrogate model using the above formula. The actual damage results are determined.

[0059] The above prediction method will be further introduced through a specific example below.

[0060] The first step is to determine the specific slope geometry model and calculate the initial geometry of the problem domain, such as... Figure 7 The image shows a slope with an upper base L1 = 20m, an initial height L2 = 5m, a lower base L3 = 25m, and a slope a = 45°. It can be modeled using standard finite element software, divided into 11,356 finite element triangular elements, forming 5,869 nodal particles. Each mesh node represents an initial particle cloud. Record the initial coordinates of all particles. .

[0061] Secondly, the range of variation of mechanical parameters was determined. Based on relevant data and preliminary geological survey data, the slope was identified as a clay slope, and the elastic modulus was estimated. The value ranges from 500 kPa to 1500 kPa, and the Poisson's ratio is... The value ranges from 0.2 to 0.4, indicating cohesion. c The value range is between 18 kPa and 22 kPa, and the internal friction angle is... The value range is between 16° and 28°. Five levels are evenly selected for each factor within its value range. Using a combination of 4 factors and 5 levels, 625 combinations of mechanical parameters are formed.

[0062] A dedicated training database is further constructed. For each parameter combination, the safety factor of the slope is first calculated using the classic strength reduction method. Then, the improved PFEM of this invention is used to simulate the complete dynamic landslide process of the unstable slope. This method employs nodal integration techniques to eliminate variable mapping errors and combines second-order conical convex optimization theory to accurately handle boundary contact, thereby ensuring the physical authenticity and computational accuracy of the generated data. For each parameter combination, the improved PFEM of this invention is used for a complete dynamic simulation. This method employs nodal integration techniques to eliminate variable mapping errors, non-slip complementary contact theory to handle boundary contact problems, and a convex optimization solver for efficient solution, thereby ensuring the physical authenticity and computational accuracy of the generated data.

[0063] The input parameters are systematically varied, and all discrete values ​​of the above parameters are combined to cover all possible cases. For each parameter combination, the proposed PFEM high-precision simulation process is executed to generate data samples for the slope stability prediction sub-network (ANN-I). Subsequently, samples of all instability conditions are selected from this data sample to form a sub-database specifically used to train the landslide failure consequence prediction sub-network (ANN-II).

[0064] Finally, the neural network was trained. For a simple two-dimensional problem, there are four varying mechanical parameters, each with five possible values, resulting in 5 × 5 × 5 × 5 = 625 parameter combinations. The generated training database was divided into training and test sets in an 8:2 ratio. The Adam optimization algorithm was used, iteratively optimizing the network weights through backpropagation to minimize the mean square error between the displacement increment predicted by the neural network and the actual displacement increment calculated by PFEM. This process continued until the model's performance metrics on the test set (e.g., coefficient of determination R² > 0.99) reached the preset accuracy requirement. Training was then complete, and the model was saved. A visualization of the prediction results can be found in [link to image]. Figure 8 .

[0065] Corresponding to the aforementioned application function implementation method embodiments, the present invention also provides a slope disaster prediction system and corresponding embodiments.

[0066] Please see Figure 9 , Figure 9 This is a schematic diagram of the module structure of a slope disaster prediction system.

[0067] Slope disaster prediction systems include: The parameter acquisition unit 91 is used to acquire the slope geometry of the target slope and determine the range of variation of the geotechnical parameters of the target slope. Database construction unit 92 is used to numerically simulate the dynamic process of landslides using an improved particle finite element method based on the variation range of slope geometry and geotechnical parameters, in order to build a training database. The improved particle finite element method uses nodal integration technology to eliminate variable mapping errors in mesh updates, uses non-slip complementary contact theory to handle boundary contact problems, and solves the problem using a convex optimization solver. The training database includes input parameters and corresponding landslide results. Model training unit 93 is used to train a composite artificial neural network model based on a training database to obtain a trained prediction model; The prediction execution unit 94 is used to obtain the working parameters of the target slope and input them into the prediction model to obtain the landslide disaster prediction results.

[0068] In one embodiment, in the numerical simulation of landslide dynamics using the improved particle finite element method, the aforementioned database construction unit 92 is specifically used for: The slope geometry is discretized into a set of particle clouds; Based on the range of variation of geotechnical parameters, multiple sets of parameter combinations are determined; For each set of parameters, perform the following simulation procedure: Assign material mechanical properties to the particle cloud based on the current parameter combination; Within each time step starting from the initial time step, the following steps are executed repeatedly until a preset termination condition is met: Reconstructing the computational boundary of the particle cloud based on the current time step; The background mesh is generated using the computational boundary as a constraint; A smooth domain is created based on the nodes of the background mesh, and strain is calculated using nodal integration techniques. Based on the calculated strain and material mechanical properties, the elastoplastic control equations and contact constraints are combined to construct a second-order cone programming problem and then solved. Update the state and position of the particle cloud based on the solution results.

[0069] In one embodiment, regarding determining multiple parameter combinations based on the variation range of geotechnical parameters, the aforementioned database construction unit 92 is specifically used for: Full factorial experimental design is performed within the range of variations of geotechnical mechanics parameters to generate parameter combinations that cover all possible scenarios.

[0070] In one embodiment, in calculating strain using nodal integration techniques, the database construction unit 92 is specifically used for: Based on the smoothing domain, the strain at each grid node is calculated as the area-weighted average of the strains of all adjacent elements; the smoothing domain is defined by the polygonal boundary formed by connecting the midpoints of each side of the triangular elements in the background grid with the centroid of the element.

[0071] In one embodiment, based on the calculated strain and material mechanical properties, the elastoplastic governing equations and contact constraints are jointly constructed into a second-order cone programming problem and solved using a second-order cone convex optimization solver. Specifically, the database construction unit 92 is used for: Based on the calculated strain and material mechanical properties, the stress state is determined; Using the Hellinger-Reissner variational principle and implicit time integration method, the elastoplastic dynamic control equations involving stress states are discretized into a variable-based minimum-maximum optimization problem. Contact constraints are introduced as non-slip complementary contact conditions for the optimization problem. By introducing auxiliary variables and performing mathematical transformations, the optimization problem involving non-slip complementary contact conditions is reconstructed into a standard form second-order cone programming problem. The convex optimization solver is invoked to solve the second-order cone programming problem.

[0072] Regarding the system in the above embodiments, the specific manner in which each unit module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0073] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for predicting slope disasters, characterized in that, include: Obtain the slope conditions of the target slope and determine the range of variation of the geotechnical parameters of the target slope; Based on the range of variation of the slope geometry and geotechnical parameters, a numerical simulation of the landslide dynamic process is performed using an improved particle finite element method to construct a training database. The improved particle finite element method uses nodal integration to eliminate variable mapping errors in mesh updates, employs non-slip complementary contact theory to handle boundary contact problems, and solves the problem using a convex optimization solver; the training database includes input parameters and corresponding landslide results. A composite artificial neural network model is trained based on the training database to obtain a trained prediction model; The working parameters of the target slope are obtained and input into the prediction model to obtain the landslide disaster prediction results.

2. The method for predicting slope disasters according to claim 1, characterized in that, Numerical simulation of landslide dynamics using an improved particle finite element method, including: The slope geometry is discretized into a set of particle clouds; Based on the range of variation of the aforementioned geotechnical parameters, multiple sets of parameter combinations are determined; For each set of parameters, perform the following simulation procedure: Assign material mechanical properties to the particle cloud based on the current parameter combination; Within each time step starting from the initial time step, the following steps are executed repeatedly until a preset termination condition is met: Reconstructing the computational boundary of the particle cloud based on the current time step; A background mesh is generated using the computational boundary as a constraint; A smoothing domain is created based on the nodes of the background mesh, and strain is calculated using nodal integration techniques. Based on the calculated strain and the mechanical properties of the material, the elastoplastic control equation and the contact constraints are combined to construct a second-order cone programming problem and then solved. Update the state and position of the particle cloud based on the solution results.

3. The method according to claim 2, characterized in that, The determination of multiple parameter combinations based on the variation range of the aforementioned geotechnical parameters includes: Full factorial experiments were designed within the range of variation of the aforementioned geotechnical parameters to generate parameter combinations that cover all possible scenarios.

4. The method according to claim 2, characterized in that, The method of calculating strain using nodal integration includes: Based on the smoothing domain, the strain at each grid node is calculated as the area-weighted average of the strains of all adjacent elements; the smoothing domain is defined by the polygonal boundary formed by connecting the midpoints of each side of the triangular element in the background grid with the centroid of the element.

5. The method according to claim 2, characterized in that, Based on the calculated strain and the aforementioned material mechanical properties, the elastoplastic governing equations and contact constraints are combined to construct a second-order cone programming problem and solved, including: Based on the calculated strain and the mechanical properties of the material, the stress state is determined; Using the Hellinger-Reissner variational principle and implicit time integration method, the elastoplastic dynamic control equations containing the stress state are discretized into a variable-based minimum-maximum optimization problem. Contact constraints are introduced as non-slip complementary contact conditions for the optimization problem. By introducing auxiliary variables and performing mathematical transformations, the optimization problem involving non-slip complementary contact conditions is reconstructed into a standard form second-order cone programming problem. The convex optimization solver is invoked to solve the second-order cone programming problem.

6. The method for predicting slope disasters according to claim 1, characterized in that, The composite artificial neural network model includes a slope stability prediction subnetwork and a landslide failure consequence prediction subnetwork; the slope stability prediction subnetwork is used to predict the safety factor of the slope based on the input parameters; the landslide failure consequence prediction subnetwork is used to predict dimensionless landslide disaster consequence indicators based on the safety factor and the input parameters.

7. The method for predicting slope disasters according to claim 1, characterized in that, The input parameters include slope depth, slope gradient, and dimensionless stability parameters; the prediction results include one or more of the following: slope safety factor, sliding distance, sliding depth, slope crest influence distance, and sliding soil volume.

8. The method according to claim 1, characterized in that, The slope geometry includes slope height, slope gradient, and slope base depth; the geotechnical parameters include cohesion, internal friction angle, and unit weight.

9. A slope disaster prediction system, characterized in that, include: The parameter acquisition unit is used to acquire the slope geometry of the target slope and determine the range of variation of the geotechnical parameters of the target slope. The database construction unit is used to numerically simulate the dynamic process of landslides using an improved particle finite element method based on the variation range of the slope geometry and geotechnical parameters, in order to construct a training database. The improved particle finite element method uses nodal integration to eliminate variable mapping errors in mesh updates, employs non-slip complementary contact theory to handle boundary contact problems, and solves the problem using a convex optimization solver; the training database includes input parameters and corresponding landslide results. The model training unit is used to train a composite artificial neural network model based on the training database to obtain a trained prediction model. The prediction execution unit is used to obtain the working condition parameters of the target slope and input them into the prediction model to obtain the landslide disaster prediction result.

10. A slope disaster prediction system according to claim 9, characterized in that, In the numerical simulation of landslide dynamics using the improved particle finite element method, the database construction unit is specifically used for: The slope geometry is discretized into a set of particle clouds; Based on the range of variation of the aforementioned geotechnical parameters, multiple sets of parameter combinations are determined; For each set of parameters, perform the following simulation procedure: Assign material mechanical properties to the particle cloud based on the current parameter combination; Within each time step starting from the initial time step, the following steps are executed repeatedly until a preset termination condition is met: Reconstructing the computational boundary of the particle cloud based on the current time step; A background mesh is generated using the computational boundary as a constraint; A smoothing domain is created based on the nodes of the background mesh, and strain is calculated using nodal integration techniques. Based on the calculated strain and the mechanical properties of the material, the elastoplastic control equation and the contact constraints are combined to construct a second-order cone programming problem and then solved. Update the state and position of the particle cloud based on the solution results.

Images