Roadbed slope stability intelligent prediction method based on deep learning

Abstract

The invention relates to a subgrade slope stability intelligent prediction method based on deep learning, and belongs to the technical field of data processing, and the method comprises the following steps: 1, constructing a multi-field coupling data collection network, and forming a dynamic coupling data set with synchronous time response; 2, a meta learning-dynamic graph Transform hybrid model is constructed, rapid adaptation of a new slope scene is achieved through meta learning, the dynamic graph Transform updates rock mass unit mechanical association in real time, the dynamic graph Transform captures multi-field coupling data features, and a Hoek-Brown criterion and a damage evolution equation are embedded to serve as double physical constraint layers; 3, a causal discovery algorithm is introduced to identify key stability factors, a multi-objective evolutionary algorithm is adopted to optimize meta-learning-dynamic graph Transform hybrid model hyper-parameters, and a dynamic loss function is constructed by using project full life cycle risk cost; 4, inputting real-time monitoring data, outputting data, and cooperatively updating multi-engineering data; the method has the beneficial effects that the causal effect in a group is quantified, hyper-parameter optimization is guided, and the prediction performance is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to an intelligent prediction method for roadbed slope stability based on deep learning. BACKGROUND

[0002] The roadbed slope is an important part of the traffic infrastructure (such as highways and railways), and its stability directly affects the safety of the engineering structure and the operation efficiency. Influenced by factors such as geological conditions, hydrological environment, climate change and vehicle load, the roadbed slope is prone to instability disasters such as landslides and collapses, which not only causes economic losses, but also may cause casualties. Therefore, accurate prediction of the stability of the roadbed slope is the key to ensuring the safe operation of the roadbed engineering.

[0003] The traditional prediction method for the stability of the roadbed slope mainly includes the limit equilibrium method, the numerical simulation method and the empirical formula method. The limit equilibrium method (such as the Swedish slice method and the Bishop method) calculates the safety factor by establishing a mechanical equilibrium equation of the slope, but this method assumes that the rock-soil body of the slope is an ideal rigid body, ignores the nonlinear mechanical properties of the rock-soil body and the spatial and temporal variability of the parameters, resulting in a deviation between the prediction result and the actual situation; the numerical simulation method (such as the finite element method and the discrete element method) can simulate the stress and strain distribution of the slope, but it needs complex parameter setting and mesh division, and has low calculation efficiency, which is difficult to meet the real-time prediction demand; the empirical formula method needs to rely on the experience summary of engineering personnel, has poor adaptability to different geological conditions, and the prediction accuracy is greatly affected by subjective factors.

[0004] With the development of sensing technology and artificial intelligence technology, machine learning methods have been applied to the prediction of the stability of the roadbed slope. For example, support vector machines and random forest models learn the mapping relationship between historical data and the stability of the slope for prediction, but such models are difficult to process multi-dimensional time series data and cannot effectively capture the dynamic rules of parameter changes over time, and the prediction performance in complex engineering scenarios still needs to be improved. SUMMARY

[0005] The application provides an intelligent prediction method for the stability of the roadbed slope based on deep learning, which is used to solve the technical problem of poor prediction performance in the existing prediction of the stability of the roadbed slope, and improves the prediction performance by quantifying the causal effect within the group and guiding the optimization of hyperparameters.

[0006] In order to achieve the above purpose, the application realizes the following technical scheme:

[0007] An intelligent prediction method for the stability of the roadbed slope based on deep learning, comprising the following steps:

[0008] Step 1: Construct a multi-field coupled data acquisition network to synchronously acquire rock mass microcrack evolution image data, in-situ stress field distribution data, hydrology-weather-seismic multi-source disturbance data, and construction disturbance time series data, forming a dynamic coupled data set with time response synchronization;

[0009] Step 2: Construct a meta-learning-dynamic graph Transformer hybrid model, where meta-learning realizes rapid adaptation to new slope scenes, dynamic graph Transformer updates rock mass unit mechanics correlations in real time, dynamic graph Transformer captures multi-field coupled data characteristics, and Hoek-Brown criterion and damage evolution equation are embedded as double physical constraint layers;

[0010] Step 3: Introduce a causal discovery algorithm to identify key stabilizing factors, use a multi-objective evolutionary algorithm to optimize the hyperparameters of the meta-learning-dynamic graph Transformer hybrid model, and construct a dynamic loss function combining the risk cost of the entire project life cycle;

[0011] Step 4: Input real-time monitoring data, output decision-level prediction results including safety factor spatiotemporal distribution, instability time probability density, potential sliding surface three-dimensional morphology, and optimal reinforcement timing, and realize multi-project data collaborative update through federated learning mechanism.

[0012] Optionally, in Step 1, introduce damage accumulation variables and time delay attenuation variables to synchronize the time response of multi-field coupled data;

[0013] Based on continuous damage mechanics, supplement rock mass damage accumulation variables to generate the cumulative process of microcrack expansion to mesostructure degradation and then to macro damage.

[0014] Further, the specific process of time delay attenuation is as follows:

[0015] First step: Determine the baseline law of attenuation without time difference through the measured attenuation function;

[0016] Second step: Perform triggering or modulation of attenuation by time delay and triggering time through the time delay triggering function;

[0017] Third step: Obtain the actual attenuation result after time delay triggering modulation by calculating the final time delay attenuation amount. By introducing time delay triggering, it is convenient to correct the disturbance data set.

[0018] Optionally, in Step 2, for the rapid adaptation of meta-learning to new slope scenes, a cyclic rapid adaptation is adopted, and for new tasks , a second-order derivative is calculated based on samples for optimization:

[0019] ;

[0020] wherein, The updated parameters; These are the parameters before the update; The learning rate is the step size for controlling parameter updates, and it is adjusted adaptively.

[0021] The first-order gradient is the loss function. Regarding parameters The first-order partial derivative measures the loss function as the parameters change. The instantaneous rate of change of is the core term of gradient descent, guiding the parameters to move in the direction of reducing loss;

[0022] The second gradient is the loss function. Regarding parameters The matrix formed by the second-order partial derivatives describes the rate of change of the first-order gradient itself, reflecting the loss function. The curvature;

[0023] Second-order gradient is introduced It considers the loss function. The curvature's influence on the update allows for more accurate parameter updates, thus enabling rapid adaptation in cycles.

[0024] Optionally, in step 2, a dynamic graph Transformer is introduced as a rock mass element diagram. The dynamic graph Transformer has nodes, and each node has the following characteristics: or ;

[0025] The calculation method for real-time updating of rock mass element mechanical associations is as follows:

[0026] ;

[0027] in, For the exponent term, and Representing the units in the rock mass and unit At any moment Spatial location vector;

[0028] It is a unit and unit At any moment The spatial distance is the magnitude of the vector, reflecting the spatial relative position of the two units;

[0029] Used to describe the attenuation of the interaction strength between two units with distance; It is a width parameter that controls the range of influence of the interaction;

[0030] For indicator functions, Representation unit and unit Feature distance; It is a threshold distance that can be determined by the physical properties of the rock mass; It is an indicator function that filters out valid unit pairs.

[0031] For at any time Below, rock mass unit and unit The effective interaction coefficient between them.

[0032] Optionally, in step 2, the Hoek-Brown criterion constraint layer of the dual physical constraint layer is corrected by three-dimensional stress state. The description of the strength of rock material under three-dimensional stress is corrected by introducing the material's own properties and empirical parameters, so as to adapt to complex engineering stress scenarios.

[0033] Damage evolution is used for seepage-stress coupling. By calculating the acceleration of damage accumulation, the driving force of stress, and the inhibition of static pressure, the influence of stress state, damage history, and volume effect on the damage evolution rate is characterized. It is a progressive damage process in the damage mechanics of continuous media.

[0034] Optionally, in step 3, the step of identifying key stabilizing factors using the causal discovery algorithm is as follows: samples from the estimated dataset fall into the leaf nodes to calculate the causal effect of the leaf nodes. In the subgroups obtained from the leaf nodes, the difference between the predicted values ​​of the outcome variable and the control group is calculated, and then these differences are averaged to obtain the causal effect on the leaf nodes.

[0035] Optionally, in step 3, the hyperparameters of the meta-learning-dynamic graph Transformer hybrid model, which uses a multi-objective evolutionary algorithm to optimize the model, are calculated using non-dominated ranking and crowding.

[0036] The non-dominated sort is divided into different fronts. The first front is divided into non-dominated solutions, the second front is divided into solutions dominated by the first front, and so on.

[0037] The calculation of crowding degree measures the dispersion of solutions within the same frontier. It is obtained by summing the normalized changes of the two-variable functions to obtain the total dynamic change of the element and guide hyperparameter optimization.

[0038] Optionally, in step 3, the dynamic loss function constructed by combining the risk costs of the entire life cycle of the project is to normalize the risk costs of each stage and then sum them up, and calculate the penalty weight of the deviation in the high-risk stage, so as to achieve a balance between prediction accuracy and economic loss control in the model.

[0039] Optionally, in step 4, the following steps are used to generate the model inference and multi-dimensional prediction results:

[0040] Step 4.2a: Real-time data input and feature matching;

[0041] Step 4.2b: Calculation of the spatiotemporal distribution of the safety factor;

[0042] Step 4.2c: Prediction of the probability density of step instability time;

[0043] Step 4.2d: Inversion of the three-dimensional morphology of the potential slip surface;

[0044] Step 4.2e: Optimal reinforcement timing assessment.

[0045] The beneficial effects of this invention are:

[0046] 1. This invention identifies key stabilizing factors by introducing a causal discovery algorithm, calculates the difference between the target variable and the predicted value of the control group, and then averages these differences to obtain the causal effect at the leaf node. Within the scope of subdivided population + estimated dataset, the average deviation between the actual result and the predicted result of the control group is used to quantify the causal signal or causal effect within the population, thereby improving the prediction performance.

[0047] 2. This invention employs a multi-objective evolutionary algorithm to optimize the hyperparameters of the meta-learning-dynamic graph Transformer hybrid model, and uses non-dominated sorting and crowding calculation to enhance the responsiveness to new tasks, guide hyperparameter optimization, and improve prediction performance. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of the workflow of the present invention. Detailed Implementation

[0050] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0051] Example 1;

[0052] like Figure 1 As shown in the figure, this embodiment provides a method for intelligent prediction of roadbed slope stability based on deep learning, including the following steps:

[0053] Step 1: Construct a multi-field coupled data acquisition network to synchronously acquire image data of rock mass micro-fracture evolution, geostress field distribution data, hydrological-meteorological-earthquake multi-source disturbance data, and construction disturbance time series data, forming a dynamic coupled dataset with synchronized time response;

[0054] Step 2: Construct a meta-learning-dynamic graph Transformer hybrid model. Meta-learning enables rapid adaptation to new slope scenarios, while dynamic graph Transformer updates the mechanical correlation of rock mass units in real time. Dynamic graph Transformer captures multi-field coupled data features and embeds the Hoek-Brown criterion and damage evolution equation as a dual physical constraint layer.

[0055] Step 3: Introduce a causal discovery algorithm to identify key stabilizing factors, use a multi-objective evolutionary algorithm to optimize the hyperparameters of the meta-learning-dynamic graph Transformer hybrid model, and construct a dynamic loss function by combining the risk cost of the entire engineering life cycle;

[0056] Step 4: Input real-time monitoring data and output decision-level prediction results including the spatiotemporal distribution of safety factor, probability density of instability time, three-dimensional morphology of potential slip surface and optimal reinforcement timing. Multi-engineering data are updated collaboratively through a federated learning mechanism.

[0057] Example 2;

[0058] Based on Example 1, in step 1, the time response of the multi-field coupled data has nonlinearity (e.g., cracks accelerate after stress exceeds the yield threshold) and time lag (e.g., groundwater pressure changes lag behind rainfall, and stress disturbances lag behind blasting). Damage accumulation variables and time lag decay variables are introduced to synchronize the time response of the multi-field coupled data.

[0059] The damage accumulation variable is used to quantify irreversible damage in multi-field coupled data. Based on continuous damage mechanics, it supplements the rock mass damage accumulation variable. This describes the cumulative process of microcrack propagation → mesoscopic structural degradation → macroscopic damage, and its evolution formula is:

[0060] ;

[0061] in, Indicates the location of the material ,time The cumulative damage variable describes the extent of development of micro-defects (such as cracks and pores) within a material, and its range is [range missing]. ; The initial damage value is derived from the rock mass integrity coefficient. Calibration, ;

[0062] It describes the time from the initial moment. At the time The process of damage accumulation, the integral is the time accumulation of the damage rate; Used to describe the degree of fatigue-related damage; The rate of fatigue-related damage is expressed as time. The derivative of represents the rate of change. The range is arrive For example, the rate of crack propagation in a material under cyclic loading; It is the weighting coefficient of fatigue damage, used to quantify the contribution of fatigue mechanism to total damage;

[0063] It describes damage related to overload; It represents the effective stress, and considers the equivalent stress under complex stress states. For example, the von Mises equivalent stress is used to describe the combined effect of multiaxial stresses. This indicates the yield strength of a material. When the stress is below the yield strength, the material undergoes mainly elastic deformation, and damage develops slowly. When the stress exceeds the yield strength, the material enters the plastic deformation stage, and damage accumulates more rapidly. It ensures that only the stress exceeding the yield point drives damage. This value is 0, meaning there is no damage accumulation under this mechanism; It is the weighting coefficient for overload damage, quantifying the degree to which stress exceeds yield contributes to the total damage.

[0064] It is determined by the accumulation of initial damage and fatigue damage and overload damage over time. It is used to simulate the damage evolution process of materials such as rock and soil under complex loads (such as cyclic load + overload), and provides a theoretical basis for structural life prediction and safety analysis.

[0065] The time-delay decay variable is introduced to correct for the time lag effect after a disturbance. Natural disturbances (e.g., rainfall → groundwater pressure) and construction disturbances (e.g., blasting → stress transfer) have time lags, so time-delay triggering is introduced. ( For the first The lag time of perturbations is used to facilitate the correction of perturbation datasets. :

[0066] ;

[0067] in, It is the final time-delay decay amount, which is the location. and time The function for position place, time The time-delay decay amount, due to the first The attenuation result generated by each time delay source Used to describe the quantitative impact of time delay on the decay process;

[0068] The decay function is a measure of the position, describing the time difference without considering the time delay triggering. place, time The decay law, Also the first The time delay duration of a time delay source is the time difference between the actual triggering and the ideal time-free triggering.

[0069] For time-delay triggering functions, which are either pulse or step functions, describing the first... Each time delay source at the trigger time After triggering, time delay For example, if the attenuation is switched or modulated, No time decay, The decay is triggered only when the time is right. It will exhibit step characteristics;

[0070] During the time delay, the first The modulation effect of each time delay source on the attenuation process is as follows:

[0071] Step 1: First pass To determine the baseline law of decay when there is no time difference;

[0072] Step 2: Reconnect Considering time delay and trigger time Triggering or modulating attenuation is equivalent to putting a time switch on attenuation; attenuation will only take effect when the time delay condition is met.

[0073] The third step, It is the actual attenuation result after the reference attenuation is triggered and modulated by time delay.

[0074] The purpose is to understand the cross-scale transmission of rock mass from microscopic fissures to mesoscopic structures to macroscopic responses. The time delay of disturbance transmission varies at different scales. For example, the time delay of microscopic fissure expansion is measured in seconds, while the time delay of macroscopic stress adjustment is measured in days. The formula is extended to a multi-scale time delay attenuation model, incorporating time delay parameters from microscopic to mesoscopic and from mesoscopic to macroscopic, to more comprehensively characterize the time effect of multi-scale coupling.

[0075] Example 3;

[0076] Based on Example 1, in step 2, meta-learning is an independent machine learning task, such as classification or regression. Each task includes a support set and a query set. The support set consists of a small number of samples (e.g., 10 images) used to quickly adapt to the new task. The query set consists of samples (e.g., 50 images) used to verify the adaptation effect.

[0077] For meta-learning to achieve rapid adaptation to new slope scenarios, a loop-based rapid adaptation method is used. For new tasks... Optimization using second derivatives calculated from samples:

[0078] ;

[0079] in, The updated parameters; These are the parameters before the update; The learning rate is the step size for controlling parameter updates, and it is adjusted adaptively.

[0080] The first-order gradient is the loss function. Regarding parameters The first-order partial derivative measures the loss function as the parameters change. The instantaneous rate of change of is the core term of gradient descent, guiding the parameters to move in the direction of reducing loss;

[0081] The second-order gradient (Hessian matrix) is the loss function. Regarding parameters The matrix formed by the second-order partial derivatives describes the rate of change of the first-order gradient itself, reflecting the loss function. curvature (e.g., loss function) (The steepness of whether it is convex or non-convex).

[0082] Second-order gradient is introduced It considers the loss function. The curvature's influence on the update allows for more accurate parameter updates, thus enabling rapid adaptation in cycles.

[0083] The first-order gradient is the basic update that ensures the direction of the gradient, moving the parameters in the direction that reduces the loss.

[0084] The second-order gradient adjusts the update direction and step size using the Hessian matrix. For example, if the loss function has a large curvature (large absolute value of the second derivative) near the current parameters, it indicates a rapid gradient change. The second-order gradient adjusts the update magnitude to prevent the step size from being too large and overshooting the optimal value, or too small and converging too slowly, allowing the model to adapt to new data. The goal is to enable faster and more accurate adaptation of parameters to new data or tasks through gradient descent with second-order curvature correction.

[0085] The dynamic graph Transformer is a rock mass element graph. The dynamic graph Transformer contains nodes, and each node has the following characteristics: or That is, the characteristics of each node are stress, strain, cohesion, internal friction angle and damage variables, etc.

[0086] For real-time updating of rock mass element mechanical associations:

[0087] ;

[0088] in, For the exponent term, and Representing the units in the rock mass (No. (time delay source) and unit (No. (a time delay source) at time... Spatial position vector (i.e.) and (Composition of coordinates)

[0089] It is a unit and unit At any moment The spatial distance is the magnitude of the vector, reflecting the spatial relative position of the two units;

[0090] This term describes how the interaction strength between two units decreases with distance. The closer the distance, the closer the exponent is to 1 (stronger interaction); the farther the distance, the closer the exponent is to 0 (weaker interaction). It is a width parameter that controls the range of influence of the interaction. The larger the distance, the more gradual the decay of the interaction (far-away units will not interact). The smaller the value, the steeper the decay (near units interact with each other).

[0091] For indicator functions, Representation unit and unit The characteristic distances (such as the effective action distance of the element, the critical failure distance, etc., need to be defined in conjunction with the specific rock mass analysis scenario); It is a threshold distance that can be determined by the physical properties of the rock mass; It is an indicator function that filters out valid unit pairs.

[0092] For at any time Below, rock mass unit and unit The effective interaction coefficient between them.

[0093] when When the indicator function is 1, the two units will interact with each other in a Gaussian-decreasing manner based on their spatial distance;

[0094] when (If the indicator function is 0), then the two units will not have a Gaussian decaying interaction based on spatial distance;

[0095] For discrete element analysis of rock masses (e.g., studying force transmission and damage correlation between elements when rock masses are deformed or damaged), distance attenuation and threshold filtering are used to more accurately simulate the local and near-field interaction characteristics of rock masses.

[0096] The Hoek-Brown criterion constraint layer with dual physical constraint layers (rock mass strength constraint) is a three-dimensional stress state correction:

[0097] ;

[0098] in, The maximum principal stress is the ultimate maximum principal stress at which the material fails. The minimum principal stress, i.e., the confining pressure, reflects the constraint effect on the material; The uniaxial compressive strength of rock is the ultimate strength of a material under uniaxial compression, which is an inherent property of the material itself. These are parameters related to the internal friction characteristics of materials, reflecting the strength contribution characteristics of materials under pressure due to internal friction, interparticle interlocking, etc. The uniaxial tensile strength of rock is the ultimate strength of a material under uniaxial tension, reflecting the material's ability to resist tensile failure. These are parameters related to the tensile properties of materials, used to correct the effect of tension on overall strength; their values ​​vary depending on the material or working condition. It is an empirical index, obtained through experimental fitting, used to match the actual strength variation law of materials, and may have different values ​​for different materials or stress states;

[0099] Under complex stress conditions, the strength of a material is affected by the confining pressure. Stretching and the material's own compressive strength properties Tensile properties and internal characteristics , , Overall impact:

[0100] Confining pressure It increases the strength of the material because confining pressure restricts internal deformation and crack propagation, making the material more difficult to break. Directly with The superposition of related terms reflects the contribution of confining pressure.

[0101] Uniaxial tensile strength It exists both as an independent item and participates in the process of... The relevant corrections reflect the weakening effect of tensile stress on the overall strength of the material, since the tensile strength of a material is usually much weaker than its compressive strength, and tension is more likely to cause failure.

[0102] , , This is to make the formula more accurately match the actual failure law of materials under three-dimensional stress. For different rocks or materials, these parameters need to be determined through experiments.

[0103] The description of the strength of rock materials under three-dimensional stress (tension, compression, and confining pressure) is modified by introducing the material's own properties and empirical parameters to correct the strength law under uniaxial stress, so as to adapt to more complex engineering stress scenarios, such as stress analysis in underground engineering and rock excavation.

[0104] Damage evolution is mainly used for seepage-stress coupling, and the specific damage variable evolution is as follows:

[0105] ;

[0106] Among them, damage variables To describe the material in the first Under a time delay source, the degree of development of internal defects (such as microcracks and voids) A value of 0 indicates no damage. A value of 1 indicates near-complete damage or material failure, used in continuum damage mechanics to characterize damage over time. The evolution law is used to analyze the fatigue, creep, or fracture processes of rock mass materials;

[0107] Damage variable The rate of change over time, i.e., the rate of damage development;

[0108] This is an acceleration term for damage accumulation; It is a material constant and is related to the intrinsic damage sensitivity of the material. This reflects the accelerating effect of damage accumulation, as damage... Increase Decreasing the denominator reduces the overall term and increases the overall term, reflecting the accelerated evolution characteristic of damage: the more severe the damage, the faster the subsequent damage develops. Damage accumulation during fatigue leads to accelerated failure. It is the primary material parameter used to control the steepness of damage acceleration.

[0109] The power-law term is the threshold term; For the first The equivalent stress corresponding to each time delay source simplifies the complex stress state into a single variable, describing the magnitude of the driving force on the damage. The damage threshold stress; These are material- or scale-dependent constants used to adjust the degree to which stress drives damage; It is the second material parameter, which can control the steepness of the damage rate change with stress after the stress exceeds the threshold. This indicates that a non-negative value is taken.

[0110] This is a static pressure suppression term. It is an exponential function; Represents static pressure or volume-dependent stress components; It is the third material parameter, used to describe the strength of the damage suppression effect of the volume-related stress component;

[0111] Volume-dependent stress components inhibit damage development; under the influence of volume-dependent stress components, materials are less prone to microcracks and voids. When the exponent increases, the overall value decreases, resulting in damage. The rate of damage decreases, which is consistent with the principle that compressive stress inhibits damage and tensile stress promotes damage. When the exponential term decreases, the overall exponential term increases, resulting in damage. The rate of increase is consistent with the principle that tensile stress inhibits damage, while compressive stress promotes damage.

[0112] Acceleration through damage accumulation The driving force of stress (the power-law term of the threshold) and the inhibitory force of static pressure (i.e., the inhibitory term of static pressure) comprehensively characterize the influence of stress state, damage history, and volume effect on the damage evolution rate, and represent the progressive damage process in continuous medium damage mechanics.

[0113] Example 4

[0114] Based on Example 1, in step 3, the stabilizing factor (target variable, i.e.) Factors (or factors) are variables that have a direct causal impact on stability (e.g., equipment reliability, process volatility). Causal discovery algorithms are used to extract the causal structure of factors leading to stability from observed data, eliminating spurious variables. .

[0115] The causal discovery algorithm used to identify key stabilizing factors is based on the estimated dataset. The samples fall into the leaf nodes to calculate the causal effect of the leaf nodes:

[0116] ;

[0117] in, This represents the estimate of a causal effect at the leaf node; This represents the estimated sample size corresponding to the leaf node, i.e., the number of samples used to calculate the estimated value of the leaf node. For those that simultaneously belong to the leaf node and are in the estimated dataset Samples in (No. Summation is performed using (each time-delay source); For the sample The result variable or target variable, i.e., the result; Based on covariates or dummy variables , for processing variables When the outcome variable is set to 0 (i.e., the control group or untreated group), The predicted value, It is a prediction function, obtained from regression and machine learning models; For outcome variables With outcome variable The difference between the predicted values, If the result variable is close to 0, then Consistent outcome variables The predicted value.

[0118] Within the subgroups obtained from the leaf nodes, the outcome variable and the target variable are calculated. Predicted values ​​compared to the control group The differences are then averaged to obtain the causal effect at the leaf node. Within the subdivided population + estimated dataset, the average deviation of the actual results from the predicted results of the control group is used to quantify the causal signal within the population.

[0119] For optimizing the hyperparameters of the meta-learning-dynamic graph Transformer hybrid model using a multi-objective evolutionary algorithm, non-dominated sorting and crowding calculation are employed:

[0120] Non-dominated sorting is divided into different fronts (Pareto levels). The first front is a non-dominated solution (i.e. no other solution is better than the two objectives), the second front is a solution dominated only by the first front, and so on.

[0121] Based on the dominance relationship, all parameter solutions are divided into different fronts (levels). The smaller the front number, the higher the quality of the parameter solution.

[0122] The first frontier (Pareto optimal frontier) consists of solutions within the set that do not dominate each other and are not dominated by any other solutions. This is the optimal set of solutions in the entire solution set and is the level that decision-makers are most concerned about.

[0123] The second front is a set of solutions that are dominated only by solutions of the first front (i.e., no solutions of the second, third, ... fronts can dominate them), and the solutions do not dominate each other. It is a set of suboptimal solutions.

[0124] From the third front onwards, and so on, the solution of each front is only dominated by the front with the smaller (better) number, and they do not dominate each other internally.

[0125] The essence of non-dominated sorting is to establish a superior-inferior order for the set of parameter solutions of multi-objective problems. It does not directly provide a final decision (which solution to choose is determined by the decision-maker based on preference), but it selects valuable solutions (i.e. solutions of each frontier) through hierarchical screening, thus avoiding decision-makers blindly choosing from a massive number of parameter solutions.

[0126] Crowding degree calculation measures the dispersion of solutions within the same frontier (avoiding solution clustering and ensuring diversity), for the objective , No. The congestion level of each time-delay source is:

[0127] ;

[0128] in, For the first Individual learning unit (the first) The degree of dynamic change of each time-delay source (e.g., attention head, sub-module, and task-related components); For the first Individual functions, i.e. These are the target parameters for the second material, and they represent the feature mappings of different parameters in meta-learning. These represent two types of core metafunctions, such as task representation functions and dynamic graph update functions; For the first The input of a metafunction at adjacent time steps or states, such as the input difference between the previous task and the next task, or the parameter correlation between the previous state and the next state in meta-learning. For the first The normalized denominator of a function is used to scale the range of variation to the [0,1] interval, eliminating the problem of incomparability of the degree of variation of different functions due to differences in their numerical ranges.

[0129] Is calculating the first The absolute change of a metafunction in adjacent states, such as the difference before and after attention weights, or the difference before and after feature representations;

[0130] It is the first The range of values ​​for the variable functions is normalized to ensure that the changes of different variable functions can be compared on the same scale (for example, if one variable function outputs [0,100] and another variable function outputs [0,1], the relative magnitude of the change is more intuitive after normalization). It involves summing the normalized changes of the two-variable functions to obtain the total dynamic change of the unit.

[0131] pass Numerical value, determine the first The magnitude of adjustment of each meta-learning unit during the meta-learning process is important. A larger magnitude indicates greater sensitivity to new tasks or states, resulting in more drastic adjustments; a smaller magnitude indicates greater stability, ensuring the dynamic nature of quantization.

[0132] like The values ​​remain consistently high (overly dynamic), requiring adjustment of hyperparameters such as update step size and regularization strength to avoid overfitting or instability; if If the learning rate is too low (lacking dynamism), it is necessary to enhance its responsiveness to new tasks, such as by increasing the learning rate or introducing a more flexible structure, which can guide hyperparameter optimization.

[0133] To construct a dynamic loss function by combining the risk cost of the entire project lifecycle. It is a summation of normalized risk costs at each stage, reflecting the penalty weight for deviations in high-risk stages:

[0134] ;

[0135] in, Risk cost loss (Loss) is one of the objectives that needs to be minimized during model training, reflecting the penalty imposed on the model by the economic losses caused by prediction bias. The total number of stages in the entire life cycle of a project (e.g., design, production, operation and maintenance, and decommissioning stages). (each stage)

[0136] For the first The triggering time of each time delay source, triggering time As the first Each life cycle stage (e.g.: For the design phase, (For the production stage)

[0137] For the first The weighting coefficients for each lifecycle stage reflect the importance of a particular stage within the entire lifecycle (e.g., the maintenance and operation stage, with its long duration and significant cost impact, receives a higher weight). Different stages have varying impacts on deviation stability and economic losses. During the maintenance and operation stage, equipment continues to run; if predicted deviations lead to failures, the losses (repair, downtime) far exceed the cost of correcting design deviations during the design stage. Therefore, the maintenance and operation stage... The impact will be greater, and greater emphasis will be placed on the accuracy of predictions at this stage.

[0138] For the first Each life cycle stage, prediction bias is The actual risk cost at that time (e.g., maintenance and downtime losses caused by downtime due to forecasting errors during the operation and maintenance phase). yes The stability prediction bias at each life cycle stage can measure the prediction bias. In the The actual economic losses incurred during this phase.

[0139] The maximum possible risk cost over the entire lifecycle (i.e., the normalization factor, to avoid excessive differences in cost values ​​at different stages, which could lead to instability in the loss function).

[0140] It is due to different stages The numerical range can vary greatly (e.g., design phase costs are in the tens of thousands of yuan, while operation and maintenance phase costs are in the millions of yuan). After normalization of the (maximum cost over the entire cycle), the costs of each stage can be compared within [0,1], avoiding the stage with the largest value from dominating the loss function.

[0141] This involves summing the weighted normalized risk costs across all stages of the entire lifecycle to obtain the final risk cost loss. During model training, it is necessary to simultaneously minimize the base prediction loss (e.g., MAE, MSE) and This allows the model to strike a balance between predictive accuracy and economic loss control.

[0142] Example 5;

[0143] Based on Example 1, in step 4, the preprocessing and standardization of real-time monitoring data adopts the following steps:

[0144] Step 4.1a: Data Access and Format Unification; Establish a multi-source monitoring data access interface to support real-time data transmission from terminals such as sensors (e.g., displacement gauges, stress sensors, piezometers), meteorological stations, and earthquake monitoring equipment. Unstructured data (e.g., raw voltage signals from sensors) and semi-structured data (e.g., CSV format data from meteorological stations) output by different devices will be uniformly converted into JSON format to ensure that data fields (including monitoring timestamps, monitoring location coordinates, data type identifiers, and numerical values) are standardized and consistent.

[0145] Step 4.1b: Abnormal data cleaning and repair; A hybrid cleaning strategy based on statistical thresholds and the isolated forest algorithm is adopted. First, the data is initially screened by setting reasonable thresholds (e.g., removing outliers exceeding ±3 standard deviations based on the equipment range and historical data fluctuation range); then, the isolated forest model is used to identify hidden abnormal data (e.g., continuous invariant values ​​and sudden jump values ​​caused by sensor failure); for missing data, the repair method is selected according to the missing data rate. When the missing rate is <5%, linear interpolation is used; when 5% ≤ missing rate ≤ 20%, LSTM time series prediction model is used for completion; when the missing rate is >20%, the monitoring point is marked and equipment maintenance reminder is triggered.

[0146] Step 4.1c: Data standardization and spatiotemporal alignment; The cleaned data is standardized by using the Z-Score standardization formula to eliminate the influence of different units (e.g., displacement unit mm, stress unit MPa). At the same time, based on the GPS coordinates and timestamps of the monitoring points, the discrete monitoring data is mapped to the preset three-dimensional grid model of the slope, so as to achieve accurate alignment of the data in the spatial dimension (grid cell) and the temporal dimension (unified time step, e.g., 10 minutes / step), forming a standardized spatiotemporal dataset.

[0147] The following steps are used for model inference and multi-dimensional prediction result generation:

[0148] Step 4.2a: Real-time data input and feature matching; The standardized spatiotemporal dataset is input into the optimized meta-learning-dynamic graph Transformer hybrid model. The model quickly matches similar features (such as slope lithology, slope ratio, and monitoring data fluctuation patterns) between the current slope scenario and the historical training scenario through the meta-learning module, and calls the pre-training parameters of the corresponding scenario; The dynamic graph Transformer module updates the mechanical correlation graph between rock mass units based on real-time data, captures the dynamic features of multi-field coupling (stress-seepage-disturbance), and performs mechanical constraint calculations in combination with a dual physical constraint layer (Hoek-Brown criterion and damage evolution equation).

[0149] Step 4.2b: Calculation of the spatiotemporal distribution of safety factor; The model calculates the safety factor of each grid cell using the simplified Bishop method based on the mechanical parameters of the rock mass unit (e.g., cohesion, internal friction angle, derived from the double physical constraint layer) and the real-time stress state (stress, seepage pressure); The safety factor values ​​between grid cells are supplemented by spatiotemporal interpolation algorithms (e.g., Kriging interpolation), generating a spatiotemporal distribution map of the safety factor of the entire slope, clearly showing the safety status of different locations (e.g., slope top, slope waist, slope toe) at different time steps (e.g., safety factor > 1.2 indicates stability, 1.0 ≤ safety factor ≤ 1.2 indicates critical stability, and safety factor < 1.0 indicates instability).

[0150] Step 4.2c: Prediction of instability time probability density; Based on the time-series change trend of the safety factor, a survival analysis model (e.g., Weibull distribution model) is introduced to transform the process of safety factor decline into an instability risk probability function; 1000 sets of random disturbance scenarios (e.g., sudden rainfall, small earthquake) are generated through Monte Carlo simulation, the time nodes of slope instability under different scenarios are calculated, the instability time probability density curve is obtained statistically, and the time interval with a high probability of instability is identified (e.g., the time period in the next 72 hours when the probability of instability reaches 30%).

[0151] Step 4.2d: 3D morphological inversion of potential slip surface; using the damage degree data of rock mass units output by the model (calculated by the damage evolution equation), select units with damage degree > threshold (e.g., damage variable D > 0.6) as candidate units of potential slip surface; use the 3D Delaunay triangulation algorithm to reconstruct the surface of the candidate units, and combine the slope geological structure (e.g., distribution of faults and weak interlayers) to correct the reconstruction results, generate a 3D morphological model of potential slip surface, and mark key parameters such as maximum depth, length, and orientation of slip surface.

[0152] Step 4.2e: Optimal reinforcement timing assessment; Construct a reinforcement cost-risk loss coupling function, where reinforcement cost includes reinforcement material costs and construction costs (which vary over time, such as increased construction costs during the rainy season), and risk loss includes economic losses caused by instability (such as road damage and equipment damage) and personnel casualty risks (calculated based on the density of personnel activity below the slope); By comparing the sum of reinforcement cost and risk loss at different time points (such as the present, the next 24 hours, and the next 48 hours), select the time point with the lowest total cost as the optimal reinforcement timing, and output the benefit ratio (the ratio of the reduction in risk after reinforcement to the reinforcement cost) corresponding to different reinforcement timings.

[0153] For collaborative data updates across multiple projects under a federated learning mechanism, the following steps are adopted:

[0154] Step 4.3a: Local model parameter extraction and encryption; Each participating engineering site (e.g., slope monitoring points on different highway sections) extracts the parameter update amount generated by the local model based on real-time monitoring data (e.g., attention weight adjustment value of dynamic graph Transformer, scene matching coefficient change value of meta-learning module); Homomorphic encryption algorithm (e.g., Paillier encryption algorithm) is used to encrypt the parameter update amount to ensure that the privacy of local data is not leaked (only the encrypted parameters are transmitted, and the original monitoring data is not transmitted).

[0155] Step 4.3b: Parameter aggregation and global model update; The Federated Learning Center node receives the encrypted parameter updates from each engineering site, and uses the Federated Avg algorithm to aggregate the parameters, that is, to allocate weights according to the sample data volume (monitoring data scale) of each site, and feeds back the aggregated global parameters to each engineering site. Each site updates its local model based on the global parameters, realizing cross-engineering collaborative optimization of the model.

[0156] Step 4.3c: Verification and iteration of the effectiveness of collaborative updates; After each collaborative update, each engineering site uses a local validation dataset (historical monitoring data that did not participate in model training) to test the model's prediction accuracy (e.g., safety factor prediction error, instability time prediction deviation); The validation results are fed back to the federated learning center, which calculates the average and variance of the prediction accuracy of each site to determine whether the collaborative update is effective (e.g., if the average prediction accuracy improvement is >5%, it is considered effective); If effective, it enters the next round of collaborative update cycle (e.g., once a week); If ineffective, the reasons are analyzed (e.g., low data quality of a certain site, unreasonable parameter weight allocation) and the collaborative strategy is adjusted (e.g., removing abnormal sites, optimizing the weight allocation formula) before re-aggregating parameters.

[0157] The feedback for decision-level output and engineering applications is as follows:

[0158] The multi-format output of results includes the prediction results such as the spatiotemporal distribution of safety factor, probability density of instability time, three-dimensional morphology of potential slip surface, and optimal reinforcement timing. These results are output in three forms: visual reports (e.g., line graphs showing the time series changes of safety factor, three-dimensional models showing slip surface morphology), text reports (including risk level assessment and reinforcement recommendations), and early warning signals (e.g., yellow warnings triggered when critical stability is reached, and red warnings triggered when the risk of instability is high). The results can be pushed to management personnel and construction teams through multiple channels such as the engineering management platform (Web), mobile APP, and on-site early warning terminals.

[0159] In engineering applications and data feedback, the construction team formulates reinforcement plans based on the optimal reinforcement timing and potential slip surface parameters (e.g., setting anti-slip piles at the slip surface location, spraying and anchoring support on the slope surface), and feeds back reinforcement construction data (e.g., construction time, reinforcement material usage, post-construction monitoring data) to the model; the reinforced data is used as new training samples to supplement the dynamic coupling dataset for subsequent iterative optimization of the model, forming a closed-loop management of prediction-decision-construction-feedback-optimization.

[0160] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope described in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent prediction of roadbed slope stability based on deep learning, characterized in that, Includes the following steps: Step 1: Construct a multi-field coupled data acquisition network to synchronously acquire image data of rock mass micro-fracture evolution, geostress field distribution data, hydrological-meteorological-earthquake multi-source disturbance data, and construction disturbance time series data, forming a dynamic coupled dataset with synchronized time response; Step 2: Construct a meta-learning-dynamic graph Transformer hybrid model. Meta-learning enables rapid adaptation to new slope scenarios, while dynamic graph Transformer updates the mechanical correlation of rock mass units in real time. Dynamic graph Transformer captures multi-field coupled data features and embeds the Hoek-Brown criterion and damage evolution equation as a dual physical constraint layer. Step 3: Introduce a causal discovery algorithm to identify key stabilizing factors, use a multi-objective evolutionary algorithm to optimize the hyperparameters of the meta-learning-dynamic graph Transformer hybrid model, and construct a dynamic loss function by combining the risk cost of the entire engineering life cycle; Step 4: Input real-time monitoring data and output decision-level prediction results including the spatiotemporal distribution of safety factor, probability density of instability time, three-dimensional morphology of potential slip surface and optimal reinforcement timing. Multi-engineering data are updated collaboratively through a federated learning mechanism.

2. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 1, characterized in that, In step 1, damage accumulation variables and time-delay decay variables are introduced to simultaneously perform time response of multi-field coupled data; Based on continuous damage mechanics, we supplement the cumulative variable of rock mass damage to generate the cumulative process of micro-cracks expanding to meso-structural deterioration and then to macro-damage.

3. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 2, characterized in that, The specific process of time-delay decay is as follows: Step 1: Determine the baseline decay pattern when there is no time difference by measuring the decay function; Step 2: Use the time-delay trigger function to trigger or modulate the decay based on the time delay and trigger time. The third step is to calculate the final time-delay attenuation amount to obtain the actual attenuation result after time-delay trigger modulation. By introducing time-delay triggering, it is easier to correct the disturbance dataset.

4. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 1, characterized in that, In step 2, the rapid adaptation of meta-learning to new slope scenarios is achieved through iterative rapid adaptation; for new tasks... Optimization is achieved by calculating the second derivative using samples: ； in, The updated parameters; The parameters were before the update; The learning rate is the step size for controlling parameter updates, and it is adjusted adaptively. The first-order gradient is the loss function. Regarding parameters The first-order partial derivative measures the loss function as the parameters change. The instantaneous rate of change of is the core term of gradient descent, guiding the parameters to move in the direction of reducing loss; The second gradient is the loss function. Regarding parameters The matrix formed by the second-order partial derivatives describes the rate of change of the first-order gradient itself, reflecting the loss function. The curvature; Second-order gradient is introduced It considers the loss function. The curvature's influence on the update allows for more accurate parameter updates, thus enabling rapid adaptation in cycles.

5. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 1, characterized in that, In step 2, a dynamic graph Transformer is introduced as a rock mass element diagram. The dynamic graph Transformer has nodes, and each node has the following characteristics: or ; The calculation method for real-time updating of rock mass element mechanical associations is as follows: ； in, For the exponent term, and Representing the units in the rock mass and unit At any moment Spatial location vector; It is a unit and unit At any moment The spatial distance is the magnitude of the vector, reflecting the spatial relative position of the two units; Used to describe the attenuation of the interaction strength between two units with distance; It is a width parameter that controls the range of influence of the interaction; For indicator functions, Representation unit and unit Feature distance; It is a threshold distance, determined by the physical properties of the rock mass; It is an indicator function that filters out valid unit pairs. For at any time Below, rock mass unit and unit The effective interaction coefficient between them.

6. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 1, characterized in that, In step 2, the Hoek-Brown criterion constraint layer with dual physical constraint layers is corrected using three-dimensional stress state. The description of the strength of rock materials under three-dimensional stress is corrected by introducing the material's own properties and empirical parameters, so as to adapt to complex engineering stress scenarios. Damage evolution is used for seepage-stress coupling. By calculating the acceleration of damage accumulation, the driving force of stress, and the inhibition of static pressure, the influence of stress state, damage history, and volume effect on the damage evolution rate is characterized. It is a progressive damage process in the damage mechanics of continuous media.

7. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 1, characterized in that, In step 3, the step of identifying key stabilizing factors using the causal discovery algorithm is as follows: samples from the estimated dataset fall into the leaf nodes to calculate the causal effect of the leaf nodes. In the subgroups obtained from the leaf nodes, the difference between the target variable and the predicted value of the control group is calculated, and then these differences are averaged to obtain the causal effect on the leaf nodes.

8. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 1, characterized in that, In step 3, the hyperparameters of the meta-learning-dynamic graph Transformer hybrid model, optimized using a multi-objective evolutionary algorithm, are calculated using non-dominated ranking and crowding degree: The non-dominated sort is divided into different fronts. The first front is divided into non-dominated solutions, the second front is divided into solutions dominated by the first front, and so on. The calculation of crowding degree measures the dispersion of solutions within the same frontier. It is obtained by summing the normalized changes of the two-variable functions to obtain the total dynamic change of the element and guide hyperparameter optimization.

9. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 1, characterized in that, In step 3, the dynamic loss function is constructed by combining the risk costs of the entire life cycle of the project. This involves normalizing the risk costs of each stage and summing them, and then calculating the penalty weight for deviations in high-risk stages. This allows the model to achieve a balance between prediction accuracy and economic loss control.

10. The intelligent prediction method for roadbed slope stability based on deep learning according to claim 1, characterized in that, In step 4, the generation of model inference and multi-dimensional prediction results is performed using the following steps: Step 4.2a: Real-time data input and feature matching; Step 4.2b: Calculation of the spatiotemporal distribution of the safety factor; Step 4.2c: Prediction of the probability density of step instability time; Step 4.2d: Inversion of the three-dimensional morphology of the potential slip surface; Step 4.2e: Optimal reinforcement timing assessment.

Images