Porous physical information neural network landslide displacement field reconstruction and hole distribution method and system

Abstract

This invention discloses a method and system for reconstructing and deploying landslide displacement fields using a porous physical information neural network. The method includes: acquiring observation data and spatial coordinates of multiple monitoring wells deployed along the main sliding direction of the landslide; independently normalizing the displacement observation data for each monitoring well, and performing global shared parameter normalization on the observation depth, observation time, and spatial coordinates; constructing a porous joint physical information neural network model; training the model using a phased training strategy, reasoning point-by-point on the prediction grid, and reconstructing the physical displacement according to a displacement scale function constructed along the main sliding direction to reconstruct a continuous displacement field; generating a synthetic twin dataset based on the trained creep parameter network, performing leave-one-out cross-validation on multiple candidate deployment configurations, determining the minimum number of monitoring wells and the maximum allowable spacing based on the average determination coefficient of the internal wells, and outputting the deployment scheme. This invention achieves continuous displacement field reconstruction across monitoring wells and provides a quantitative basis for monitoring well deployment density.

Description

Technical Field

[0001] This invention relates to the fields of landslide geological hazard monitoring and geological engineering technology, and in particular to a porous physical information neural network method and system for landslide displacement field reconstruction and pore layout. Background Technology

[0002] Landslides are one of the most common geological hazards in mountainous highways, railways, and urban construction. Effective early warning relies on continuous monitoring of slope deformation. While surface monitoring methods such as synthetic aperture radar interferometry, GPS, and lidar can obtain the distribution of surface displacement in a planar manner, they cannot detect the internal structure of the slope. The depth and geometry of the slip surface that control the instability mechanism are not visible in surface observations. Borehole inclinometers fill this observation gap by measuring the cumulative lateral displacement at different depths and are the main means of identifying the location of the slip surface and characterizing the deep creep process.

[0003] However, the high cost of drilling and equipment installation objectively limits the spatial coverage density of boreholes. In engineering practice, a landslide on a scale of several hundred meters usually requires only three to five inclinometer boreholes, with a spacing of more than one hundred meters between adjacent boreholes. This has led to a long-standing engineering question that lacks a quantitative answer: how many boreholes are needed, and at what spacing, to reliably reconstruct a continuous underground displacement field.

[0004] To address the spatial analysis of underground displacement fields in landslides, existing technologies primarily employ the following methods: (1) Single-hole independent modeling method: For the depth-time displacement data of each inclinometer hole, a time series model or physical constraint model is established separately, and the displacement evolution and slip surface depth at the hole location are given for each hole. In this method, the hole models are isolated from each other, and the displacement estimation ability is not set for the hole location. It cannot output a continuous global displacement field, nor can it answer whether the inter-hole interpolation is reliable. (2) Geostatistical interpolation method: Spatial interpolation of displacement between holes is performed using methods such as Kriging interpolation. This method is oriented towards static scalar fields and cannot handle dynamic displacement fields coupled with depth, time and space. It does not contain any creep constitutive constraints, and the interpolation results lack physical constraints and have poor extrapolation ability when the hole spacing is large. (3) Numerical calculation method: The slope model is established by finite element method or finite difference method, and the complete displacement field is modeled in the forward direction. This method requires complete constitutive parameters and boundary conditions, but these data are usually not available for in-service monitored slopes, and parameter inversion itself depends on sufficiently dense observations; However, none of the existing methods mentioned above treat the borehole density itself as a diagnosable and optimizable object. Therefore, the existing technologies have the following technical problems that urgently need to be solved: First, it lacks the ability to reconstruct the continuous displacement field across boreholes, and the location of the unmarked borehole becomes a blind spot for monitoring, making it impossible to continuously depict the spatial extension shape of the slip surface. Secondly, there is a lack of a unified expression for the dynamic process of depth-time-space coupling, and spatial interpolation and temporal prediction are separated into two unsupported links; Third, the reliance on complete constitutive parameters and boundary conditions does not match the data conditions of in-service monitored slopes, making it difficult to provide real-time results during the monitoring period.

[0005] Fourth, the displacement magnitudes of different parts of the same landslide can differ by two orders of magnitude. If the data from each borehole are directly fed into a unified loss function for joint training, the loss will be dominated by the large displacement boreholes, and the displacement patterns of the small displacement boreholes will be submerged, making it impossible to guarantee the reconstruction accuracy of weak signal parts.

[0006] Fifth, there is a lack of quantitative criteria for determining the density of boreholes. Too few boreholes create monitoring blind spots, while too many boreholes lead to wasted exploration investment. In engineering, borehole placement can only be based on experience, and there are no quantitative guidelines on whether to increase the density in geologically abnormal sections or on which side the boundary boreholes should be placed first.

[0007] Sixth, it failed to distinguish the different requirements of time-dimensional prediction and spatial-dimensional interpolation for the amount of observation information. In the early stages of monitoring when the number of wells was insufficient, it was impossible to obtain reliable spatial interpolation results or to use existing single-well time-series data to achieve early warning. The time-series allocation of monitoring resources lacked a basis. Summary of the Invention

[0008] In view of the shortcomings of the prior art, the present invention provides a method and system for reconstructing and locating the landslide displacement field using a porous physical information neural network, in order to solve the problems in the prior art where the location of the borehole cannot be continuously reconstructed, the inter-bore interpolation lacks physical constraints, the weak signal is submerged due to the difference in the magnitude of the porous displacement, the borehole layout density lacks quantitative criteria, and the time dimension prediction and spatial dimension interpolation are not distinguished.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for reconstructing and pore-layouting landslide displacement fields using a porous physical information neural network, comprising the following steps: Observational data of multiple monitoring holes deployed along the main sliding direction of the landslide are obtained, and the spatial coordinates of each monitoring hole along the main sliding direction are obtained; The observation data includes cumulative lateral displacement observation data, the observation depth corresponding to the cumulative lateral displacement observation data, and the observation time; The displacement observation data is independently normalized according to the monitoring hole to obtain the normalized displacement, and the observation depth, the observation time and the spatial coordinates are normalized with global shared parameters to obtain the normalized variables; The normalized variables include normalization depth, normalization time, and normalized spatial coordinates; Construct a multi-pore joint physical information neural network model and establish a composite loss function; The porous joint physical information neural network model includes a displacement network and a creep parameter network; Based on the normalized variables and normalized shifts, a phased training strategy is adopted to train the physical information neural network model based on the composite loss function, thereby obtaining the trained multi-pore joint physical information neural network model. Multiple input coordinates are generated within the range of the normalized variable. Each input coordinate is input into the trained displacement network to obtain the normalized predicted displacement at each input coordinate. The normalized predicted displacement is then denormalized to obtain the actual displacement at each input coordinate, thereby forming a three-dimensional continuous displacement field. The preset normalized depth and normalized spatial coordinates are input into the trained creep parameter network to obtain the corresponding creep parameters and generate a synthetic twin dataset superimposed with spatial heterogeneity features. Obtain geological data of the landslide, and set multiple candidate deployment configurations based on the geological data; Based on the synthetic twin dataset, leave-one-out cross-validation is performed on each of the candidate deployment configurations to obtain the average determination coefficient of each candidate deployment configuration. The candidate deployment configurations whose average determination coefficients meet the preset conditions are determined as deployment schemes and output.

[0010] In an optional implementation, the method further includes: The strain field is obtained by automatically differentiating the continuous displacement field along the depth direction, and the position of the sliding surface is determined based on the concentrated zone of the strain field along the depth direction. Using the trained multi-pore joint physical information neural network model, extrapolation prediction is performed along the time direction to obtain the extrapolated displacement prediction value.

[0011] In an optional implementation, the displacement observation data is independently normalized for each monitoring hole to obtain normalized displacement, and the observation depth, observation time, and spatial coordinates are normalized using globally shared parameters to obtain normalized variables. This specifically includes the following steps: Based on the minimum and maximum values ​​of the displacement observation data of each monitoring hole, the displacement observation data of the monitoring hole is linearly mapped to a preset interval to obtain the normalized displacement of the monitoring hole. The monitoring hole is the borehole where the borehole surveying instrument is located; The observation depth is normalized by using the minimum and maximum values ​​of the observation depths among all monitoring wells as normalization parameters shared across the entire region, thus obtaining the normalized depth. The observation time is normalized by using the minimum and maximum values ​​of the observation time of all monitoring wells as normalization parameters shared across the entire region, thus obtaining the normalized time. Using the minimum and maximum values ​​of the spatial coordinates of all monitoring wells as the normalization parameters shared across the entire domain, the spatial coordinates are normalized to obtain normalized spatial coordinates.

[0012] In an optional implementation, a multi-pore joint physical information neural network model is constructed, and a composite loss function is established, specifically including the following steps: A multi-pore joint physical information neural network model is constructed based on the physical information neural network. The displacement network takes the normalized variable as input and outputs a normalized predicted displacement. The creep parameter network takes the normalized depth and the normalized spatial coordinates as input and outputs creep parameters that vary with depth and spatial coordinates. The creep parameters are Burgers viscoelastic creep model parameters, which include initial elastic displacement parameters, steady-state creep rate parameters, transient creep amplitude parameters, and transient decay rate parameters. A data fidelity loss term is established to constrain the normalized predicted displacement to approximate the normalized displacement by penalizing the deviation between the normalized predicted displacement and the normalized displacement. A creep constitutive loss term is established to constrain the temporal evolution of the displacement to conform to the Burgers viscoelastic creep model by penalizing the deviation between the normalized predicted displacement derivative with respect to time and the theoretical creep rate determined by the creep parameters. A strain smoothing loss term is established to constrain strain smoothing along the depth direction by penalizing the mean square value of the third derivative of the normalized predicted displacement with respect to the normalized depth at the placement point. A creep constitutive loss term is established. By restoring the normalized predicted displacement to physical displacement according to the displacement scale and differentiating it with respect to physical time, the physical displacement rate is obtained. The deviation between the physical displacement rate and the theoretical creep rate determined by the creep parameters is penalized to constrain the time evolution of the displacement to conform to the Burgers viscoelastic creep model. A displacement monotonicity loss term is established, which is used to constrain the displacement from the front edge of the landslide to the rear edge of the landslide by penalizing the square mean of the first-order partial derivative of the normalized predicted displacement with respect to the normalized spatial coordinates after one-sided rectification. The data fidelity loss term, the creep constitutive loss term, the strain smoothing loss term, the creep parameter spatial smoothing loss term, and the displacement monotonicity loss term constitute a composite loss function.

[0013] In an optional implementation, using the normalized variables as input, a phased training strategy is employed to train the physical information neural network model based on the composite loss function, resulting in a trained multi-pore joint physical information neural network model. This specifically includes the following steps: With minimizing the data fidelity loss term as the optimization objective, the porous joint physical information neural network model is pre-trained to obtain the pre-trained porous joint physical information neural network model. The creep constitutive loss term, the strain smoothing loss term, the creep parameter space smoothing loss term, and the displacement monotonicity loss term are modulated by using course weights that increase from zero to one with each training round. The modulated loss terms are then superimposed with the data fidelity loss term to obtain the modulated composite loss function. With minimizing the modulated composite loss function as the optimization objective, the pre-trained multi-pore joint physical information neural network model is fine-tuned to obtain the fine-tuned multi-pore joint physical information neural network model. A quasi-Newton optimizer is used to refine the fine-tuned porous joint physical information neural network model with the goal of minimizing the composite loss function, resulting in a trained porous joint physical information neural network model.

[0014] In an optional implementation, a preset normalized depth and normalized spatial coordinates are input into the trained creep parameter network to obtain the corresponding creep parameters and generate a synthetic twin dataset superimposed with spatial heterogeneity features. This specifically includes the following steps: The preset normalized depth and normalized spatial coordinates are input into the trained creep parameter network to obtain the corresponding creep parameters, thereby forming a two-dimensional creep parameter field. Based on the two-dimensional creep parameter field, a reference trend for each creep parameter along the main sliding direction is set; The baseline trend decreases from the leading edge of the landslide to the trailing edge. The baseline trend corresponding to each creep parameter is multiplied by the heterogeneity multiplier to obtain the superimposed creep parameters, so that the spatial heterogeneity features are superimposed on each creep parameter; The heterogeneous multipliers include sinusoidal wave terms and local anomaly terms; Based on the preset depth shape function, a synthetic displacement field is constructed according to the superimposed creep parameters; The synthetic twin dataset is obtained by sampling at preset monitoring hole positions, depths, and time points on the synthetic displacement field.

[0015] In an optional implementation, based on the synthetic twin dataset, leave-one-out cross-validation is performed on each of the candidate deployment configurations to obtain the average determination coefficient of each candidate deployment configuration. The candidate deployment configurations whose average determination coefficients satisfy preset conditions are determined as deployment schemes and output. Specifically, this includes the following steps: The synthetic displacement data is globally standardized using the mean and standard deviation of the synthetic displacement data of all monitoring holes in the synthetic twin dataset to obtain standard synthetic displacement data. For each monitoring hole in the candidate deployment configuration, the monitoring hole is sequentially designated as a reserved hole, and leave-one cross-validation is performed to obtain the determination coefficient of each monitoring hole in the candidate deployment configuration. The monitoring holes located at both ends in the candidate deployment configuration are determined as boundary holes, and the remaining monitoring holes are determined as internal holes. The average value of the determination coefficients corresponding to all the internal holes is taken as the average determination coefficient of the candidate deployment configuration. Leave-one-out cross-validation is performed on the current candidate deployment configuration in ascending order of the number of monitoring holes to obtain the average coefficient of determination of the internal holes of the candidate deployment configuration. Determine whether the average determination coefficient of the internal hole is less than a preset threshold. If the number is less than the required number of monitoring holes, leave-one-out cross-validation will be performed on the next candidate deployment configuration in ascending order of the number of monitoring holes. If it is not less than, then the current candidate deployment configuration is determined as the selected deployment configuration, and the number of monitoring holes corresponding to the current candidate deployment configuration and the average distance between adjacent monitoring holes are determined as the minimum number of monitoring holes and the maximum allowable distance, respectively. In the selected deployment configuration, the section where the monitoring hole with the lowest determination coefficient in the internal borehole is located is identified as a suspected abnormal section or a section that needs to be densified, and is verified in conjunction with geological data; The spacing between the monitoring holes in the reviewed sections is increased to less than half of the maximum allowable spacing, and the boundary holes on the side with smaller displacement magnitude are laid out first to form the layout scheme. The specific steps of leave-one-out cross-validation include: The multi-pore joint physical information neural network model is trained using the standard synthetic displacement data corresponding to the remaining monitoring holes (excluding the reserved holes) in the candidate layout configuration in the synthetic twin dataset. The normalized coordinates of the opening in the synthetic twin dataset are input into the trained multi-pore joint physical information neural network model to obtain the displacement prediction value corresponding to the opening. Based on the standard synthetic displacement data and displacement prediction value corresponding to the reserved hole, the determination coefficient of the reserved hole is calculated in the same standardized domain.

[0016] In an optional implementation, the strain field is obtained by automatically differentiating the continuous displacement field along the depth direction. The slip surface position is determined based on the concentrated zone of the strain field along the depth direction. Then, an extrapolation prediction is performed along the time direction using a trained multi-pore joint physical information neural network model to obtain the extrapolated displacement prediction value. Specifically, the following steps are included: The strain field is obtained by automatically differentiating the continuous displacement field along the depth direction. For each monitoring hole, the distribution curve of the strain field along the depth direction at each time point is calculated, and the distribution curve of the same monitoring hole at each time point is averaged along the time direction to obtain the time-averaged strain profile of the monitoring hole. Locate local maxima in the time-averaged strain profile and determine whether the strain value of the local maxima exceeds a preset strain threshold. If it exceeds the limit, the depth corresponding to the local maximum point is determined as the sliding surface position; If it does not exceed the limit, then the local maximum point is excluded; The displacement observation data of each monitoring well were divided into training and verification segments according to time sequence. The multi-well joint physical information neural network model was retrained using the displacement observation data corresponding to the training segments of all monitoring wells. Using the retrained multi-pore joint physical information neural network model, the displacement at each time point in the test section is predicted, and the extrapolated displacement prediction value at each time point in the test section is obtained. Calculate the extrapolation determination coefficient between each extrapolated displacement prediction value and the displacement observation data corresponding to the test section, and use the extrapolation determination coefficient to characterize the extrapolation accuracy.

[0017] Secondly, the present invention provides a porous physical information neural network system for landslide displacement field reconstruction and pore layout, comprising: The data acquisition module is used to acquire observation data from multiple monitoring holes deployed along the main sliding direction of the landslide, and to acquire the spatial coordinates of each monitoring hole along the main sliding direction. The observation data includes cumulative lateral displacement observation data, the observation depth corresponding to the cumulative lateral displacement observation data, and the observation time; The data processing module is used to independently normalize the displacement observation data according to the monitoring hole to obtain normalized displacement, and to perform global shared parameter normalization on the observation depth, the observation time and the spatial coordinates to obtain normalized variables; The normalized variables include normalization depth, normalization time, and normalized spatial coordinates; The model building module is used to construct a multi-pore joint physical information neural network model and establish a composite loss function; The porous joint physical information neural network model includes a displacement network and a creep parameter network; The model training module is used to train the physical information neural network model based on the composite loss function according to the normalized variables and normalized displacements, using a phased training strategy, to obtain the trained multi-pore joint physical information neural network model. The displacement field reconstruction module is used to generate multiple input coordinates within the range of the normalized variable, input each input coordinate into the trained displacement network to obtain the normalized predicted displacement at each input coordinate, and perform inverse normalization on the normalized predicted displacement to obtain the actual displacement at each input coordinate, so as to form a three-dimensional continuous displacement field. The dataset generation module is used to input the preset normalized depth and normalized spatial coordinates into the trained creep parameter network to obtain the corresponding creep parameters and generate a synthetic twin dataset superimposed with spatial heterogeneity features. The candidate deployment module is used to acquire geological data of the landslide and set multiple candidate deployment configurations based on the geological data. The deployment scheme module is used to perform leave-one-out cross-validation on each of the candidate deployment configurations based on the synthetic twin dataset, obtain the average determination coefficient of each candidate deployment configuration, and determine and output the candidate deployment configurations whose average determination coefficients meet the preset conditions as deployment schemes.

[0018] In an optional implementation, the system further includes: The slip surface identification module is used to automatically differentiate the continuous displacement field along the depth direction to obtain the strain field, and determine the slip surface position based on the concentrated zone of the strain field along the depth direction. The extrapolation prediction module is used to perform extrapolation prediction along the time direction using the trained multi-pore joint physical information neural network model to obtain the extrapolated displacement prediction value.

[0019] The beneficial effects of the embodiments provided by the present invention include: This invention expands the input of the displacement network from a two-dimensional depth-time model to a three-dimensional depth-time-space model, and constructs a multi-pore joint physical information neural network model including a displacement network and a creep parameter network. This enables a single model to be jointly trained on multiple monitoring holes and output displacement estimates in the region between monitoring holes, thereby realizing continuous displacement field reconstruction across monitoring holes and eliminating monitoring blind spots where no monitoring holes are located.

[0020] This invention constructs a composite loss function, enabling spatial interpolation and temporal prediction to be completed on the same continuous field. This overcomes the shortcomings of geostatistical interpolation methods, which cannot handle depth-time-space coupled dynamic processes and lack physical constraints. By outputting creep parameters that vary with depth and spatial coordinates through a creep parameter network, the displacement field can be reconstructed without prior complete constitutive parameters and boundary conditions. This adapts to the data conditions of in-service monitored slopes.

[0021] This invention avoids the dominant loss function of large displacement monitoring holes by independently normalizing displacement observation data according to the monitoring holes, and enables the displacement patterns of weak signal monitoring holes to be effectively learned, thereby improving the reconstruction accuracy of weak signal parts in multi-monitoring hole joint reconstruction.

[0022] This invention generates a synthetic twin dataset with superimposed spatial heterogeneity features based on trained creep parameters, and sets multiple candidate layout configurations based on geological data. Leave-one-out cross-validation is performed on each candidate layout configuration, and the layout scheme is determined by the average coefficient of determination meeting the preset conditions. This enables quantitative diagnosis and optimization of borehole layout density, avoiding monitoring blind spots or exploration investment waste caused by borehole layout based on experience. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0024] Figure 1 A schematic flowchart of the porous physical information neural network landslide displacement field reconstruction and pore placement method in the embodiments of this specification is shown; Figure 2 The diagram illustrates the architecture and training flowchart of the multi-pore joint physical information neural network model in the embodiments of this specification. Figure 3 A schematic diagram of multi-scale normalization processing in an embodiment of this specification is shown; wherein, Figure 3 (a) Comparison of the magnitudes of the cumulative lateral displacements of the three monitoring holes before normalization; Figure 3 (b) is a displacement distribution diagram of the three monitoring holes after independent normalization; Figure 4 This document illustrates a schematic diagram of the course weight scheduling and three-stage training strategy in an embodiment of this specification; wherein, Figure 4 (a) is a piecewise linear scheduling curve of course weights with training rounds; Figure 4 (b) is a diagram showing the number of rounds and learning rate configurations for the three-stage training; Figure 5 The following diagram shows the accuracy of the joint reconstruction of three monitoring holes in an embodiment of this specification; wherein, Figure 5 (a) is a diagram of the reconstruction determination coefficients for each monitoring well; Figure 5 (b) shows the reconstructed root mean square error diagram for each monitoring well; Figure 6 The following is a global displacement field distribution map of the synthetic twin dataset in an embodiment of this specification; wherein, Figure 6 (a) is a contour map of displacement in the depth-space plane at t=20d; Figure 6 (b) is a contour map of displacement in the depth-space plane at t=80d; Figure 7 This specification shows a graph illustrating the relationship between the number of monitoring holes and the average determination coefficient of internal holes in the embodiments of this specification. Figure 8 A graph showing the time-direction extrapolation prediction accuracy in the embodiments of this specification is provided. Figure 9 This document shows a bar chart illustrating the coefficient of determination for cross-validation with one hole left in each of the seven-hole configurations in an embodiment of this specification. Figure 10 The diagram shows the structural block diagram of the porous physical information neural network landslide displacement field reconstruction and pore layout system in the embodiments of this specification. Detailed Implementation

[0025] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without requiring some of these specific details. The following description of embodiments is merely intended to provide a better understanding of the invention by illustrating examples of the invention.

[0026] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appear in a specific order. However, it should be clearly understood that these operations may not be performed in the order they appear herein, or they may be performed in parallel. The operation numbers, such as S1, S2, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be performed sequentially or in parallel.

[0027] Example 1

[0028] like Figures 1-2 As shown, this embodiment provides a method for reconstructing and pore placement of landslide displacement field using a porous physical information neural network, including the following steps: S1. Obtain observation data from multiple monitoring holes deployed along the main sliding direction of the landslide, and obtain the spatial coordinates of each monitoring hole along the main sliding direction; The observation data includes cumulative lateral displacement observation data, the observation depth corresponding to the cumulative lateral displacement observation data, and the observation time. In some embodiments, the cumulative lateral displacement observation data is collected by a borehole inclinometer installed at the monitoring hole; Optionally, multiple monitoring holes are arranged at intervals along the main sliding direction of the landslide, and each monitoring hole is equipped with a borehole inclinometer; Wherein, the spatial coordinates are the horizontal distance between each monitoring hole along the main sliding direction; the observation depth and observation time are the depth and time when the borehole inclinometer collects the cumulative lateral displacement observation data.

[0029] S2. The displacement observation data is independently normalized according to the monitoring hole to obtain the normalized displacement, and the observation depth, the observation time and the spatial coordinates are normalized with global shared parameters to obtain the normalized variables; The normalization variables include normalization depth, normalization time, and normalization space coordinates; For example, step S2 specifically includes the following steps: S201. Based on the minimum and maximum values ​​of the displacement observation data of each monitoring hole, linearly map the displacement observation data of the monitoring hole to a preset interval to obtain the normalized displacement of the monitoring hole. Specifically, the normalized displacement of the monitoring hole is expressed as follows: ; In the formula, The normalized displacement of the k-th monitoring hole; This represents the displacement observation value of the k-th monitoring hole; This represents the minimum value of the cumulative lateral displacement observation data for the k-th monitoring well; This represents the maximum value of the cumulative lateral displacement observation data for the k-th monitoring well; The preset interval is [0,1]. S202. Using the minimum and maximum values ​​of the observation depths among all monitoring wells as normalization parameters shared across the entire region, the observation depths are normalized to obtain the normalized depth. Specifically, the expression for the normalization depth is: ; In the formula, Normalized depth; For observation depth; This is the minimum value within the observed depth; The maximum value within the observed depth; S203. Using the minimum and maximum values ​​of the observation times of all monitoring wells as the normalization parameters shared across the entire region, the observation times are normalized to obtain the normalized time. Specifically, the expression for normalized time is: ; In the formula, Normalized time; For observation time; This is the minimum value during the observation period; The maximum value during the observation period; S204. Using the minimum and maximum values ​​of the spatial coordinates of all monitoring holes as the normalization parameters shared across the entire domain, normalize the spatial coordinates to obtain normalized spatial coordinates. Specifically, the expression for normalized spatial coordinates is: ; In the formula, Normalized spatial coordinates; To monitor the spatial coordinates of the hole along the main sliding direction; The maximum value among all the spatial coordinates of the monitoring holes; In this embodiment, the observation depth, observation time, and spatial coordinates are normalized using globally shared parameters, so that the observation points of multiple monitoring points fall into the same continuous normalized coordinate system, thereby enabling the model to learn spatial correlations across monitoring holes. The displacement is normalized independently for each monitoring hole, so that the model learns the normalized displacement evolution pattern of each monitoring hole rather than the absolute magnitude, thereby avoiding the large displacement monitoring point dominating the loss function.

[0030] S3. Construct a multi-pore joint physical information neural network model based on physical information neural network, and establish a composite loss function; Among them, the porous joint physical information neural network model includes a displacement network and a creep parameter network; For example, step S3 specifically includes the following steps: S301. Construct a multi-pore joint physical information neural network model based on physical information neural network; Specifically, the displacement network takes the normalized variable as input and outputs a normalized predicted displacement; the creep parameter network takes the normalized depth and the normalized spatial coordinates as input and outputs creep parameters that vary with depth and spatial coordinates.

[0031] In some embodiments, the creep parameters are Burgers viscoelastic creep model parameters, including initial elastic displacement parameters, steady-state creep rate parameters, transient creep amplitude parameters, and transient decay rate parameters; wherein, the parameters in the creep parameters are all bivariate continuous functions of normalized depth and normalized spatial coordinates.

[0032] In this embodiment, the output layer of the creep parameter network applies a softplus activation function to the steady-state creep rate parameter, transient creep amplitude parameter, and transient decay rate parameter to ensure that the above three parameters are strictly positive; no positive value constraint is applied to the initial elastic displacement parameter to accommodate the positive and negative offset of the observation baseline.

[0033] In a preferred embodiment, the displacement network consists of 8 fully connected hidden layers, each containing 128 neurons, with an output dimension of 1; the creep parameter network consists of 4 fully connected hidden layers, each containing 64 neurons, with an output dimension of 4, corresponding sequentially to the initial elastic displacement parameter, steady-state creep rate parameter, transient creep amplitude parameter, and transient decay rate parameter.

[0034] In both the displacement network and the creep parameter network, each connected hidden layer uses the hyperbolic tangent function as the activation function, and its expression is: ; In the formula, a is the input to the activation function; The expression for the softplus function is: ; Burgers' viscoelastic creep model expresses the creep displacement that decays over time as: ; In the formula, u is the creep displacement; A is the initial elastic displacement parameter in mm; B is the steady-state creep rate parameter in mm / d; C is the transient creep amplitude parameter in mm; D is the transient decay rate parameter in 1 / d; and t is time in d. In this embodiment, the displacement network is responsible for representing the continuous displacement field. Its input is three-dimensional normalized coordinates, so it can be differentiable in the three directions of depth, time and space. The derivative fields of each order can be obtained through automatic differentiation. The creep parameter network is responsible for the spatial inversion of constitutive parameters. Its input does not contain time. Therefore, the inverted creep parameters only change with depth and spatial position and do not drift with time, which is consistent with the physical properties of creep parameters.

[0035] Compared to a single-hole model that only uses depth and time as input, the displacement network in this embodiment adds normalized spatial coordinates to the input, and the network capacity is correspondingly increased from 6 layers × 64 neurons to 8 layers × 128 neurons to accommodate the increased spatial dimension and multi-scale displacement patterns across monitoring holes.

[0036] S302. Establish a data fidelity loss term, and constrain the normalized predicted displacement to approximate the normalized displacement by penalizing the deviation between the normalized predicted displacement and the normalized displacement. Specifically, the data fidelity loss item is: ; In the formula, For data fidelity loss; N is the total number of observed data points; This is the normalized predicted displacement output by the displacement network at the i-th observation point; The displacement observation value of the i-th observation point is obtained by independent normalization according to the borehole.

[0037] S303. Establish a creep constitutive loss term by restoring the normalized predicted displacement to physical displacement according to the displacement scale and differentiating it with respect to physical time to obtain the physical displacement rate. Penalize the deviation between the physical displacement rate and the theoretical creep rate determined by the creep parameters to constrain the time evolution of the displacement to conform to the Burgers viscoelastic creep model. Specifically, the creep constitutive loss term is: ; In the formula, For creep constitutive loss; Let be the first partial derivative of the displacement with respect to time at the i-th observation point, i.e., the displacement rate, in mm / d; , , These are the steady-state creep rate parameter, transient creep amplitude parameter, and transient decay rate parameter output by the creep parameter network at the i-th observation point, respectively. Let be the time of the i-th observation point.

[0038] The creep constitutive loss term is obtained by differentiating the Burgers viscoelastic creep model with respect to time, i.e., the theoretical displacement rate is... .

[0039] S304. Establish a strain smoothing loss term by penalizing the mean square value of the third derivative of the normalized predicted displacement with respect to the normalized depth at the placement point to constrain strain smoothing along the depth direction. Specifically, the strain smoothing loss term is: ; In the formula, For strain smoothness loss; M is the number of placement points. The third partial derivative of the normalized predicted displacement with respect to the normalized depth at the j-th configuration point is obtained by automatically differentiating the displacement network output three times consecutively along the normalized depth direction.

[0040] The configuration points are generated by Latin hypercube sampling within the coordinate domain formed by the normalized variables, and are resampled in each training round. It should be noted that the third-order partial derivative is penalized instead of the second-order partial derivative in this embodiment in order to suppress high-frequency oscillations while allowing strain to form concentrated bands along the depth, so as not to impair the strain peak value on which the slip surface identification depends.

[0041] S305. Establish a creep parameter spatial smoothing loss term, and constrain the continuous change of each creep parameter along the main slip direction by penalizing the first-order partial derivatives of the steady-state creep rate parameter, the transient creep amplitude parameter, and the transient decay rate parameter with respect to the normalized spatial coordinates. Specifically, the creep parameter spatial smoothing loss term is: ; In the formula, For creep parameter space smoothing loss; , , These are the first-order partial derivatives of the steady-state creep rate parameter, transient creep amplitude parameter, and transient decay rate parameter at the j-th configuration point with respect to the normalized spatial coordinates.

[0042] The initial elastic displacement parameters do not need to be included in the strain smoothing loss term. The initial elastic displacement parameters absorb the observation baseline offset and do not need to be continuous along the main slip direction.

[0043] S306. Establish a displacement monotonicity loss term, and constrain the displacement not to increase from the front edge of the landslide to the rear edge of the landslide along the main sliding direction by penalizing the square mean of the first-order partial derivative of the normalized predicted displacement with respect to the normalized spatial coordinates after one-sided rectification.

[0044] Specifically, the displacement monotonicity loss term is: ; ; In the formula, This is due to displacement monotonicity loss; Let be the monotonic default value at the j-th configuration point; Let be the first-order partial derivative of the normalized predicted displacement with respect to the normalized spatial coordinates at the j-th configuration point; In this embodiment, single-sided rectification is used so that the constraint only penalizes the portion of the displacement that increases from the leading edge to the trailing edge along the main sliding direction, without interfering with the normal decreasing trend.

[0045] It should be noted that the four physical constraints have a clear division of labor in the three coordinate directions: the creep constitutive loss term constrains the evolution law in the time direction, the strain smoothing loss term constrains the continuity in the depth direction, the creep parameter spatial smoothing loss term constrains the parameter continuity along the main sliding direction, and the displacement monotonicity loss term injects the spatial prior of the traction landslide displacement decreasing from the leading edge to the trailing edge; among them, the creep parameter spatial smoothing loss term and the displacement monotonicity loss term are the key constraints that distinguish multi-hole joint modeling from single-hole modeling. Through the creep parameter spatial smoothing loss term and the displacement monotonicity loss term, the spatial coordinates are transformed from hole position labels into interpolable continuous variables.

[0046] S307. A composite loss function is constructed using the data fidelity loss term, the creep constitutive loss term, the strain smoothing loss term, the creep parameter spatial smoothing loss term, and the displacement monotonicity loss term. Specifically, the composite loss function is: ; In the formula, L is the composite loss function; Weights for strain smoothness loss; Weights for creep constitutive loss; Weights for smoothing the creep parameter space; Weights for displacement monotonicity loss; As the course weight; In this embodiment, the course weight is a piecewise linear function of the training rounds. When the training round is less than the warm-up termination round, the course weight is 0; when the training round is not less than the warm-up termination round and less than the climb termination round, the course weight increases linearly from 0 to 1; when the training round is not less than the climb termination round, the course weight is 1.

[0047] The expression for course weights is as follows: ;

[0048] In the formula, e represents the training round; The preheating cycle is terminated; The climb ends the cycle.

[0049] It should be noted that in the early stages of model training, only the data fidelity loss term is used to optimize the porous joint physical information neural network model, allowing the displacement network to first fit the basic displacement pattern. As the training rounds increase, the course weights gradually increase, and creep constitutive loss term, strain smoothing loss term, creep parameter space smoothing loss term, and displacement monotonicity loss term are gradually introduced according to the course weights to avoid instability of higher-order derivative terms in the early stages of training. When the training rounds reach the climbing termination round, the creep constitutive loss term, strain smoothing loss term, creep parameter space smoothing loss term, and displacement monotonicity loss term all participate in optimization with the maximum weight, thereby ensuring the data fitting accuracy while making the displacement field satisfy physical constraints such as creep constitutive, strain smoothing, creep parameter space continuity, and displacement monotonicity.

[0050] S4. Using the normalized variables as input, a phased training strategy is adopted to train the physical information neural network model based on the composite loss function to obtain the trained multi-pore joint physical information neural network model. For example, step S4 specifically includes the following steps: S401. With minimizing the data fidelity loss term as the optimization objective, the porous joint physical information neural network model is pre-trained to obtain the pre-trained porous joint physical information neural network model. In some embodiments, step S401 is the data pre-training stage, in which the optimization objective is to minimize the data fidelity loss term. An adaptive moment estimation optimizer is used in conjunction with cosine annealing learning rate scheduling to enable the displacement network to first fit the basic displacement pattern.

[0051] S402. Using course weights that increase from zero to one with each training round, the creep constitutive loss term, the strain smoothing loss term, the creep parameter space smoothing loss term, and the displacement monotonicity loss term are modulated, and the modulated loss terms are superimposed with the data fidelity loss term to obtain the modulated composite loss function. S403. With minimizing the modulated composite loss function as the optimization objective, the pre-trained multi-pore joint physical information neural network model is fine-tuned to obtain the fine-tuned multi-pore joint physical information neural network model. In some embodiments, steps S402 and S403 are physical constraint fine-tuning stages, with the optimization objective of minimizing the modulated composite loss function. Differential learning rates are set for the displacement network and the creep parameter network, and the learning rate of the displacement network is lower than that of the creep parameter network.

[0052] In this embodiment, the purpose of setting the differential learning rate is as follows: after pre-training, the displacement network is close to the data manifold and needs to maintain the fitting accuracy with a small learning rate, while the creep parameter network is not supervised in the pre-training stage and needs to quickly invert the constitutive parameters with a large learning rate.

[0053] S404. Using a quasi-Newton optimizer, with minimizing the composite loss function as the optimization objective, the fine-tuned porous joint physical information neural network model is refined to obtain the trained porous joint physical information neural network model.

[0054] In some embodiments, step S404 is a quasi-Newton refinement stage, in which a finite-memory quasi-Newton optimizer with strong Wolfe line search is used to globally refine the composite loss function.

[0055] It should be noted that the role of the phased strategy is as follows: when optimizing directly with the composite loss function from random initialization, the gradient direction given by the higher-order derivative term has no physical meaning before the displacement field is formed, which can easily lead the network into a trivial solution. Therefore, the displacement field is first formed with the data fidelity loss term, and then physical constraints are gradually introduced, which can simultaneously obtain fitting accuracy and physical consistency.

[0056] S5. Generate multiple input coordinates within the range of the normalized variable, input each input coordinate into the trained displacement network to obtain the normalized predicted displacement at each input coordinate, and perform inverse normalization on the normalized predicted displacement to obtain the actual displacement at each input coordinate, so as to form a three-dimensional continuous displacement field.

[0057] For example, step S5 specifically includes the following steps: S501. Generate an input grid within the range of the normalized depth, the normalized time, and the normalized spatial coordinates. The input grid includes multiple input grid points. The input coordinates of the input grid points include the normalized depth, normalized time, and normalized spatial coordinates of the input grid points; S502. Input the input coordinates of each input grid point into the trained displacement network to obtain the normalized predicted displacement at each input grid point. S503. Based on the spatial coordinates of each monitoring hole and its displacement normalization parameters, construct a continuous distribution of displacement normalization parameters along the main sliding direction; Specifically, the displacement normalization parameters include the minimum displacement observation value and displacement amplitude of each monitoring hole; The displacement amplitude is the difference between the maximum and minimum displacement observations, and its calculation formula is as follows: ; In the formula, This refers to the displacement amplitude; This represents the maximum displacement observation value; This is the minimum displacement observation value; In some embodiments, for locations where no monitoring hole is provided, piecewise linear interpolation or conformal interpolation is used to construct the minimum displacement observation value and the displacement amplitude; For locations outside the spatial coordinate range of all monitoring holes, the displacement normalization parameter of the nearest end monitoring hole is used; S504. Based on the continuous distribution of the displacement normalization parameters, the normalized predicted displacements are inversely normalized to obtain the actual physical displacements at each input grid point. The inverse normalization function expression is: ; In the formula, This represents the actual physical displacement after restoration. The normalized predicted displacement output by the displacement network; In some embodiments, a non-negative constraint is applied to the obtained actual physical displacement so that the actual displacement conforms to the non-negative physical characteristics of the cumulative lateral displacement.

[0058] S505. Arrange the actual physical displacements according to the input coordinates of each input grid point to form a three-dimensional continuous displacement field; When querying or automatically differentiating the three-dimensional continuous displacement field, the same coordinate transformation and displacement scale transformation as in steps S501 to S504 can be used.

[0059] S6. Input the preset normalized depth and normalized spatial coordinates into the trained creep parameter network to obtain the corresponding creep parameters and generate a synthetic twin dataset superimposed with spatial heterogeneity features. For example, step S6 specifically includes the following steps: S601. Fix the weights of the trained creep parameter network, and input the preset normalized depth and normalized spatial coordinates into the trained creep parameter network to obtain the creep parameters at the corresponding coordinates, so as to form a two-dimensional creep parameter field. Optionally, the preset normalized depth and normalized spatial coordinates refer to multiple coordinate points preset within the value range of the normalized variable; wherein, the multiple coordinate points can be generated according to a preset resolution, for example, taking values ​​at equal intervals within the value range of the normalized depth and normalized spatial coordinates to form a two-dimensional coordinate grid.

[0060] S602. Take the value of the two-dimensional creep parameter field at the center depth of the slip surface, or take the average value within a preset slip surface depth window centered on the center depth of the slip surface, so as to obtain a one-dimensional reference parameter that varies along the main slip direction. S603. Based on the one-dimensional reference parameters, set the reference trend of each creep parameter along the main sliding direction; Among them, the baseline trend decreases from the leading edge of the landslide to the trailing edge of the landslide. Specifically, the baseline trend of creep parameters is as follows: ; ; ; ; In the formula, , , , These are the baseline trends of the initial elastic displacement, steady-state creep rate, transient creep amplitude, and transient decay rate before superposition of heterogeneity, respectively. , , , These are the values ​​at the leading edge; , , , These are the decreasing coefficients of the initial elastic displacement, steady-state creep rate, transient creep amplitude, and transient decay rate along the main slip direction, respectively. These are normalized spatial coordinates. Lower bound constraints on parameters ensure that the baseline trends of steady-state creep rate, transient creep amplitude, and transient decay rate are all positive.

[0061] S604. Multiply the baseline trend corresponding to each creep parameter by the heterogeneity multiplier to obtain the superimposed creep parameters, so that each creep parameter is superimposed with the spatial heterogeneity feature; Among them, the heterogeneity multipliers include sinusoidal wave terms and local anomaly terms; Specifically, the expression for the heterogeneity multiplier is: ; in, Let be the heterogeneity multiplier for parameter q, where q takes the values ​​A, B, C, or D; Let q be the amplitude of the sinusoidal heterogeneity. The characteristic heterogeneous wavelength; Let q be the phase of the parameter, and each parameter takes a different phase to avoid the perturbations of the four parameters being completely correlated; Let q be the local anomaly intensity. This is the location of the local anomaly center, used to simulate local structures such as faults or weak interlayers; This represents the width of a local anomaly.

[0062] Preferably, by limiting the sinusoidal heterogeneity amplitude to remain positive, or by applying positive truncation to the superimposed steady-state creep rate, transient creep amplitude, and transient decay rate, the synthesized data can be guaranteed to meet the physical range of the constitutive parameters.

[0063] Specifically, the expression for the superimposed creep parameters is as follows: ; ; ; ; In the formula, These are the initial elastic displacement parameters after superposition; These are the superimposed steady-state creep rate parameters; These are the parameters of the transient creep amplitude after superposition; The parameters represent the transient decay rate after superposition. In this embodiment, the baseline trend is modulated by the heterogeneity multiplier, so that the superimposed creep parameter is further superimposed with sinusoidal fluctuations and local anomalies on the baseline trend that decreases along the main sliding direction, thereby constituting the spatial heterogeneity feature.

[0064] S605. Based on the preset depth shape function, construct a synthetic displacement field according to the superimposed creep parameters; In some embodiments, a synthetic displacement field is constructed based on a preset depth shape function and the product of the analytical solution of the Burgers viscoelastic creep model along the time direction and the depth shape function. Specifically, the preset depth shape function is a Gaussian depth shape function centered on the surface depth, and its function expression is: ; ; In the formula, Preset depth shape function; For depth; The depth of the smooth surface center in the synthetic twin data; The amplitude of the undulation of the sliding surface depth along the main sliding direction; The Gaussian width of the depth shape function; This is the horizontal distance from the leading edge of the landslide along the main sliding direction; This is the maximum value of the horizontal distance; Specifically, the analytical solution of the Burgers viscoelastic creep model along the time direction is the creep displacement that decays with time, and its expression is: ; In the formula, This is the analytical solution of Burgers' viscoelastic creep model along the time direction; t is time.

[0065] Specifically, the expression for the synthesized displacement field is: ; In the formula, To synthesize twin displacements; It is Gaussian noise, and its standard deviation is taken as η times the standard deviation of the hole displacement, where η can be 0.02; S606. Sampling is performed on the synthetic displacement field according to the preset monitoring hole positions, depths and time points to obtain the synthetic twin dataset.

[0066] It should be noted that on a real landslide site, it is impossible to test the interpolation effect under different monitoring hole densities by increasing or decreasing the number of monitoring holes. Therefore, this step uses the creep parameters output by the creep parameter network after training in S5 as a benchmark to construct a synthetic twin dataset that is consistent with the real site in terms of statistical characteristics. On this synthetic twin dataset, a subset of monitoring holes can be arbitrarily extracted to simulate different numbers of monitoring holes and monitoring hole spacing, thereby transforming the monitoring hole density into a controllable variable.

[0067] To enable the synthetic twin dataset to effectively test interpolation capabilities, during its generation process, the heterogeneity multiplier is multiplied by the baseline trend, thereby incorporating nonlinear spatial features not assumed by the multi-pore joint physical information neural network model, placing the model in an intentionally mismatched state. If the synthetic data fully conforms to the model assumptions, the model can accurately interpolate the displacement between monitoring holes based on its spatial smoothness assumption, causing the interpolation accuracy under different monitoring hole densities to tend to be consistent. This would result in the density criterion being unable to effectively distinguish the interpolation capabilities of each candidate deployment configuration, and thus unable to determine the minimum number of monitoring holes and the maximum allowable spacing.

[0068] S7. Obtain geological data of the landslide, and set multiple candidate deployment configurations based on the geological data; Specifically, based on the geological data of the landslide, the characteristic heterogeneity wavelength of the landslide is determined, and half of the characteristic heterogeneity wavelength is taken as the initial candidate spacing. Multiple candidate layout configurations with different numbers of monitoring holes and different spacings are generated in ascending order of the number of monitoring holes.

[0069] The geological data includes geological exploration data or data from pre-monitoring wells.

[0070] S8. Based on the synthetic twin dataset, leave-one-out cross-validation is performed on each of the candidate deployment configurations to obtain the average determination coefficient of each candidate deployment configuration. The candidate deployment configurations whose average determination coefficients meet the preset conditions are determined as deployment schemes and output.

[0071] For example, step S8 specifically includes the following steps: S801. Using the mean and standard deviation of the synthetic displacement data of all monitoring holes in the synthetic twin dataset, the synthetic displacement data is globally standardized to obtain standard synthetic displacement data. Specifically, the globally standardized expression is: ; In the formula, The standardized composite displacement data is the globally standardized composite displacement data; u is the cumulative lateral displacement observation data of any monitoring hole. and These represent the mean and standard deviation of the cumulative lateral displacement observation data from all monitoring wells, respectively.

[0072] The global standardized parameters are estimated only from the training holes other than the exit holes, and the same mean and standard deviation are used to transform the synthetic displacement data of the training holes and the exit holes to avoid leakage of exit hole information.

[0073] S802. For each monitoring hole in the candidate deployment configuration, the monitoring hole is sequentially designated as a reserved hole, and leave-one cross-validation is performed to obtain the determination coefficient of each monitoring hole in the candidate deployment configuration. In some embodiments, the specific steps of leave-one cross-validation in step S802 include: The multi-pore joint physical information neural network model is trained using the standard synthetic displacement data corresponding to the remaining monitoring holes (excluding the reserved holes) in the candidate layout configuration in the synthetic twin dataset. The normalized coordinates of the opening in the synthetic twin dataset are input into the trained multi-pore joint physical information neural network model to obtain the displacement prediction value corresponding to the opening. Based on the standard synthetic displacement data and displacement prediction value corresponding to the reserved hole, the determination coefficient of the reserved hole is calculated in the same standardized domain. Specifically, the determination coefficient of the h-th outlet hole is calculated according to the following formula: ; In the formula, The number of data points for the h-th exit hole; This represents the predicted displacement value of the model at the i-th data point in the h-th opening; The composite twin displacement at the i-th data point in the h-th outlet hole; The mean of all the synthetic twin displacements of the h-th exit hole; is the determination coefficient for the h-th exit hole.

[0074] In this embodiment, by using globally standardized synthetic displacement data, the multi-hole joint physical information neural network model can learn the real variation law of displacement magnitude along the main sliding direction, thereby having the ability to predict the absolute magnitude of the position of the unmonitored hole.

[0075] S803. The monitoring holes located at both ends in the candidate deployment configuration are determined as boundary holes, the remaining monitoring holes are determined as internal holes, and the average value of the determination coefficients corresponding to all the internal holes is taken as the average determination coefficient of the candidate deployment configuration. Among them, when a candidate configuration has only two monitoring holes, there are no internal holes, and the average determination coefficient of the internal holes in this configuration is marked as invalid. Specifically, the arithmetic mean of the determination coefficients of all the internal holes is used as the average determination coefficient of the internal holes of the candidate layout configuration. Boundary holes are not included in the arithmetic mean calculation because the prediction of boundary holes is spatial extrapolation rather than spatial interpolation. If the determination coefficients of boundary holes are included in the arithmetic mean, it will mask the true interpolation capability of each candidate layout configuration.

[0076] S804. Perform steps S802 and S803 on the current candidate deployment configuration in ascending order of the number of monitoring holes to obtain the average determination coefficient of the candidate deployment configuration. Determine whether the average coefficient of determination is less than a preset threshold; If the number is less than the required number, then steps S802 and S803 are executed for the next candidate deployment configuration in ascending order of the number of monitoring holes. If it is not less than the preset threshold, the current candidate deployment configuration that is not less than the preset threshold for the first time is determined as the selected deployment configuration, and the number of monitoring holes and the average distance between adjacent monitoring holes corresponding to the current candidate deployment configuration are determined as the minimum number of monitoring holes and the maximum allowable distance, respectively.

[0077] If all candidate deployment configurations are less than the preset threshold, a prompt will be output indicating that the preset threshold cannot be met within the current candidate range, and a prompt will be made to expand the candidate deployment configuration range; optionally, the preset threshold is 0.8.

[0078] S805. In the selected deployment configuration, the section where the monitoring hole with the lowest determination coefficient in the internal borehole is located is identified as a suspected abnormal section or a section that needs to be densified, and is verified in conjunction with geological data. Specifically, a hole-by-hole determination coefficient analysis is performed on the selected deployment configuration that meets the preset threshold. The section with the lowest determination coefficient among the internal monitoring holes is identified as a suspected abnormal section or a section that needs to be densified. The analysis is then verified in conjunction with geological data, boundary conditions, and observation quality. S806. The spacing between the monitored holes in the section after verification is increased to less than half of the maximum allowable spacing, and the boundary holes on the side with smaller displacement magnitude are laid out first to form the layout scheme.

[0079] In some embodiments, the method further includes: S9. Automatically differentiate the continuous displacement field along the depth direction to obtain the strain field, and determine the sliding surface position based on the concentrated zone of the strain field along the depth direction. For example, step S9 specifically includes the following steps: S901. Automatically differentiate the continuous displacement field along the depth direction to obtain the strain field; Specifically, the functional expression for the strain field is: ; ; In the formula, Let Z be the physical strain of the k-th monitoring hole at normalized depth z and normalized time t; The physical displacement after being restored according to the displacement scaling function; For strain field; The reconstructed displacement after inverse normalization; The strain field is obtained through automatic differential analysis without using finite differences, thus avoiding amplification of observation noise and achieving implicit noise reduction of inclinometer observation data.

[0080] S902. For each monitoring hole, calculate the distribution curve of the strain field along the depth direction at each time point, and average the distribution curve of the same monitoring hole at each time point along the time direction to obtain the time-averaged strain profile of the monitoring hole. Specifically, the time-averaged strain profile, the preset strain threshold, and the candidate slip surface depth set are calculated according to the following formulas: ; ; ; In the formula, This is the time-averaged strain profile of the k-th monitoring well; This represents the number of time sampling points; This is the strain threshold coefficient; The preset strain threshold; This represents the maximum value of the time-averaged strain profile along the depth direction at the k-th monitoring well; Let be the set of candidate slip surface depths for the k-th monitoring hole; The candidate slip surface depth.

[0081] Preferably, when the candidate set contains only one element, that element is determined as the slip surface depth; when the candidate set contains multiple elements, multiple candidate slip surface depths are output, or the depth with the largest strain peak is selected.

[0082] S903. Locate the local maximum point in the time-averaged strain profile and determine whether the strain value of the local maximum point exceeds the preset strain threshold. If it exceeds the limit, the depth corresponding to the local maximum point is determined as the sliding surface position; If it does not exceed the limit, then the local maximum point is excluded.

[0083] In some embodiments, the method further includes: S10. Using the trained multi-pore joint physical information neural network model, extrapolation prediction is performed along the time direction to obtain the extrapolated displacement prediction value.

[0084] S1001. Divide the displacement observation data of each monitoring hole into training segments and verification segments according to the time sequence, and retrain the multi-hole joint physical information neural network model with the displacement observation data corresponding to the training segments of all monitoring holes. S1002. Using the retrained multi-pore joint physical information neural network model, the displacement at each time point in the test section is predicted to obtain the extrapolated displacement prediction value at each time point in the test section. S1003. Calculate the extrapolation determination coefficient between each extrapolated displacement prediction value and the displacement observation data corresponding to the test section, and use the extrapolation determination coefficient to characterize the extrapolation accuracy.

[0085] Application Example 1

[0086] To further illustrate the technical solution of this invention, an application example is provided as follows: This application example uses a massive, thick-layered, traction-type landslide at kilometer marker K61 of an under-construction highway in southern China as its application scenario. The landslide has a tongue-shaped planform, a main sliding direction of 311 degrees, a main axis length of 370m, a transverse axis width of 290m, a height difference of approximately 100m between the front and rear edges, and an affected area of ​​approximately 1.0 × 10⁻⁶ m². 5 m 2 The volume of the sliding body is approximately 1.5 × 10⁻⁶. 6 m 3 The landslide body consisted of Quaternary gravelly clay, silty clay, and weathered debris, with a thickness of 15m to 20m. The sliding bed was Upper Triassic argillaceous siltstone. After the initial instability, the slope entered a creep stage with a decreasing displacement rate.

[0087] Three monitoring wells were set up sequentially from the front edge to the rear edge along the main sliding direction of the landslide. A borehole inclinometer was installed in each monitoring well to collect cumulative lateral displacement observation data. The observation data of each monitoring well are shown in Table 1.

[0088] Table 1: Statistical analysis of observation data from the three monitoring wells

[0089] All three monitoring wells were used for observations at 15 depth layers from November 2025 to February 2026, yielding a total of 4076 data points. The distances between adjacent wells were 119.4m and 164.1m, respectively, with an average distance of 141.8m. The monitoring period coincided with the dry season, with a cumulative rainfall of only 86.48mm. The slope showed no significant rainfall response; therefore, no external force term related to rainfall was included in the physical constraints of this application example.

[0090] According to the normalization method described in the aforementioned specific implementation, in this application example, the minimum observation depth is 1m and the maximum observation depth is 30m; the minimum observation time is 0d and the maximum observation time is 100d; the maximum spatial coordinate is 283.5m.

[0091] Because the displacement magnitudes of the leading and trailing edge monitoring holes differ significantly, if absolute displacement is directly used for joint training, the monitoring hole with the larger displacement will dominate the loss function. To illustrate the effect of independent normalization per monitoring hole used in this application example, the displacement distribution before and after normalization is compared, and the results are as follows: Figure 3 As shown; in, Figure 3 (a) Comparison of the cumulative lateral displacement magnitudes of the three monitoring holes before normalization. It can be seen that the average displacements of the leading edge monitoring holes and the trailing edge monitoring holes differ by about two orders of magnitude. Figure 3 (b) shows the displacement distribution after independent normalization of the monitoring holes. It can be seen that the displacements of the three monitoring holes all fall within the range of 0 to 1, and the magnitude of the contribution of each monitoring hole to the loss function tends to be consistent.

[0092] According to the aforementioned specific implementation method for constructing a multi-pore joint physical information neural network model, in this application example, the displacement network consists of 8 fully connected hidden layers, each containing 128 neurons, with an output dimension of 1; the creep parameter network consists of 4 fully connected hidden layers, each containing 64 neurons, with an output dimension of 4.

[0093] Figure 4This diagram illustrates the course weight scheduling and three-stage training strategy in this embodiment. In this application embodiment, the total number of observation data points N is 4076, and the number of configuration points M is 5000. The strain smoothness loss weight is 0.1, the creep constitutive loss weight is 0.01, the creep parameter space smoothness loss weight is 0.05, and the displacement monotonicity loss weight is 0.05. The course weight parameters include the preheating termination round. Take 1000, climb to terminate round The four weighting coefficients mentioned above, set to 4000, were determined through preliminary experiments on the leading-edge monitoring hole ZK61340-2; among them, the course weights are arranged in a piecewise linear manner with the training rounds as follows: Figure 4 As shown in (a), the number of rounds and learning rate configurations for the three-stage training are as follows: Figure 4 As shown in (b).

[0094] Training was conducted according to the three-stage training strategy described in the aforementioned specific implementation method, with the random seed fixed at 42, and the total training time on a single NVIDIA RTX3090 graphics processor was approximately 30 minutes.

[0095] According to the continuous displacement field reconstruction method described in the aforementioned specific implementation method, a depth-time-space continuous displacement field covering multiple monitoring holes and the area between the monitoring holes is obtained.

[0096] The reconstruction accuracy after joint training of the three monitoring holes is shown in Table 2. To visually demonstrate the reconstruction effect of each monitoring hole, Figure 5 (a) shows the reconstruction determination coefficients for each monitoring well. Figure 5 (b) shows the reconstructed root mean square error for each monitoring well.

[0097] Table 2: Accuracy of Joint Reconstruction of Three Monitoring Holes

[0098] As shown in Table 2, the reconstruction determination coefficients of the three monitoring holes range from 0.954 to 0.993. The leading edge monitoring hole ZK61340-2 has the highest determination coefficient, with its root mean square error less than 3% of its displacement range. The middle monitoring hole ZK61340-3 has the lowest determination coefficient among the three, because there are three competing sliding surfaces in this monitoring hole, resulting in the superposition of multiple deformation signals. The trailing edge monitoring hole ZK61340-4 has an average displacement of only 1.34 mm, yet its determination coefficient still reaches 0.972, indicating that independent normalization by monitoring hole effectively prevents the strong leading edge signal from suppressing the weak trailing edge signal during training.

[0099] According to the synthetic twin dataset generation method described in the aforementioned specific implementation, the parameter values ​​in this application example are as follows: For the leading edge values ​​of the creep parameter reference trends, the initial elastic displacement reference is 50 mm, the steady-state creep rate reference is 0.5 mm / d, the transient creep amplitude reference is 30 mm, and the transient decay rate reference is 0.05 / d; for the decreasing coefficients, the initial elastic displacement reference is 0.8, the steady-state creep rate reference is 0.9, the transient creep amplitude reference is 0.9, and the transient decay rate reference is 0.03 / d; for the sinusoidal heterogeneity amplitude, the sinusoidal heterogeneity amplitudes of the initial elastic displacement reference, steady-state creep rate reference, and transient creep amplitude reference are all 0.20, and the sinusoidal heterogeneity amplitude of the transient decay rate reference is 0.20. The amplitude of the string heterogeneity is set to 0.05; the characteristic heterogeneity wavelength is set to 120 m; in terms of phase, the phase of the initial elastic displacement reference is set to 0 rad, the phase of the steady-state creep rate reference is set to 1.2 rad, the phase of the transient creep amplitude reference is set to 2.5 rad, and the phase of the transient decay rate reference is set to 3.8 rad; in terms of local anomaly intensity, the local anomaly intensity of the initial elastic displacement reference is set to 0.20, the local anomaly intensity of the steady-state creep rate reference is set to 0.30, the local anomaly intensity of the transient creep amplitude reference is set to 0.15, and the local anomaly intensity of the transient decay rate reference is set to −0.075; the local anomaly center position is set to 160 m; the local anomaly width is set to 40 m; the sliding surface reference depth is set to 15 m; the sliding surface depth fluctuation amplitude is set to 2 m; the Gaussian width is set to 3 m; and the standard deviation of the Gaussian noise is set to 0.02 times the standard deviation of the displacement of the monitoring hole.

[0100] The synthetic twin dataset consists of seven monitoring wells, SYN-1 to SYN-7, spaced at equal intervals from 0m to 283.5m. The horizontal distances along the main sliding direction are 0m, 47.25m, 94.5m, 141.75m, 189.0m, 236.25m, and 283.5m, respectively. Fifteen points are taken in the depth direction from 1m to 30m, and 90 points are taken in the time direction from 0d to 100d. Each monitoring well has 1350 data points, for a total of 9450 data points.

[0101] To demonstrate the spatial distribution characteristics of the synthetic twin dataset and how they change over time, Figure 6 The displacement contour lines of the synthetic twin dataset in the depth-space plane are given; where, Figure 6 (a) shows the displacement contour lines at t=20d. Figure 6 (b) shows the displacement contour lines at t=80d.

[0102] Depend on Figure 6It can be seen that the displacement is the largest at the leading edge, concentrated near the center depth of the slip surface along the depth direction, and generally decreases towards the trailing edge. Due to the superposition of sinusoidal heterogeneity at a wavelength of 120m and local anomaly at 160m, a local rebound occurs at a displacement of about 120m to 130m, forming secondary closed contour lines. The contour line shape remains similar between the two time points, while the amplitude increases overall and the increase decreases with time, which is consistent with the deformation characteristics of the deceleration creep stage.

[0103] Following the candidate deployment configuration generation method described in the aforementioned specific implementation, in this application example, the characteristic heterogeneity wavelength λh is set to 120m, and the initial candidate spacing is 60m. Three candidate deployment configurations are generated in ascending order of the number of monitoring holes: a 3-monitoring-hole configuration selects SYN-1, SYN-4, and SYN-7, with an average spacing of 141.8m between adjacent monitoring holes; a 5-monitoring-hole configuration selects SYN-1, SYN-2, SYN-4, SYN-5, and SYN-7, with an average spacing of 70.9m between adjacent monitoring holes; and a 7-monitoring-hole configuration selects all 7 monitoring holes, with an average spacing of 47.2m between adjacent monitoring holes.

[0104] According to the leave-one-out cross-validation method described in the aforementioned specific implementation, leave-one-out cross-validation is performed on each candidate deployment configuration on the synthetic twin dataset, with a preset threshold of 0.8.

[0105] The leave-one-out cross-validation results for the three candidate deployment configurations are shown in Table 3. To visually demonstrate the trend of the average coefficient of determination of the internal boreholes as the number of monitoring boreholes changes, Figure 7 The relationship curve between the number of monitoring holes and the average coefficient of determination of the internal holes is given;

[0106] Table 3: Monitoring Hole Layout Density and Interpolation Accuracy

[0107] As shown in Table 3, the average determination coefficient of internal boreholes exhibits a threshold-like jump rather than a gradual improvement with the change in the number of monitoring boreholes. The average determination coefficient of internal boreholes with a 3-monitoring-hole configuration is only 0.53, indicating spatial interpolation failure; the average determination coefficient of internal boreholes with a 5-monitoring-hole configuration rises to 0.87, exceeding the preset threshold; the average determination coefficient with a 7-monitoring-hole configuration further increases to 0.92, but the increase relative to the 5-monitoring-hole configuration is only 0.05, indicating significantly diminishing marginal returns. Based on this, the minimum number of monitoring boreholes required for reliable reconstruction of the displacement field of this landslide is determined to be 5, and the maximum allowable spacing is 70.9 m.

[0108] As a comparison, the same leave-one-out cross-validation was performed on real three-well data, and the average coefficient of determination was -29.5. The average coefficient of determination for the synthetic twin dataset with a three-well configuration was -1.45. Both failed in the same direction, indicating that the interpolation failure under the three-well configuration stemmed from the inherent lack of information due to sparse spatial sampling, rather than a model implementation error. The failure of the real data was more severe, indicating that the complexity of the real strata is higher than that of the simplified parameterization of the synthetic twin dataset. Therefore, the number of monitoring wells required for the real site will not be lower than the lower limit given by the synthetic twin dataset.

[0109] According to the slip surface identification and time extrapolation method described in the aforementioned specific implementation method, in this application example, the preset strain threshold is 0.5 times the maximum value of the time-averaged strain profile; the displacement observation data of each monitoring hole is divided into a training set and a test set according to time sequence, with a division ratio of 70% and 30%, that is, the data from the first 67 to 70 days is used as the training segment, and the data from the next 28 to 30 days is used as the test segment to output the extrapolated displacement prediction value.

[0110] The extrapolation results are shown in Table 4. To demonstrate the degree of agreement between the extrapolated predicted values ​​in the time direction and the observed values ​​in the test section, Figure 8 Extrapolated prediction curves for each monitoring well within the inspection section are given.

[0111] Table 4: Prediction Accuracy of Time Direction Extrapolation

[0112] As shown in Table 4, the coefficients of determination for the test sets of all three monitoring wells exceed 0.75, indicating that the creep constitutive constraints have an effective physical regularization effect on the temporal evolution of the network, suppressing overfitting to the training period. Temporal extrapolation can be performed using only time-series data from a single monitoring well, without being limited by the number of spatial monitoring wells, and therefore can be used for early warning in the early stages of monitoring.

[0113] According to the output method of the monitoring hole layout scheme described in the aforementioned specific implementation method, a monitoring hole-by-monitoring hole determination coefficient analysis is performed on the selected layout configuration that meets the preset threshold.

[0114] The results of the cross-validation with one monitoring well left out for each of the seven monitoring wells are shown in Table 5. To facilitate comparison of the interpolation reliability of each monitoring well, Figure 9 A bar chart of the leave-one-out cross-validation determination coefficients for each monitoring well under a 7-well configuration is presented.

[0115] Table 5: Seven-well configuration, one well left for each monitoring well cross-validation results

[0116] As shown in Table 5, the extrapolation reliability of boundary holes is determined by the signal magnitude rather than spatial distance: the extrapolation determination coefficient of the leading-edge boundary hole SYN-1 is 0.861; the extrapolation determination coefficient of the trailing-edge boundary hole SYN-7 is as low as 0.061. Therefore, boundary holes should be preferentially placed on the side with smaller displacement magnitude; in this application example, boundary holes are preferentially placed at the trailing edge. The determination coefficients of the internal holes SYN-4 and SYN-5 adjacent to the local anomaly section are 0.937 and 0.841, respectively, which are about 5% and 15% lower than those of SYN-2 and SYN-3, which are far from the anomaly section (0.986 and 0.984). Among them, SYN-5 has the lowest determination coefficient of all internal holes. Therefore, for anomaly sections with x ranging from 140m to 190m, the local spacing should be reduced from the maximum allowable spacing of 70.9m to half of that, i.e., below 35.45m; in this application example, 35m is used.

[0117] Based on the above diagnostic results, the monitoring well layout scheme for this landslide in this application example is as follows: At least five monitoring wells are deployed along the main sliding direction, with an average spacing between adjacent wells not exceeding 70.9 m; in the geological anomaly section with an x-axis of approximately 160 m, the spacing is locally increased to below 35 m; boundary wells are prioritized on the low-displacement side at the rear edge. In the initial monitoring phase, when the number of monitoring wells is less than five, early warning is implemented by extrapolating along the time direction. Once the monitoring well network reaches the aforementioned density, full-domain displacement field reconstruction and continuous slip surface identification are carried out, forming a phased monitoring strategy of sparse early warning followed by denser reconstruction.

[0118] In summary, this application example verifies that the method of the present invention can reconstruct the continuous displacement field across monitoring holes under real landslide monitoring data conditions through a multi-hole joint physical information neural network. Furthermore, it quantitatively determines the minimum number of monitoring holes and the maximum allowable spacing on a synthetic twin dataset through leave-one-out cross-validation, providing a quantifiable technical basis for landslide underground displacement field monitoring and monitoring hole layout design.

[0119] Example 2

[0120] like Figure 10 As shown, this embodiment provides a porous physical information neural network system for landslide displacement field reconstruction and pore layout, including: The data acquisition module 101 is used to acquire observation data from multiple monitoring holes deployed along the main sliding direction of the landslide, and to acquire the spatial coordinates of each monitoring hole along the main sliding direction. The observation data includes cumulative lateral displacement observation data, the observation depth corresponding to the cumulative lateral displacement observation data, and the observation time; The data processing module 102 is used to independently normalize the displacement observation data according to the monitoring hole to obtain normalized displacement, and to perform global shared parameter normalization on the observation depth, the observation time and the spatial coordinates to obtain normalized variables; The normalized variables include normalization depth, normalization time, and normalized spatial coordinates; Model building module 103 is used to build a porous joint physical information neural network model and establish a composite loss function; The porous joint physical information neural network model includes a displacement network and a creep parameter network; The model training module 104 is used to train the physical information neural network model based on the composite loss function according to the normalized variables and normalized displacements, using a phased training strategy, to obtain the trained multi-pore joint physical information neural network model. The displacement field reconstruction module 105 is used to generate multiple input coordinates within the range of the normalized variable, input each input coordinate into the trained displacement network to obtain the normalized predicted displacement at each input coordinate, and perform inverse normalization on the normalized predicted displacement to obtain the actual displacement at each input coordinate, so as to form a three-dimensional continuous displacement field. The dataset generation module 106 is used to input the preset normalized depth and normalized spatial coordinates into the trained creep parameter network to obtain the corresponding creep parameters and generate a synthetic twin dataset superimposed with spatial heterogeneity features. The candidate deployment module 107 is used to acquire geological data of the landslide and set multiple candidate deployment configurations based on the geological data. The deployment scheme module 108 is used to perform leave-one-out cross-validation on each of the candidate deployment configurations based on the synthetic twin dataset, obtain the average determination coefficient of each candidate deployment configuration, and determine and output the candidate deployment configurations whose average determination coefficients meet the preset conditions as deployment schemes.

[0121] In some embodiments, the system further includes: The slip surface identification module 109 is used to automatically differentiate the continuous displacement field along the depth direction to obtain the strain field, and determine the slip surface position based on the concentrated zone of the strain field along the depth direction. The extrapolation prediction module 110 is used to perform extrapolation prediction along the time direction using the trained multi-pore joint physical information neural network model to obtain the extrapolated displacement prediction value.

[0122] Example 3

[0123] This embodiment provides an electronic device, including at least one control processor and a memory for communicatively connecting to the at least one control processor; Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device.

[0124] In some implementations, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0125] The non-transient software program and instructions required to implement the porous physical information neural network landslide displacement field reconstruction and aperture arrangement method of the above embodiments are stored in memory. When executed by the processor, the porous physical information neural network landslide displacement field reconstruction and aperture arrangement method of the above embodiments is executed, for example, the method described above is executed. Figure 1 Method steps S1 to S10; The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] Example 4

[0127] This embodiment provides a computer-readable storage medium storing computer-executable instructions for causing a computer to execute a porous physical information neural network method for landslide displacement field reconstruction and pore placement as described in Embodiment 1.

[0128] It should be noted that the computer-readable storage medium in this embodiment may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.

[0129] More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0130] In this embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this embodiment, the computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0131] The computer-readable signal medium may also be any computer-readable storage medium other than a computer-readable storage medium, which can send, propagate or transmit a program for use by or in connection with an instruction execution system, apparatus or device.

[0132] The computer program contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof. The aforementioned computer-readable storage medium can be used to write a computer program for performing this embodiment in one or more programming languages ​​or combinations thereof. These programming languages ​​include object-oriented programming languages—such as Python, Java, and C++—and conventional procedural programming languages—such as C or similar programming languages. The program can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0133] The above description is merely a preferred embodiment of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description, and thus all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention.

Claims

1. A porous physical information neural network method for landslide displacement field reconstruction and pore placement, characterized in that, Includes the following steps: Observational data of multiple monitoring holes deployed along the main sliding direction of the landslide are obtained, and the spatial coordinates of each monitoring hole along the main sliding direction are obtained; The observation data includes cumulative lateral displacement observation data, the observation depth corresponding to the cumulative lateral displacement observation data, and the observation time; The displacement observation data is independently normalized according to the monitoring hole to obtain the normalized displacement, and the observation depth, the observation time and the spatial coordinates are normalized with global shared parameters to obtain the normalized variables; The normalized variables include normalization depth, normalization time, and normalized spatial coordinates; Construct a multi-pore joint physical information neural network model and establish a composite loss function; The porous joint physical information neural network model includes a displacement network and a creep parameter network; Based on the normalized variables and normalized shifts, a phased training strategy is adopted to train the physical information neural network model based on the composite loss function, thereby obtaining the trained multi-pore joint physical information neural network model. Multiple input coordinates are generated within the range of the normalized variable. Each input coordinate is input into the trained displacement network to obtain the normalized predicted displacement at each input coordinate. The normalized predicted displacement is then denormalized to obtain the actual displacement at each input coordinate, thereby forming a three-dimensional continuous displacement field. The preset normalized depth and normalized spatial coordinates are input into the trained creep parameter network to obtain the corresponding creep parameters and generate a synthetic twin dataset superimposed with spatial heterogeneity features. Obtain geological data of the landslide, and set multiple candidate deployment configurations based on the geological data; Based on the synthetic twin dataset, leave-one-out cross-validation is performed on each of the candidate deployment configurations to obtain the average determination coefficient of each candidate deployment configuration. The candidate deployment configurations whose average determination coefficients meet the preset conditions are determined as deployment schemes and output.

2. The method according to claim 1, characterized in that, The method further includes: The strain field is obtained by automatically differentiating the continuous displacement field along the depth direction, and the position of the sliding surface is determined based on the concentrated zone of the strain field along the depth direction. Using the trained multi-pore joint physical information neural network model, extrapolation prediction is performed along the time direction to obtain the extrapolated displacement prediction value.

3. The method according to claim 1, characterized in that, The displacement observation data is independently normalized for each monitoring hole to obtain normalized displacement. Furthermore, the observation depth, observation time, and spatial coordinates are normalized using globally shared parameters to obtain normalized variables. This process specifically includes the following steps: Based on the minimum and maximum values ​​of the displacement observation data of each monitoring hole, the displacement observation data of the monitoring hole is linearly mapped to a preset interval to obtain the normalized displacement of the monitoring hole. The monitoring hole is the borehole where the borehole surveying instrument is located; The observation depth is normalized by using the minimum and maximum values ​​of the observation depths among all monitoring wells as normalization parameters shared across the entire region, thus obtaining the normalized depth. The observation time is normalized by using the minimum and maximum values ​​of the observation time of all monitoring wells as normalization parameters shared across the entire region, thus obtaining the normalized time. Using the minimum and maximum values ​​of the spatial coordinates of all monitoring wells as the normalization parameters shared across the entire domain, the spatial coordinates are normalized to obtain normalized spatial coordinates.

4. The method according to claim 1, characterized in that, Constructing a multi-pore joint physical information neural network model and establishing a composite loss function includes the following steps: A multi-pore joint physical information neural network model is constructed based on the physical information neural network. The displacement network takes the normalized variable as input and outputs a normalized predicted displacement. The creep parameter network takes the normalized depth and the normalized spatial coordinates as input and outputs creep parameters that vary with depth and spatial coordinates. The creep parameters are Burgers viscoelastic creep model parameters, which include initial elastic displacement parameters, steady-state creep rate parameters, transient creep amplitude parameters, and transient decay rate parameters. A data fidelity loss term is established to constrain the normalized predicted displacement to approximate the normalized displacement by penalizing the deviation between the normalized predicted displacement and the normalized displacement. A creep constitutive loss term is established. By restoring the normalized predicted displacement to physical displacement according to the displacement scale and differentiating it with respect to physical time, the physical displacement rate is obtained. The deviation between the physical displacement rate and the theoretical creep rate determined by the creep parameters is penalized to constrain the time evolution of the displacement to conform to the Burgers viscoelastic creep model. A strain smoothing loss term is established to constrain strain smoothing along the depth direction by penalizing the mean square value of the third derivative of the normalized predicted displacement with respect to the normalized depth at the placement point. A creep parameter spatial smoothing loss term is established, which constrains the continuous change of each creep parameter along the main slip direction by penalizing the first-order partial derivatives of the steady-state creep rate parameter, the transient creep amplitude parameter, and the transient decay rate parameter with respect to the normalized spatial coordinates. A displacement monotonicity loss term is established, which is used to constrain the displacement from the front edge of the landslide to the rear edge of the landslide by penalizing the square mean of the first-order partial derivative of the normalized predicted displacement with respect to the normalized spatial coordinates after one-sided rectification. The data fidelity loss term, the creep constitutive loss term, the strain smoothing loss term, the creep parameter spatial smoothing loss term, and the displacement monotonicity loss term constitute a composite loss function.

5. The method according to claim 4, characterized in that, Using the normalized variables as input, and employing a phased training strategy, the physical information neural network model is trained based on the composite loss function to obtain the trained multi-pore joint physical information neural network model. Specifically, the process includes the following steps: With minimizing the data fidelity loss term as the optimization objective, the porous joint physical information neural network model is pre-trained to obtain the pre-trained porous joint physical information neural network model. The creep constitutive loss term, the strain smoothing loss term, the creep parameter space smoothing loss term, and the displacement monotonicity loss term are modulated by using course weights that increase from zero to one with each training round. The modulated loss terms are then superimposed with the data fidelity loss term to obtain the modulated composite loss function. With minimizing the modulated composite loss function as the optimization objective, the pre-trained multi-pore joint physical information neural network model is fine-tuned to obtain the fine-tuned multi-pore joint physical information neural network model. A quasi-Newton optimizer is used to refine the fine-tuned porous joint physical information neural network model with the goal of minimizing the composite loss function, resulting in a trained porous joint physical information neural network model.

6. The method according to claim 1, characterized in that, The preset normalized depth and normalized spatial coordinates are input into the trained creep parameter network to obtain the corresponding creep parameters, and a synthetic twin dataset with superimposed spatial heterogeneity features is generated. The specific steps include: The preset normalized depth and normalized spatial coordinates are input into the trained creep parameter network to obtain the corresponding creep parameters, thereby forming a two-dimensional creep parameter field. Based on the two-dimensional creep parameter field, a reference trend for each creep parameter along the main sliding direction is set; The baseline trend decreases from the leading edge of the landslide to the trailing edge. The baseline trend corresponding to each creep parameter is multiplied by the heterogeneity multiplier to obtain the superimposed creep parameters, so that the spatial heterogeneity features are superimposed on each creep parameter; The heterogeneous multipliers include sinusoidal wave terms and local anomaly terms; Based on the preset depth shape function, a synthetic displacement field is constructed according to the superimposed creep parameters; The synthetic twin dataset is obtained by sampling at preset monitoring hole positions, depths, and time points on the synthetic displacement field.

7. The method according to claim 1, characterized in that, Based on the synthetic twin dataset, leave-one-out cross-validation is performed on each of the candidate deployment configurations to obtain the average coefficient of determination for each candidate deployment configuration. The candidate deployment configurations whose average coefficient of determination meets the preset conditions are determined as deployment schemes and output. The specific steps include: The synthetic displacement data is globally standardized using the mean and standard deviation of the synthetic displacement data of all monitoring holes in the synthetic twin dataset to obtain standard synthetic displacement data. For each monitoring hole in the candidate deployment configuration, the monitoring hole is sequentially designated as a reserved hole, and leave-one cross-validation is performed to obtain the determination coefficient of each monitoring hole in the candidate deployment configuration. The monitoring holes located at both ends in the candidate deployment configuration are determined as boundary holes, and the remaining monitoring holes are determined as internal holes. The average value of the determination coefficients corresponding to all the internal holes is taken as the average determination coefficient of the candidate deployment configuration. Leave-one-out cross-validation is performed on the current candidate deployment configuration in ascending order of the number of monitoring holes to obtain the average coefficient of determination of the internal holes of the candidate deployment configuration. Determine whether the average determination coefficient of the internal hole is less than a preset threshold. If the number is less than the required number of monitoring holes, leave-one-out cross-validation will be performed on the next candidate deployment configuration in ascending order of the number of monitoring holes. If it is not less than, then the current candidate deployment configuration is determined as the selected deployment configuration, and the number of monitoring holes corresponding to the current candidate deployment configuration and the average distance between adjacent monitoring holes are determined as the minimum number of monitoring holes and the maximum allowable distance, respectively. In the selected deployment configuration, the section where the monitoring hole with the lowest determination coefficient in the internal borehole is located is identified as a suspected abnormal section or a section that needs to be densified, and is verified in conjunction with geological data; The spacing between the monitoring holes in the reviewed sections is increased to less than half of the maximum allowable spacing, and the boundary holes on the side with smaller displacement magnitude are laid out first to form the layout scheme. The specific steps of leave-one-out cross-validation include: The multi-pore joint physical information neural network model is trained using the standard synthetic displacement data corresponding to the remaining monitoring holes (excluding the reserved holes) in the candidate layout configuration in the synthetic twin dataset. The normalized coordinates of the opening in the synthetic twin dataset are input into the trained multi-pore joint physical information neural network model to obtain the displacement prediction value corresponding to the opening. Based on the standard synthetic displacement data and displacement prediction value corresponding to the reserved hole, the determination coefficient of the reserved hole is calculated in the same standardized domain.

8. The method according to claim 2, characterized in that, The strain field is obtained by automatically differentiating the continuous displacement field along the depth direction. The position of the slip surface is determined based on the concentrated zone of the strain field along the depth direction. Then, the extrapolated displacement prediction is obtained by using the trained multi-pore joint physical information neural network model along the time direction. The specific steps include: The strain field is obtained by automatically differentiating the continuous displacement field along the depth direction. For each monitoring hole, the distribution curve of the strain field along the depth direction at each time point is calculated, and the distribution curve of the same monitoring hole at each time point is averaged along the time direction to obtain the time-averaged strain profile of the monitoring hole. Locate local maxima in the time-averaged strain profile and determine whether the strain value of the local maxima exceeds a preset strain threshold. If it exceeds the limit, the depth corresponding to the local maximum point is determined as the sliding surface position; If it does not exceed the limit, then the local maximum point is excluded; The displacement observation data of each monitoring well were divided into training and verification segments according to time sequence. The multi-well joint physical information neural network model was retrained using the displacement observation data corresponding to the training segments of all monitoring wells. Using the retrained multi-pore joint physical information neural network model, the displacement at each time point in the test section is predicted, and the extrapolated displacement prediction value at each time point in the test section is obtained. Calculate the extrapolation determination coefficient between each extrapolated displacement prediction value and the displacement observation data corresponding to the test section, and use the extrapolation determination coefficient to characterize the extrapolation accuracy.

9. A porous physical information neural network system for landslide displacement field reconstruction and pore layout, characterized in that, include: The data acquisition module is used to acquire observation data from multiple monitoring holes deployed along the main sliding direction of the landslide, and to acquire the spatial coordinates of each monitoring hole along the main sliding direction. The observation data includes cumulative lateral displacement observation data, the observation depth corresponding to the cumulative lateral displacement observation data, and the observation time; The data processing module is used to independently normalize the displacement observation data according to the monitoring hole to obtain normalized displacement, and to perform global shared parameter normalization on the observation depth, the observation time and the spatial coordinates to obtain normalized variables; The normalized variables include normalization depth, normalization time, and normalized spatial coordinates; The model building module is used to construct a multi-pore joint physical information neural network model and establish a composite loss function; The porous joint physical information neural network model includes a displacement network and a creep parameter network; The model training module is used to train the physical information neural network model based on the composite loss function according to the normalized variables and normalized displacements, using a phased training strategy, to obtain the trained multi-pore joint physical information neural network model. The displacement field reconstruction module is used to generate multiple input coordinates within the range of the normalized variable, input each input coordinate into the trained displacement network to obtain the normalized predicted displacement at each input coordinate, and perform inverse normalization on the normalized predicted displacement to obtain the actual displacement at each input coordinate, so as to form a three-dimensional continuous displacement field. The dataset generation module is used to input the preset normalized depth and normalized spatial coordinates into the trained creep parameter network to obtain the corresponding creep parameters and generate a synthetic twin dataset superimposed with spatial heterogeneity features. The candidate deployment module is used to acquire geological data of the landslide and set multiple candidate deployment configurations based on the geological data. The deployment scheme module is used to perform leave-one-out cross-validation on each of the candidate deployment configurations based on the synthetic twin dataset, obtain the average determination coefficient of each candidate deployment configuration, and determine and output the candidate deployment configurations whose average determination coefficients meet the preset conditions as deployment schemes.

10. The system according to claim 9, characterized in that, The system also includes: The slip surface identification module is used to automatically differentiate the continuous displacement field along the depth direction to obtain the strain field, and determine the slip surface position based on the concentrated zone of the strain field along the depth direction. The extrapolation prediction module is used to perform extrapolation prediction along the time direction using the trained multi-pore joint physical information neural network model to obtain the extrapolated displacement prediction value.

Images