A method and device for constructing a slope creep deformation proxy model and a storage medium

Abstract

A method, apparatus, and storage medium for constructing a surrogate model for slope creep deformation are disclosed. The method includes: constructing a set of numerical samples, each sample containing a first constitutive parameter, external boundary conditions, a first intermediate latent feature, and a first creep deformation time series; the first constitutive parameter includes: static and dynamic first constitutive parameters and steady-state parameters; the first intermediate latent feature is a time-dimensional correlation feature generated during the nonlinear calculation of the constitutive model; performing the following operations on each sample: inputting the static first constitutive parameter into a first learning network to obtain the predicted intermediate latent feature; inputting the dynamic first constitutive parameter, the predicted intermediate latent feature, and the steady-state parameter into a second learning network to obtain the predicted creep deformation time series; optimizing a combined learning network constructed from the first and second learning networks based on the sample, the predicted intermediate latent feature, and the predicted creep deformation time series; and training the combined learning network based on the set of numerical samples and the optimized parameters to obtain the surrogate model.

Description

Technical Field

[0001] This article relates to slope digital twin technology, and more particularly to a method, device and storage medium for constructing a slope creep deformation proxy model. Background Technology

[0002] Instability of reservoir bank slopes not only threatens the safety of people's lives and property, but also affects the normal operation of infrastructure such as hydropower stations. Using numerical simulation to assess slope stability is crucial for the timely identification of potential landslide hazard areas, the prediction of possible landslide events, and the implementation of effective disaster prevention and mitigation measures.

[0003] The reliability of numerical simulations depends on whether the considered physical mechanisms match the actual situation and whether the constitutive parameter values ​​are reasonable. Inverting constitutive parameters based on field deformation monitoring data is an important method to improve the reliability of numerical simulations. Among these methods, surrogate models reflecting slope deformation serve as an efficient simulation tool, acting as a bridge between constitutive parameters and creep deformation during the inversion process. However, with the increasing complexity of field engineering conditions, traditional surrogate models have proven insufficient in handling complex physical processes, often failing to accurately reflect the interaction between variable boundary conditions and underlying physical mechanisms. Summary of the Invention

[0004] This application provides a method, apparatus, and storage medium for constructing a slope creep deformation proxy model, which can improve the prediction accuracy of slope deformation.

[0005] The method for constructing a slope creep deformation proxy model provided in this application includes:

[0006] A set of numerical samples is constructed for training the slope creep deformation surrogate model. Each numerical sample in the set contains a first constitutive parameter, external boundary conditions causing slope creep deformation, a first intermediate latent feature, and a first slope creep deformation time series. The first constitutive parameter is determined by the original constitutive parameters of the creep constitutive model reflecting the slope creep deformation mechanism, including: a time-independent static first constitutive parameter, a time-dependent dynamic first constitutive parameter, and a steady parameter. The first slope creep deformation time series is determined by the slope creep deformation response obtained by the slope computation grid through the slope creep constitutive model. The first intermediate latent feature is a time-dependent feature generated during the nonlinear calculation process of the slope creep constitutive model.

[0007] For each numerical sample, the following operations are performed: The static first constitutive parameters, which are not correlated with the time dimension of the first constitutive parameters, are input into the first learning network to obtain predicted intermediate latent features; the dynamic first constitutive parameters, which are correlated with the time dimension of the first constitutive parameters, the predicted intermediate latent features, and the steady-state parameters in the first constitutive parameters are input into the second learning network to obtain the predicted slope creep deformation time series; based on the numerical sample, the predicted intermediate latent features, and the predicted slope creep deformation time series, the parameters of the combined learning network constructed by the first and second learning networks are optimized.

[0008] The combined learning network is trained based on the numerical sample set and the final optimized parameters of the combined learning network, and the trained combined learning network is used as the surrogate model for slope creep deformation.

[0009] The non-transient computer-readable storage medium provided in this application embodiment stores a computer program that can be executed by a processor to implement a method for constructing a slope creep deformation proxy model.

[0010] The apparatus for constructing a slope creep deformation proxy model provided in this application includes a memory and a processor. The memory stores a computer program, and when the computer program is read and executed by the processor, it can realize the method for constructing a slope creep deformation proxy model.

[0011] Traditional slope creep deformation surrogate models are typically static, meaning the input parameters are static parameters that do not change over time, and the output parameters are also static deformation increments. Such models cannot accurately reflect the interaction between changing boundary conditions and intrinsic physical mechanisms. In contrast, the slope creep deformation surrogate model constructed in this application includes dynamic external boundary conditions that cause slope creep deformation, as well as time-related dynamic parameters, in its input parameters. The output parameters are also a time series of slope creep deformation that changes over time. Therefore, it can accurately reflect the time-dependent deformation of the slope and improve the prediction accuracy of slope creep deformation.

[0012] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. Other advantages of this application can be realized and obtained by means of the solutions described in the description and the accompanying drawings. Attached Figure Description

[0013] The accompanying drawings are used to provide an understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0014] Figure 1 A flowchart illustrating the method for constructing a slope creep deformation proxy model provided in this application embodiment;

[0015] Figure 2 An example diagram illustrating the construction of a slope creep deformation proxy model provided in this application;

[0016] Figure 3 An example diagram of a computational grid for a reservoir bank slope is provided in this application;

[0017] Figure 4 An example diagram of a full-process creep deformation curve is provided for this application;

[0018] Figure 5 An example diagram of a Transformer physics mechanism enhancement module provided in this application;

[0019] Figure 6 An example diagram of a multi-layer stacked GRU network provided in this application;

[0020] Figure 7 Example diagram comparing the predicted and actual values ​​of creep deformation time series at typical GNSS monitoring points using a slope creep deformation surrogate model provided in this application;

[0021] Figure 8 A schematic diagram of a construction device for a slope creep deformation proxy model provided in an embodiment of this application. Detailed Implementation

[0022] This application describes several embodiments, but these descriptions are exemplary and not restrictive, and it will be apparent to those skilled in the art that many more embodiments and implementations are possible within the scope of the embodiments described herein. Although many possible combinations of features are shown in the drawings and discussed in the detailed description, many other combinations of the disclosed features are also possible. Unless specifically limited, any feature or element of any embodiment may be used in combination with, or may replace, any feature or element of any other embodiment.

[0023] This application includes and contemplates combinations of features and elements known to those skilled in the art. The embodiments, features, and elements disclosed in this application can also be combined with any conventional features or elements to form unique inventive solutions. Any feature or element of any embodiment can also be combined with features or elements from other inventive solutions to form another unique inventive solution. Therefore, it should be understood that any feature shown and / or discussed in this application can be implemented individually or in any suitable combination. Therefore, the embodiments are not limited except by the limitations imposed by the appended claims and their equivalents. Furthermore, various modifications and changes can be made within the scope of the appended claims.

[0024] Furthermore, in describing representative embodiments, the specification may have presented methods and / or processes as a specific sequence of steps. However, the method or process should not be limited to the specific order of steps described herein, to the extent that it does not depend on such a specific order. As will be understood by those skilled in the art, other sequences of steps are also possible. Therefore, the specific order of steps set forth in the specification should not be construed as a limitation of the claims. Moreover, the claims concerning the method and / or process should not be limited to the steps performed in the written order, and those skilled in the art will readily understand that these orders can be varied and still remain within the spirit and scope of the embodiments of this application.

[0025] This application provides a method for constructing a slope creep deformation proxy model, such as... Figure 1 As shown, the method includes:

[0026] Step S100 constructs a set of numerical samples for training the slope creep deformation proxy model. Each numerical sample in the set contains a first constitutive parameter, external boundary conditions causing slope creep deformation, a first intermediate implicit feature, and a first slope creep deformation time series.

[0027] The first constitutive parameter is determined by the original constitutive parameters of the slope creep constitutive model that reflects the slope creep deformation mechanism, including: static first constitutive parameters that are not associated with the time dimension, dynamic first constitutive parameters that are associated with the time dimension, and steady parameters;

[0028] A constitutive model is a mathematical model describing the mechanical behavior of a material, reflecting the stress-strain relationship when the material is subjected to force. Constitutive parameters are physical or empirical parameters in the constitutive model, which quantitatively describe the mechanical properties of the material, such as elastic modulus and yield strength. The static first constitutive parameter cannot be directly represented by a time series, as it involves a highly nonlinear calculation process in the constitutive model (which usually reflects a complex physical mechanism). The dynamic first constitutive parameter can be directly represented by a time series. Steady parameters in the constitutive model are usually material parameters that do not change with time.

[0029] The external boundary conditions may include hydraulic boundary conditions (such as actual water storage process curves and precipitation change process curves) and temperature boundary conditions.

[0030] The first slope creep deformation time series is determined by the slope creep deformation response calculated by the slope computational grid through the slope creep constitutive model; the slope computational grid is a mathematical model that discretizes the continuous physical space into a large number of small units by dividing the reservoir bank slope into grids; the interaction between each small unit is calculated according to physical laws, thereby simulating the mechanical behavior of the entire reservoir bank slope;

[0031] The first intermediate implicit feature is the time-dimensional correlation feature generated during the nonlinear calculation process of the slope creep constitutive model;

[0032] Step S101 performs the following operations for each numerical sample:

[0033] Input the static first constitutive parameters that are not related to the time dimension in the first constitutive parameters into the first learning network to obtain the predicted intermediate latent features;

[0034] The dynamic first constitutive parameter associated with the time dimension in the first constitutive parameter, the predicted intermediate latent features, and the steady parameters in the first constitutive parameter are input into the second learning network to obtain the predicted slope creep deformation time series.

[0035] Based on the numerical samples, the predicted intermediate hidden features, and the predicted slope creep deformation time series, the parameters of the combined learning network constructed by the first learning network and the second learning network are optimized.

[0036] Step S102: The combined learning network is trained based on the numerical sample set and the final optimized parameters of the combined learning network, and the trained combined learning network is used as the surrogate model for slope creep deformation.

[0037] Traditional slope creep deformation surrogate models are typically static, meaning the input parameters are static parameters that do not change over time, and the output parameters are also static deformation increments. Such models cannot accurately reflect the interaction between changing boundary conditions and intrinsic physical mechanisms. In contrast, the slope creep deformation surrogate model constructed in this application includes dynamic external boundary conditions that cause slope creep deformation, as well as time-related dynamic parameters, in its input parameters. The output parameters are also a time series of slope creep deformation that changes over time. Therefore, it can accurately reflect the time-dependent deformation of the slope and improve the prediction accuracy of slope creep deformation.

[0038] In one exemplary embodiment, the method further includes:

[0039] Before constructing the numerical sample set for training the slope creep deformation proxy model, a slope computational grid and a slope creep constitutive model are constructed. For example, a slope finite difference model computational grid can be constructed based on engineering geology, hydrogeology, and topographic exploration data. The computational grid includes geological structures such as bedrock, weathered rock mass, slip zone, toppling fault zone, and landslide body. The bottom boundary adopts triaxial displacement constraint, and the lateral boundary adopts normal displacement constraint.

[0040] Correspondingly, the numerical sample set used to train the slope creep deformation surrogate model includes:

[0041] Based on the slope computational grid and the slope creep constitutive model, a set of numerical samples is constructed for training the slope creep deformation surrogate model.

[0042] For example, constructing the slope creep constitutive model includes:

[0043] Acquire full-process slope creep deformation data, which includes slope data before and after slope creep; determine the slope creep constitutive model based on the full-process slope creep deformation data.

[0044] The acquisition of full-process slope creep deformation data includes:

[0045] If the monitored slope creep deformation data is from after the slope creep has occurred, backtrack the monitoring timeline to include slope data prior to the occurrence of the slope creep; then stitch the monitored slope creep deformation data with the backtracked slope data to obtain the full-process slope creep deformation data. For example, the slope creep may include one or more of the following: slope slippage, slope trailing edge tensile creep, lateral edge tensile creep, slope surface tensile creep, and shear creep.

[0046] As an example, the slope creep is described as slow creep. Assuming that the deformation monitoring data obtained through the Global Navigation Satellite System (GNSS) is data after the slow creep has occurred, for the missing historical data before the slow creep, InSAR remote sensing monitoring can be used for retrospective analysis, and a linear correction model of the InSAR observation results can be established based on the deformation monitoring data obtained by the GNSS. The corrected InSAR observation results and the deformation monitoring data obtained by the GNSS are then stitched together to obtain the full-process slope creep deformation data based on the GNSS monitoring points on the slope before and after the slope creep.

[0047] For example, determining the slope creep constitutive model based on the entire process slope creep deformation data includes:

[0048] Based on environmental data (such as engineering geology, hydrogeology, and topographic exploration data) and the entire process of slope creep deformation data, the factors leading to slope creep deformation are identified.

[0049] The creep deformation mechanism of the slope is determined based on the factors that cause the creep deformation of the slope.

[0050] Based on the aforementioned slope creep deformation mechanism, an elastic-viscoplastic slope creep constitutive model considering the aforementioned slope creep deformation mechanism is established.

[0051] The elastic-viscoplastic creep constitutive model integrates three different mechanical behaviors: elasticity, viscosity, and plasticity, and can more comprehensively reflect the creep characteristics of materials under different conditions.

[0052] In an exemplary embodiment, constructing the numerical sample set for training the slope creep deformation surrogate model includes:

[0053] By sampling the original constitutive parameters of the slope creep constitutive model, N is obtained. S The first constitutive parameter of the i-th group; where the first constitutive parameter of the i-th group is denoted as Indicates the static first constitutive parameter; Represents the dynamic first constitutive parameter; θ i This represents the steady-state parameter; for example, N can be obtained by Latin hypercube sampling of the original constitutive parameters. SThe first constitutive parameter of the group can be generated by Latin hypercube sampling, which generates sample points uniformly distributed in the parameter space. The determination method of the original constitutive parameter can be divided into theoretical analysis method and experimental measurement method. The theoretical analysis method relies on the principles of physics and materials science and existing data models, and infers its original constitutive parameter by understanding the deformation mechanism of the material. The experimental measurement method obtains the original constitutive parameter by conducting a series of mechanical tests on the material sample, such as tensile test, compression test, torsion test, etc., and recording the stress-strain response of the material under different conditions.

[0054] For the first constitutive parameter Φ of the i-th group i , will the Convert to time series j represents the time series number, with a total of N. X A time series, where T represents the length of the time series; based on the first constitutive parameter Φ of the i-th group. i The corresponding slope creep deformation response was used to determine the first slope creep deformation time series from the first monitoring point. and the first intermediate latent feature from the second monitoring point Where u represents the sequence number of the first monitoring point, and there are a total of N points. Y There are N first monitoring points, where v represents the sequence number of the second monitoring point, for a total of N. Z The first monitoring point is usually a GNSS monitoring point in the field. The purpose of constructing a slope creep deformation surrogate model is to quickly invert slope parameters, and the target of the inversion is the actual monitored deformation. Therefore, the location of the first monitoring point is usually chosen to correspond to the location of the GNSS monitoring point. The purpose of setting a second monitoring point is to provide a time-series representation (i.e., intermediate implicit features) reflecting complex physical mechanisms during the calculation process with strong nonlinearity in the creep constitutive model. These intermediate implicit features are used to construct the physical loss. By considering the physical process during the training of the slope creep deformation surrogate model, the model accuracy is improved. Therefore, for convenience, during the numerical sample calculation process, the spatial coordinates of the second monitoring point and the first monitoring point can coincide. That is, for a monitoring point on the same spatial coordinate, both the first slope creep deformation time series and the first intermediate implicit feature are derived. However, this does not mean that the first and second monitoring points must coincide in spatial coordinates; they can also not coincide in spatial coordinates.

[0055] Will as well as The i-th numerical sample is obtained by combining the results. in, This represents the external boundary conditions that cause slope creep deformation, where k represents the number of the external boundary conditions, and there are a total of N. H External boundary conditions.

[0056] In an exemplary embodiment, the first learning network is a Transformer physics mechanism enhancement model with a position encoder. The step of inputting the time-independent static first constitutive parameters from the first constitutive parameters into the first learning network to obtain predicted intermediate latent features includes:

[0057] The static first constitutive parameters are input into the constructed Transformer physics mechanism enhancement model with a position encoder to obtain the predicted intermediate latent features.

[0058] The Transformer physics mechanism model refers to a model with a self-attention mechanism. This mechanism enables the model to process all positions of the input data simultaneously, thereby capturing the long-range dependencies of the static first constitutive parameters. This application's embodiments introduce a position encoder into the Transformer physics mechanism model to enhance its functionality. The enhanced Transformer physics mechanism model with a position encoder utilizes the self-attention mechanism and the position encoder to transform static constitutive parameters that are not correlated with the time dimension into dynamic intermediate latent features, thus realizing the transformation of static constitutive parameters from other dimensions to the time dimension.

[0059] In one exemplary embodiment, the method for constructing the Transformer physics enhancement model with a position encoder includes:

[0060] The input data of the Transformer physics enhancement model is position-encoded using a position encoder;

[0061] The position-encoded input data is passed through multiple parallel self-attention layers;

[0062] The output data of the multiple parallel self-attention layers are concatenated through a multi-head self-attention layer;

[0063] The concatenated data is passed sequentially through a feedforward neural network layer and a linear transformation layer to obtain the output data of the Transformer physics mechanism enhancement model.

[0064] For example, based on the Transformer physics mechanism enhancement model with a position encoder constructed by the method described in the above embodiments, the static first constitutive parameter is input into the Transformer physics mechanism enhancement model to obtain the predicted intermediate latent features, including:

[0065] The first constitutive parameter in the i-th group of numerical samples As the input feature vector of the Transformer physics enhancement module, i = i = 1, 2, ..., N S ;

[0066] The input feature vector is processed using a position encoder. Assign location information;

[0067] The position-encoded input feature vector Multiple parallel self-attention layers are used to compute the correlation between different locations;

[0068] The output data of multiple parallel self-attention layers are passed through a multi-head self-attention layer to simultaneously focus on information from different parts of the input feature vector;

[0069] The output data of the multi-head self-attention layer is input into the feedforward neural network;

[0070] The data output from the feedforward neural network is then input into the linear transformation layer to obtain the predicted intermediate hidden features.

[0071] In an exemplary embodiment, the second learning network is a multi-layer stacked GRU network. The step of inputting the dynamic first constitutive parameter associated with the time dimension of the first constitutive parameter, the predicted intermediate latent features, and the steady-state parameters in the first constitutive parameter into the second learning network to obtain the predicted slope creep deformation time series includes:

[0072] The predicted intermediate latent features and the dynamic original features affecting slope creep deformation determined by the dynamic first constitutive parameter are input into a multilayer stacked GRU network. For example, each GRU layer in the multilayer stacked GRU network is followed by a Dropout operation. The Dropout operation is a regularization technique widely used in deep learning. By randomly setting the output of some neurons in the network to zero, overfitting can be prevented. The Dropout operation not only reduces the complex co-adaptation relationship between neurons, but also enhances the generalization ability of the network. In practical applications, especially when training data is limited, the use of the Dropout operation can significantly improve the performance of the multilayer stacked GRU network.

[0073] The output data of the multi-layer stacked GRU network and the constant parameters are input into the fully connected layer.

[0074] The output data of the fully connected layer is used as the predicted slope creep deformation time series.

[0075] In an exemplary embodiment, the parameter optimization of the combined learning network constructed from the first and second learning networks based on the numerical samples, the predicted intermediate latent features, and the predicted slope creep deformation time series includes:

[0076] The prediction loss of the intermediate latent features is determined based on the first intermediate latent features of the numerical samples and the predicted intermediate latent features;

[0077] The prediction loss of the slope creep deformation time series is determined based on the first slope creep deformation time series of the numerical sample and the predicted slope creep deformation time series.

[0078] The combined loss is determined by combining the predicted loss of the intermediate implicit features and the predicted loss of the slope creep deformation time series.

[0079] The parameters of the ensemble learning network are optimized based on the numerical samples and the ensemble loss.

[0080] In an exemplary embodiment, a time-dependent weight adjustment mechanism can be introduced when determining the prediction loss of the slope creep deformation time series, so that the prediction loss of the slope creep deformation time series focuses more on the accuracy of predicting the slope creep deformation time series. An exemplary prediction loss function for the slope creep deformation time series incorporating the time-dependent weight adjustment mechanism is shown below. It can be represented as:

[0081]

[0082] Where T represents the length of the time series; ω(t) represents the weight function associated with time step t; and u represents the sequence number of the first monitoring point (e.g., a GNSS monitoring point), for a total of N. Y The first monitoring point; Here, the Huber loss function is used to control sensitivity to large errors. This represents the predicted deformation value corresponding to the i-th numerical sample obtained at time step t from the u-th first monitoring point (i.e., the predicted slope creep deformation time series obtained through the multi-layer stacked GRU network). This represents the actual deformation value corresponding to the i-th numerical sample obtained at time step t through the u-th first monitoring point (i.e., the first slope creep deformation time series contained in the numerical sample);

[0083] For example, the weight function ω(t) associated with time step t can be expressed as follows:

[0084]

[0085] Where t is the time step, T represents the length of the time series, a and b are weight coefficients, and γ is the non-linear rate of weight growth.

[0086] In an exemplary embodiment, determining the combined loss by combining the prediction loss of the intermediate implicit features and the prediction loss of the slope creep deformation time series includes:

[0087] The combination loss is determined by the following formula:

[0088]

[0089] in, The prediction loss function represents the slope creep deformation time series corresponding to the i-th numerical sample; Let represent the prediction loss function for the intermediate latent feature corresponding to the i-th numerical sample. Z represents the predicted value of the intermediate latent feature corresponding to the i-th numerical sample. i This represents the actual value of the intermediate latent feature corresponding to the i-th numerical sample. Let represent the combined loss function corresponding to the i-th numerical sample; α is a hyperparameter between 0 and 1, used to adjust the weight of the two loss functions in the combined loss function. If α is closer to 1, the combined loss function will place more emphasis on the accuracy of the slope creep deformation time series prediction; if α is closer to 0, the combined loss function will place more emphasis on the accuracy of the intermediate hidden feature prediction.

[0090] In an exemplary embodiment, optimizing the parameters of the combined learning network based on the numerical samples and the combined loss includes:

[0091] Based on the numerical samples and the combined loss, a Bayesian optimizer is used to optimize the parameters of the combined learning network. The Bayesian optimizer can model the dependencies between parameters through a Gaussian process, which makes it suitable for combined learning networks with high computational cost, such as combined learning networks composed of Transformer physical mechanism enhancement models and multi-layer stacked GRU networks.

[0092] For example, when the combined learning network includes a Transformer physical mechanism enhancement model and a multi-layer stacked GRU network, the parameters for optimization may include: the feature dimension d_model of the Transformer physical mechanism enhancement module, the number of heads nhead for multi-head self-attention, and the number of encoders num_encoder_layers; the number of layers num_GRU of the multi-layer stacked GRU network, the number of hidden units hidden_sizes per layer, and the dropout rate dropout_rates after each layer; the weight coefficients a and b of the weight function ω(t), the non-linearity rate γ of weight growth, the threshold δ in the Huber loss function, and the combined loss function. The hyperparameter α; the batch size of the training model, batch_size; and the number of iterations, num_epochs.

[0093] As can be seen from the technical solutions of the above embodiments, the embodiments of this application have the following beneficial effects:

[0094] (1) By using the Transformer physics mechanism enhancement module, static constitutive parameters that are not related to the time dimension can be effectively converted into intermediate implicit features that are related to the time dynamics. This allows the complex dynamic changes driven by the physics mechanism to be captured, thereby improving the sensitivity and prediction accuracy of the slope creep deformation proxy model constructed in this application embodiment to the actual physical process. This not only enhances the understanding of the input data structure of the slope creep deformation proxy model constructed in this application embodiment, but also improves the prediction results of the slope creep deformation time series.

[0095] (2) By using a multi-layer stacked GRU network, multiple types of input data can be processed simultaneously. This design can not only adapt to time-independent steady parameters, but also handle parameters that change over time. In addition, it can also integrate the complex intermediate latent feature time series extracted by the Transformer physics mechanism enhancement module. This integration provides a comprehensive data processing framework, allowing the slope creep deformation proxy model constructed in this application embodiment to capture more detailed temporal dynamic changes and their relationship with slope creep deformation;

[0096] (3) By constructing a combined loss function, the slope creep deformation proxy model constructed in this application embodiment not only emphasizes the accuracy of creep deformation prediction, but also emphasizes the accurate extraction of intermediate hidden features, ensuring that the slope creep deformation proxy model can maintain sensitivity to physical processes while learning data, thereby improving the reliability and generalization ability of prediction; in addition, the introduced time-dependent weight adjustment mechanism can enable the slope creep deformation proxy model to focus on the accuracy of prediction of key deformation stages.

[0097] (4) The slope creep deformation proxy model constructed in the embodiments of this application can be used to quickly invert slope parameters, provide engineers and decision-makers with immediate slope stability risk assessment, reduce potential safety hazards, and has high application value.

[0098] The following section uses the example of a reservoir bank slope in my country as an engineering case study to provide a detailed explanation of the solutions described in the above embodiments:

[0099] like Figure 2 As shown, the steps for constructing a slope creep deformation surrogate model include:

[0100] Step S1: Based on engineering geology, hydrogeology and topographic exploration data, construct a computational grid for the reservoir bank slope; combine on-site slope creep deformation monitoring data to analyze and determine the deformation mechanism of the reservoir bank slope, and establish a slope creep constitutive model that considers the deformation mechanism of the reservoir bank slope.

[0101] Step S2: Based on the reservoir bank slope computational grid and slope creep constitutive model obtained in Step S1, construct a set of numerical samples. Each numerical sample in the set contains a first constitutive parameter, external boundary conditions, a first intermediate implicit feature, and a first slope creep deformation time series. The first constitutive parameter includes: a static first constitutive parameter that is not correlated in the time dimension, a dynamic first constitutive parameter that is correlated in the time dimension, and a steady parameter.

[0102] Step S3: Construct a Transformer physics mechanism enhancement module, using a self-attention mechanism and a position encoder to transform the time-independent static first constitutive parameters into dynamic predictive intermediate latent features;

[0103] Step S4: Construct a multi-layer stacked GRU (Gate Recurrent Unit) network, with each GRU layer followed by a Dropout operation, and an outer fully connected layer following the last GRU layer; input the time series features (such as the water-induced parameter degradation curve) obtained by substituting the dynamic first constitutive parameters associated with the time dimension into the mathematical expression (i.e., the original features in the figure) and the predicted intermediate latent features obtained in step S3 as the time series input to the multi-layer stacked GRU network; use the outer fully connected layer as the input layer for the steady parameters; the output of the multi-layer stacked GRU network is the predicted slope creep deformation time series;

[0104] Step S5: Construct a combined loss function that considers the prediction loss of the intermediate implicit features predicted in step S3 and the prediction loss of the predicted slope creep deformation time series in step S4, and introduce a time-dependent weight adjustment mechanism in the process of calculating the prediction loss of the predicted slope creep deformation time series in step S4.

[0105] Step S6: Based on the numerical sample set obtained in step S2 and the combined loss function constructed in step S5, the hyperparameters of the Transformer-multilayer stacked GRU network composed of steps S3 and S4 are optimized using a Bayesian optimizer.

[0106] Step S7: Based on the numerical sample set obtained in step S2 and the optimal hyperparameters obtained in step S6, train the Transformer-multilayer stacked GRU network composed of steps S3 and S4 to obtain the slope creep deformation surrogate model.

[0107] The input features of the slope creep deformation surrogate model are constitutive parameters and external boundary conditions, and the output feature of the slope creep deformation surrogate model is the slope creep deformation time series.

[0108] The following is a detailed explanation of steps S1-S7:

[0109] Step S1 includes:

[0110] Step S1-1:

[0111] Based on engineering geology, hydrogeology, and topographic survey data, a finite difference model computational grid for the reservoir bank slope was constructed. This grid includes geological structures such as bedrock, weathered rock mass, slip zone, toppling fault zone, and landslide body. The bottom boundary employs triaxial displacement constraints, and the lateral boundary employs normal displacement constraints. The computational grid is as follows: Figure 3 As shown;

[0112] Step S1-2:

[0113] The reservoir bank slope began to slowly creep due to the impact of the hydropower station's water storage at the end of May 2019. The on-site GNSS emergency monitoring system was completed and put into use on July 7, 2019.

[0114] For the missing slope deformation data before July 7, 2019, InSAR remote sensing monitoring was used to trace back the data, and a linear correction model of the InSAR observation results was established based on the deformation data monitored by GNSS in the field.

[0115] The corrected InSAR observation results (from December 10, 2018 to July 6, 2019) and the deformation data monitored by GNSS (from July 7, 2019 to September 15, 2019) were stitched together to obtain the creep deformation curve of the reservoir bank slope GNSS monitoring points before and after the slope sliding (from December 10, 2018 to September 15, 2019).

[0116] The creep deformation curves of the GNSS monitoring points before and after the slope sliding are shown in the figure. Figure 4As shown in the figure, G01-G21 represent 21 GNSS monitoring points; due to the influence of various factors such as slope geometry, geological structure, and monitoring point settings, the deformation data obtained from different monitoring points vary.

[0117] Steps S1-3:

[0118] Based on the engineering geological, hydrogeological, and topographical exploration data in step S1-1 and the creep deformation curves of the GNSS monitoring points obtained in step S1-2 before and after the slope sliding, the deformation mechanism of the reservoir bank slope can be analyzed and determined.

[0119] After undergoing multiple historical evolutions, the slope was in a critical stability state. After the reservoir was filled, the rising water level reduced the mechanical properties of the accumulation at the front edge of the slope, breaking the critical stability of the slope under natural conditions. Under the action of gravity, the existing sliding surface of the slope underwent overall deformation and adjustment, causing tensile cracks at the rear and side edges of the slope, and forming multiple tensile cracks and shear joints on the slope surface, which led to slope instability. However, after experiencing continuous large deformation, the slope gradually stabilized again under the condition of no external force constraint. From a microscopic perspective, this can be explained as follows: after the slope triggered sliding, the sliding zone was destroyed and the friction strength decreased. In the process of continuous sliding and self-adjustment, the sliding zone gradually compressed and compacted, and the friction strength increased, even exceeding the initial friction strength at the time of sliding, thus leading to the self-stabilization of the slope.

[0120] Steps S1-4:

[0121] Based on the deformation mechanism of the reservoir bank slope determined in steps S1-3, an elastic-viscoplastic creep constitutive model considering the deformation mechanism of the reservoir bank slope is established within the framework of internal variable thermodynamics.

[0122] The complexity of the elastic-viscoplastic creep constitutive model is manifested in the following aspects:

[0123] On the one hand, the evolution equation f(ε) of the friction coefficient f() in the elastic-viscoplastic creep constitutive model is controlled by the static parameters velo_a, velo_b, velo_c, velo_d, and velo_e and is expressed piecewise, as shown in the following equation:

[0124]

[0125] It is evident that f(ε) is related to the shear strain coefficient ε, rather than to time t. The evolution equation f(ε) needs to be transformed into f(t), thus involving complex dimensional transformations in the creep calculation process.

[0126] On the other hand, the Mohr-Coulomb (MC) strength criterion is generally not applicable to the elastic-viscoplastic creep constitutive model of reservoir bank slopes. Instead, the Drucker-Prager (DP) strength criterion, which considers the plastic flow and shear failure characteristics of rocks under three-dimensional stress, is more suitable. Since the friction coefficient f is a parameter of the MC strength criterion, it is necessary to convert f(t) into parameters a(t) and R(t) applicable to the DP strength criterion. These a(t) and R(t) are intermediate variables in the elastic-viscoplastic creep constitutive model, involving a highly nonlinear calculation process and belonging to intermediate implicit features.

[0127] Step S2 includes:

[0128] Step S2-1:

[0129] Based on the elastic-viscoplastic creep constitutive model considering the creep deformation mechanism established in steps S1-4, Latin hypercube sampling is performed on the original constitutive parameters of the model within a reasonable range to obtain N. S =10000 sets of first constitutive parameters;

[0130] The first constitutive parameter of each group can be expressed as: Where i represents the parameter group number, and there are a total of N. S Group parameters;

[0131] This represents a static parameter array in each group of first constitutive parameters that implicitly contains complex physical mechanisms and cannot be directly represented by time series, such as the control parameters [velo_a, velo_b, velo_c, velo_d, velo_e] of the evolution equation f(ε);

[0132] This represents the array of equation parameters in each group of first constitutive parameters that can be directly represented by time series, such as the control parameters [A] in the water-induced degradation effect equation. f B f D c A E B E ];

[0133] θ i This represents the time-invariant parameter in the first constitutive parameter group, such as the flow coefficient k2vis;

[0134] Step S2-2:

[0135] Each set of first constitutive parameters generated in step S2-1 The command flow is integrated into the finite difference calculation software, where equation parameters need to be... Convert the equation expression to time series format in advance, for example. For [A] f B f D c A E B E ]hour, The conversion expression is as follows:

[0136]

[0137]

[0138]

[0139] in, Here are the equations for the water-induced degradation effect, representing the friction coefficient, cohesion, and elastic modulus, respectively, where t represents the cumulative number of days since the water was impounded (December 10, 2018), and A... f B f To control the equations for the water-induced degradation effect on friction coefficient and cohesion, D c To control the parameters affecting the equation for water-induced degradation of cohesion, A E B E To control the parameters affecting the equation for the water-induced degradation effect of elastic modulus;

[0140] Based on the above transformation expression, we can obtain... Where j represents the time series number, with a total of N. X = 3 time series, T = 280 indicates the length of the time series (from December 10, 2018 to September 15, 2019, a total of 280 days);

[0141] Command flow also needs to consider external boundary conditions. Where k represents the number of the external boundary condition, and there are N such conditions. H There are several external boundary conditions, where T represents the length of the time series; in this example, only the change in water level is considered as an external boundary condition, therefore N H =1, T=280 (from December 10, 2018 to September 15, 2019, a total of 280 days);

[0142] Step S2-3:

[0143] Calculate the first constitutive parameter for each group based on the command stream constructed in step S2-2. The corresponding creep deformation response of the reservoir bank slope;

[0144] For each set of first constitutive parameters Corresponding creep deformation response of reservoir bank slope:

[0145] Record the creep deformation time series of GNSS monitoring points Where u represents the serial number of the GNSS monitoring point, with a total of N. Y = 21 GNSS monitoring points, T=280 indicates the length of the time series (from December 10, 2018 to September 15, 2019, a total of 280 days);

[0146] Records and Related time series with intermediate hidden features that can reflect complex physical mechanisms In this example This refers to the parameter a(t) that satisfies the DP intensity criterion described in steps S1-4 for the 12 monitoring points. This represents the parameter R(t) satisfying the DP intensity criterion described in steps S1-4 for 24 monitoring points, where v represents the serial number of the monitoring point, for a total of N. Z = 24 monitoring points, T = 280 indicates the length of the time series (from December 10, 2018 to September 15, 2019, a total of 280 days);

[0147] Step S2-4:

[0148] The first constitutive parameters obtained in step S2-1 are... and the results obtained in step S2-2 and the results obtained in step S2-3 The combination yields a numerical sample containing the first constitutive parameter, boundary conditions, intermediate implicit features, and creep deformation time series.

[0149] Step S3 includes:

[0150] Step S3-1:

[0151] Step S2-1 As the input feature vector of the Transformer physics mechanism enhancement module;

[0152] The above input feature vector is processed using a position encoder. Assign location information;

[0153] For example, the position encoder is calculated using the following formula:

[0154] PE(pos,2n)=sin(pos / 10000 2n / dmodel )

[0155] PE(pos,2n+1)=cos(pos / 10000 2n / dmodel )

[0156] Where pos represents the input feature vector The position of a certain element in the feature vector, where n represents the dimension index and d model It is the dimension of the embedded vector;

[0157] Step S3-2:

[0158] The input feature vector after position encoding in step S3-1 The correlation between different locations is calculated using multiple parallel self-attention layers;

[0159] For example, the computation of each self-attention layer can be represented as:

[0160]

[0161] Among them, Q m K m V m These are the query, key, and value matrices of the m-th self-attention layer, respectively, and d. k It is the dimension of the key. V represents the transpose of the key matrix, softmax() represents the softmax function, and V m The input feature vector after position encoding Obtained through linear transformation;

[0162] Step S3-3:

[0163] The outputs of multiple parallel self-attention layers in step S3-2 are spliced ​​together to obtain a multi-head self-attention layer, so as to simultaneously focus on information from different parts of the sequence.

[0164] For example, the multi-head self-attention layer can be represented as:

[0165] MultiHead(Q,K,V)=Concat(head1,…,head m ,…,head h W O

[0166] Where Q, K, and V are the complete query, key, and value matrix, respectively, and Concat() is a concatenation operation that merges all header outputs; head m W represents the output of the m-th self-attention layer, and there are a total of h self-attention layers; O It is the weight matrix for the output linear transformation;

[0167] Step S3-4:

[0168] After the multi-head self-attention layer in step S3-3, a feedforward neural network (FNN) is constructed, which can be represented as:

[0169] FNN(x) = max(0, xW1+b1)W2+b2

[0170] Where x represents the activation from the self-attention mechanism or the output of the previous encoder layer; max(0,z) represents the ReLU activation function, which applies a non-linear activation to each element z; W1 and W2 are the weight matrices in the feedforward neural network; b1 and b2 are the bias terms in the feedforward neural network.

[0171] Step S3-5:

[0172] The data processed by the feedforward neural network in steps S3-4 is passed through the last linear transformation layer to map the encoder output to the intermediate latent feature time series. The expected dimension;

[0173] For example, the last linear transformation layer can be represented as:

[0174]

[0175] in, The predicted sequence is the intermediate implicit feature time series that can reflect complex physical mechanisms, as described in steps S2-3; H encoder It is the output of the Transformer physics enhancement module, that is, the input feature vector after processing by self-attention mechanism and position encoding. High-level abstract representation; W latent From the output of the Transformer physics enhancement module to the intermediate latent feature time series Dimensional mapping weights; b latent It is the bias term in the mapping process.

[0176] The diagram of the Transformer physics enhancement module in this example is shown below. Figure 5 As shown.

[0177] Step S4 includes:

[0178] Step S4-1:

[0179] Create a multi-layer stacked GRU (Gate Recurrent Unit) network and apply a Dropout operation after each GRU layer to prevent overfitting;

[0180] For example, each GRU unit updates its state using the following formula:

[0181] R t =σ(W r ·[H t-1 ,X t ]+U r ·X t )

[0182] Z t =σ(W z ·[H t-1 ,X t ]+U z ·X t )

[0183]

[0184]

[0185] Among them, R t Z t These represent the reset gate and update gate at the current time step t, respectively. H t These represent the candidate hidden state and the final hidden state at the current time step t, respectively; σ represents the sigmoid function, responsible for activating the reset gate and update gate; tanh is the hyperbolic tangent function, used to activate the candidate hidden state; ⊙ represents the Hadamard product; X t H represents the input feature vector at the current time step t; t-1 W represents the final hidden state at the previous time step t-1; r W z W h These represent the reset gate, update gate, and candidate hidden state corresponding to H at the current time step t, respectively. t-1 The weight matrix; U r U z U h These represent the reset gate, update gate, and candidate hidden state corresponding to X at the current time step t, respectively. t The weight matrix;

[0186] Step S4-2:

[0187] The original features (in step S2-2) ) and predict intermediate latent features (obtained in steps S3-5) Feature merging is performed and used as input to the multi-layer stacked GRU network in step S4-1;

[0188] Step S4-3:

[0189] Finally, a fully connected layer is added to the multi-layer stacked GRU network constructed in step S4-1 to fuse the time-independent steady parameter θ described in step S2-1. i The final output of the multi-layer stacked GRU network, used for dimensional transformation, is the GNSS monitoring point creep deformation time series described in step S2-2. The predicted sequence;

[0190] The schematic diagram of the multi-layer stacked GRU network in this example is as follows: Figure 6 As shown.

[0191] Step S5 includes:

[0192] Step S5-1:

[0193] Construct the output of the linear transformation layer described in steps S3-5 (predicting intermediate hidden feature time series) loss function

[0194] The loss function of the output of the linear transformation layer It can be represented as:

[0195]

[0196] Where T = 280 represents the length of the time series (from December 10, 2018 to September 15, 2019, a total of 280 days); v represents the sequence number of the intermediate hidden feature monitoring points, totaling N. Z =24 intermediate hidden feature monitoring points; This represents the predicted feature value corresponding to the i-th numerical sample obtained at time step t for the v-th intermediate latent feature monitoring point. This represents the actual value of the feature corresponding to the i-th numerical sample obtained at time step t for the v-th intermediate hidden feature monitoring point;

[0197] Step S5-2:

[0198] The final output of the multi-layer stacked GRU network described in step S4-3 (time series of creep deformation at GNSS monitoring points) The loss function of the predicted sequence is used, and a time-dependent weight adjustment mechanism is introduced to make the model pay more attention to the accuracy of the predicted sequence at a specific stage.

[0199] The loss function that introduces a time-dependent weight adjustment mechanism It can be represented as:

[0200]

[0201] Where T = 280 represents the length of the time series (from December 10, 2018 to September 15, 2019, a total of 280 days); ω(t) represents the weighting function associated with time step t, used to guide the slope creep deformation surrogate model to focus on the accuracy of predictions for key deformation stages; u represents the sequence number of the GNSS monitoring points, totaling N. Y =21 GNSS monitoring points; Here, the Huber loss function is used to control sensitivity to large errors. This represents the deformation prediction value obtained by the u-th GNSS monitoring point at time step t, corresponding to the i-th numerical sample. This represents the actual deformation value corresponding to the i-th numerical sample obtained at time step t from the u-th GNSS monitoring point;

[0202] The weighting function ω(t) associated with time step t is shown in the following equation:

[0203]

[0204] Where t is the time step, T represents the length of the time series, a and b are weight coefficients, and γ is the non-linear growth rate of the weights;

[0205] The Huber loss function As shown in the following formula:

[0206]

[0207] Where δ is the threshold in the Huber loss function;

[0208] Step S5-3:

[0209] Using step S5-1 and in step S5-2 Constructing the combined loss function The expression for balancing the intermediate implicit features of the slope creep surrogate model and the creep deformation at GNSS monitoring points is as follows:

[0210]

[0211] Here, α is a hyperparameter between 0 and 1, used to adjust the proportion of the above two losses in the total loss. If α is closer to 1, the constructed slope creep deformation surrogate model will pay more attention to the accuracy of the GNSS monitoring point creep deformation time series prediction; if α is closer to 0, the constructed slope creep deformation surrogate model will pay more attention to the accuracy of the intermediate hidden feature time series prediction.

[0212] Step S6 includes:

[0213] Based on the combined loss function constructed in step S5-3 and the set of numerical samples obtained in step S2-4 The hyperparameters of the slope creep deformation surrogate model (Transformer-multilayer stacked GRU network) composed of steps S3 and S4 were optimized using a Bayesian optimizer.

[0214] The final optimization results of the hyperparameters in this example are as follows:

[0215] The Transformer physics enhancement module has the following feature dimensions: d_model = 128, nhead = 16, and num_encoder_layers = 6.

[0216] The number of layers in the multi-layer stacked GRU network is num_GRU=4, the number of hidden units per layer is hidden_sizes=[128,128,128,128], and the dropout rate after each layer is dropout_rates=[0.3,0.3,0.2,0.2].

[0217] The weighting coefficients of the weighting function ω(t) are a = 1 and b = 3, the non-linearity of weight growth is γ = 3, the threshold in the Huber loss function is δ = 0.4, and the combined loss function... The hyperparameter α = 0.6;

[0218] The batch size for training the model is 96, and the number of iterations is 20.

[0219] Step S7 includes:

[0220] Based on the combined loss function constructed in step S5-3 and the set of numerical samples obtained in step S2-4 The optimal hyperparameters obtained in step S6 are used to train the surrogate model (Transformer-multilayer stacked GRU network) jointly composed of steps S3 and S4, thereby obtaining a surrogate model of slope creep deformation between slope constitutive parameters, boundary conditions and creep deformation time series.

[0221] The slope creep surrogate model in this example compares the predicted and actual values ​​of creep deformation time series at typical GNSS monitoring points. Figure 7 As shown in the figure, the predicted and actual values ​​of the creep deformation time series are basically consistent.

[0222] This application also provides a non-transient computer-readable storage medium storing a computer program that can be executed by a processor to implement the method for constructing a slope creep deformation proxy model as described in any of the previous embodiments.

[0223] This application also provides a device for constructing a slope creep deformation surrogate model, such as... Figure 8 As shown, it includes a memory 801 and a processor 802. The memory 801 stores a computer program. When the computer program is read and executed by the processor 802, it can implement the method for constructing a slope creep deformation proxy model as described in any of the previous embodiments.

[0224] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term "computer storage medium" includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

Claims

1. A method for constructing a slope creep deformation surrogate model, the method comprising: A set of numerical samples is constructed for training the slope creep deformation surrogate model. Each numerical sample in the set contains a first constitutive parameter, external boundary conditions causing slope creep deformation, a first intermediate latent feature, and a first slope creep deformation time series. The first constitutive parameter is determined by the original constitutive parameters of the creep constitutive model reflecting the slope creep deformation mechanism, including: a time-independent static first constitutive parameter, a time-dependent dynamic first constitutive parameter, and a steady parameter. The first slope creep deformation time series is determined by the slope creep deformation response obtained by the slope computation grid through the slope creep constitutive model. The first intermediate latent feature is a time-dependent feature generated during the nonlinear calculation process of the slope creep constitutive model. For each numerical sample, the following operations are performed: The static first constitutive parameters, which are not correlated with the time dimension of the first constitutive parameters, are input into the first learning network to obtain predicted intermediate latent features; the dynamic first constitutive parameters, which are correlated with the time dimension of the first constitutive parameters, the predicted intermediate latent features, and the steady-state parameters in the first constitutive parameters are input into the second learning network to obtain the predicted slope creep deformation time series; based on the numerical sample, the predicted intermediate latent features, and the predicted slope creep deformation time series, the parameters of the combined learning network constructed by the first and second learning networks are optimized. The combined learning network is trained based on the numerical sample set and the final optimized parameters of the combined learning network, and the trained combined learning network is used as the surrogate model for slope creep deformation.

2. The construction method according to claim 1, characterized in that, The method further includes: Before constructing the numerical sample set for training the slope creep deformation surrogate model, a slope computational grid and the slope creep constitutive model are constructed. Correspondingly, the numerical sample set used to train the slope creep deformation surrogate model includes: Based on the slope computational grid and the slope creep constitutive model, a set of numerical samples is constructed for training the slope creep deformation surrogate model; The construction of the slope creep constitutive model includes: Acquire full-process slope creep deformation data, which includes slope data before and after slope creep occurs; The slope creep constitutive model is determined based on the full-process slope creep deformation data.

3. The construction method according to claim 2, characterized in that, The acquisition of full-process slope creep deformation data includes: If the monitored slope creep deformation data is data after the slope creep occurred, backtrack the monitoring timeline to include slope data before the slope creep occurred. The monitored slope creep deformation data is stitched together with the backtracked slope data to obtain the full-process slope creep deformation data.

4. The construction method according to claim 1, characterized in that, The numerical sample set used to train the slope creep deformation surrogate model includes: By sampling the original constitutive parameters of the slope creep constitutive model, N is obtained. S The first constitutive parameter of the i-th group; the first constitutive parameter of the i-th group is expressed as Indicates the static first constitutive parameter; Represents the dynamic first constitutive parameter; θ i Indicates a constant parameter; For the first constitutive parameter Φ of the i-th group i , will the Convert to time series j represents the time series number, with a total of N. X A time series, where T represents the length of the time series; based on the first constitutive parameter Φ of the i-th group. i The corresponding slope creep deformation response was used to determine the first slope creep deformation time series from the first monitoring point. and the first intermediate latent feature from the second monitoring point Where u represents the sequence number of the first monitoring point, and there are a total of N points. Y There are N first monitoring points, where v represents the sequence number of the second monitoring point, for a total of N. Z A second monitoring point; Will as well as The i-th numerical sample is obtained by combining the results. in, This represents the external boundary conditions that cause slope creep deformation, where k represents the number of the external boundary conditions, and there are a total of N. H External boundary conditions.

5. The construction method according to claim 1, characterized in that, The step of inputting the time-independent static first constitutive parameters from the first constitutive parameters into the first learning network to obtain predicted intermediate latent features includes: The static first constitutive parameters are input into the constructed Transformer physics mechanism enhancement model with a position encoder to obtain the predicted intermediate latent features.

6. The construction method according to claim 5, characterized in that, The method for constructing the Transformer physics enhancement model with a position encoder includes: The input data of the Transformer physics enhancement model is position-encoded using a position encoder; The position-encoded input data is passed through multiple parallel self-attention layers; The output data of the multiple parallel self-attention layers are concatenated through a multi-head self-attention layer; The concatenated data is passed sequentially through a feedforward neural network layer and a linear transformation layer to obtain the output data of the Transformer physics mechanism enhancement model.

7. The construction method according to claim 1, characterized in that, The step of inputting the dynamic first constitutive parameter associated with the time dimension of the first constitutive parameter, the predicted intermediate latent features, and the steady parameters in the first constitutive parameter into the second learning network to obtain the predicted slope creep deformation time series includes: The predicted intermediate latent features and the dynamic original features affecting slope creep deformation determined by the dynamic first constitutive parameter are input into a multi-layer stacked GRU network; the dynamic original features affecting slope creep deformation determined by the dynamic first constitutive parameter refer to the time series features obtained by substituting the dynamic first constitutive parameter associated with the time dimension into the mathematical expression. The output data of the multi-layer stacked GRU network and the constant parameters are input into the fully connected layer. The output data of the fully connected layer is used as the predicted slope creep deformation time series.

8. The construction method according to claim 7, characterized in that, Each GRU layer in the multi-layer stacked GRU network is followed by a Dropout operation.

9. The construction method according to claim 1, characterized in that, The parameter optimization of the combined learning network constructed from the first and second learning networks based on the numerical samples, the predicted intermediate latent features, and the predicted slope creep deformation time series includes: The prediction loss of the intermediate latent features is determined based on the first intermediate latent features of the numerical samples and the predicted intermediate latent features; The prediction loss of the slope creep deformation time series is determined based on the first slope creep deformation time series of the numerical sample and the predicted slope creep deformation time series. The combined loss is determined by combining the predicted loss of the intermediate implicit features and the predicted loss of the slope creep deformation time series. The parameters of the ensemble learning network are optimized based on the numerical samples and the ensemble loss.

10. The construction method according to claim 9, characterized in that, The prediction loss of the slope creep deformation time series, determined based on the first slope creep deformation time series of the numerical samples and the predicted slope creep deformation time series, is achieved by the following formula: Where T represents the length of the time series; ω(t) represents the weight function associated with time step t; and u represents the sequence number of the first monitoring point, for a total of N. Y The first monitoring point; Huber loss function; This represents the predicted slope creep deformation time series obtained at time step t for the u-th first monitoring point, corresponding to the i-th numerical sample. This represents the first slope creep deformation time series obtained at time step t for the u-th first monitoring point, corresponding to the i-th numerical sample. Let represent the prediction loss function for the slope creep deformation time series corresponding to the i-th numerical sample.

11. The construction method according to claim 10, characterized in that, The weight function ω(t) associated with time step t is implemented by the following formula: Where a and b are weighting coefficients, and γ is the non-linear growth rate of the weights.

12. The construction method according to claim 9, characterized in that, The combined loss, determined by combining the predicted loss of the intermediate implicit features and the predicted loss of the slope creep deformation time series, is achieved through the following formula: in, This represents the prediction loss for the slope creep deformation time series corresponding to the i-th numerical sample. Y represents the predicted slope creep deformation time series corresponding to the i-th numerical sample. i This represents the time series of creep deformation of the first slope corresponding to the i-th numerical sample. This represents the prediction loss of the intermediate latent feature corresponding to the i-th numerical sample. Z represents the predicted intermediate latent feature corresponding to the i-th numerical sample. i This represents the first intermediate latent feature corresponding to the i-th numerical sample. Let represent the combined loss function corresponding to the i-th numerical sample, and let α represent the hyperparameter, where 0 ≤ α ≤ 1.

13. A non-transient computer-readable storage medium storing a computer program that can be executed by a processor to implement a method for constructing a slope creep deformation proxy model as described in any one of claims 1 to 12.

14. An apparatus for constructing a slope creep deformation proxy model, comprising a memory and a processor, wherein the memory stores a computer program, and the computer program, when read and executed by the processor, is capable of implementing the method for constructing a slope creep deformation proxy model as described in any one of claims 1 to 12.

Images