Geological disaster hidden danger risk fine identification method and device, equipment and medium

Abstract

The application relates to a fine identification method, device, equipment and medium for geological disaster hidden danger risk. The method comprises the following steps: acquiring multi-source monitoring data and external driving data of a target area, and performing spatio-temporal alignment and fusion processing on the multi-source monitoring data to obtain a fusion deformation observation field; inputting the fusion deformation observation field and the external driving data into a pre-trained double-branch deep neural network model to obtain a first displacement prediction field, a second displacement prediction field, a stress tensor field and a pore water pressure field, and performing bidirectional coupling calibration on the double-branch deep neural network model based on a composite loss function to obtain a calibrated spatialized physical parameter field; based on the calibrated spatialized physical parameter field, generating a physical multi-quantity field by using a physical mechanism branch, and calculating a spatialized safety coefficient field based on the physical multi-quantity field to obtain a geological disaster hidden danger risk zoning map of the target area. The method can improve the accuracy and reliability of geological disaster hidden danger risk identification.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster risk identification technology, and in particular relates to a refined method, device, equipment and medium for identifying potential geological disaster risks. Background Technology

[0002] With the development of technologies in the field of geological disaster prevention and mitigation, Physical Information Neural Network (PINN) technology has emerged. By using physical equations as loss functions to apply soft constraints to the network, physical knowledge is injected into the model, providing a solution for early identification of geological disasters.

[0003] In traditional technologies, numerical simulation methods based on physical mechanisms mainly rely on accurate constitutive models and key parameters. They construct physical models to perform numerical calculations, thereby enabling the analysis and judgment of disaster risks. Data-driven methods based on deep learning, on the other hand, rely on a large amount of monitoring data and use network models to mine potential patterns in the data, thereby completing the prediction and identification of disasters.

[0004] However, the aforementioned methods rely on key parameters that are highly heterogeneous in space and difficult to obtain accurately, which makes model parameter tuning difficult and computationally time-consuming, making it difficult to meet the actual needs of large-scale, real-time monitoring and early warning. Data-driven methods based on deep learning lack effective constraints from physical mechanisms, and are prone to producing absurd predictions that violate physical laws when outside the distribution of training data or in extreme scenarios. Furthermore, the models have poor interpretability and weak generalization ability. Summary of the Invention

[0005] Therefore, it is necessary to provide a refined method, device, equipment, and medium for identifying geological hazard risks that can be co-evolved through physical mechanisms and data-driven interaction and mutual correction, in order to address the aforementioned technical problems.

[0006] Firstly, this application provides a refined method for identifying potential geological hazards, including:

[0007] The system acquires multi-source monitoring data and external driving data for the target area, and performs spatiotemporal alignment and fusion processing on the multi-source monitoring data to obtain a fused deformation observation field. The multi-source monitoring data includes area InSAR deformation data and point GNSS displacement data; the external driving data includes meteorological and hydrological data.

[0008] The fused deformation observation field and external driving data are input into a pre-trained bi-branch deep neural network model to obtain the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field; the bi-branch deep neural network model includes a data-driven branch and a physical mechanism branch based on the spatialized physical parameter field;

[0009] Based on the composite loss function, the bidirectional coupling calibration of the dual-branch deep neural network model is performed by fusing the deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field to obtain the calibrated spatialized physical parameter field.

[0010] Based on the calibrated spatialized physical parameter field, a physical multi-quantity field is generated by using the physical mechanism branch based on the fused deformation observation field and external driving data. The spatialized safety coefficient field is then calculated based on the physical multi-quantity field to obtain the geological hazard risk zoning map of the target area.

[0011] In one embodiment, spatiotemporal alignment and fusion processing is performed on multi-source monitoring data to obtain a fused deformation observation field, including:

[0012] By using a regression model based on Gaussian processes, point-based GNSS displacement data is spatially interpolated onto the same spatial grid as the area-based InSAR deformation data to obtain the GNSS interpolated displacement field.

[0013] The residuals of the planar InSAR deformation data and the GNSS interpolated displacement field at the overlapping spatial locations are calculated, and a spatialized data reliability weight map is generated based on the spatial statistical characteristics of the residuals.

[0014] The displacement prediction field based on the physical mechanism branch uses a spatialized data reliability weight map to weight and fuse the GNSS interpolated displacement field and the planar InSAR deformation data to obtain the fused deformation observation field.

[0015] In one embodiment, based on a composite loss function, a bidirectional coupling calibration of the dual-branch deep neural network model is performed by fusing the deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field to obtain the calibrated spatialized physical parameter field, including:

[0016] Based on the fusion of the deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field, a composite loss function is obtained. The composite loss function includes a data fitting loss term, a physical equation residual loss term, and an inter-branch consistency loss term. The data fitting loss term represents the difference between the first displacement prediction field and the fusion deformation observation field. The physical equation residual loss term represents the degree to which the stress tensor field and the pore water pressure field violate the preset seepage-stress coupling control equation. The inter-branch consistency loss term represents the difference between the first displacement prediction field and the second displacement prediction field.

[0017] The weights and spatialized physical parameter fields of the physical mechanism branch are frozen, and gradient descent iterations are performed a predetermined number of times with the goal of minimizing the composite loss function to obtain the updated weights of the data-driven branch.

[0018] Freeze the weights of the updated data-driven branch, unlock the spatialized physical parameter field, and perform a predetermined number of gradient descent iterations based on the composite loss function to obtain the calibrated spatialized physical parameter field.

[0019] In one embodiment, the weights of the updated data-driven branch are frozen, the weights of the physical mechanism branch and the spatialized physical parameter field are unlocked, and a predetermined number of gradient descent iterations are performed based on the composite loss function to obtain the calibrated spatialized physical parameter field, including:

[0020] Based on the inter-branch consistency loss term, the difference between the displacement components of the first displacement prediction field and the second displacement prediction field at each spatial location is extracted to obtain the prediction difference field tensor.

[0021] The predicted difference field tensor and the spatialized physical parameter field are input into the pre-trained parameter calibration model to obtain the parameter adjustment field.

[0022] Based on the parameter adjustment field, and the gradients calculated for the spatialized physical parameter field by combining the residual loss term and the inter-branch consistency loss term of the physical equation, the spatialized physical parameter field is updated according to the preset mixing ratio to obtain the calibrated spatialized physical parameter field.

[0023] In one embodiment, the physical equation residual loss term represents the degree to which the stress tensor field and pore water pressure field violate the preset seepage-stress coupling control equations, including:

[0024] The seepage-stress coupling control equation is obtained through the following formula:

[0025]

[0026]

[0027] in, The effective stress tensor is calculated by the elastoplastic constitutive model from the stress tensor field output by the physical mechanism branch. Biot coefficient; The pore water pressure field is output from the physical mechanism branch; Density; It is the acceleration due to gravity; For the permeability tensor of the spatialized physical parameter field; The dynamic viscosity of water; This is the water storage coefficient; This is the second displacement prediction field output by the physical mechanism branch; For externally driven data source and sink items.

[0028] In one embodiment, based on the calibrated spatialized physical parameter field, a physical multi-quantity field is generated using a physical mechanism branch based on the fused deformation observation field and external driving data. A spatialized safety factor field is then calculated based on the physical multi-quantity field to obtain a geological hazard risk zoning map of the target area, including:

[0029] By inputting the fused deformation observation field and external driving data into the physical mechanism branch based on the calibrated spatialized physical parameter field, the corrected stress tensor field and the corrected pore water pressure field are obtained.

[0030] The strength parameters corresponding to each spatial location in the target region are read from the calibrated spatialized physical parameter field. Combined with the corrected stress tensor field and the corrected pore water pressure field, the local safety factor of each spatial location is calculated to obtain the spatialized safety factor field. The strength parameters include cohesion and friction angle.

[0031] Based on preset threshold mapping rules, spatial statistics are performed on the spatialized safety coefficient field and mapped and transformed to obtain a geological hazard risk zoning map.

[0032] In one embodiment, the method further includes:

[0033] Obtain scenario-driven data sequences for a future preset time period; the scenario-driven data includes rainfall forecast data for the future preset time period;

[0034] Using physical multi-quantum fields as initial conditions, the future scenario-driven data sequence is input as a source and sink term into the physical mechanism branch based on the calibrated spatialized physical parameter field to obtain the displacement prediction field, stress tensor field and pore water pressure field for a future preset time period.

[0035] The spatialized safety factor field corresponding to the future time period is calculated based on the displacement prediction field, stress tensor field and pore water pressure field of the future preset time period, and the safety factor time series transformation field is obtained by combining the spatialized safety factor field of the current time period.

[0036] Secondly, this application also provides a refined identification device for potential geological disaster risks, including:

[0037] The data monitoring module is used to acquire multi-source monitoring data and external driving data of the target area, and to perform spatiotemporal alignment and fusion processing on the multi-source monitoring data to obtain a fused deformation observation field; the multi-source monitoring data includes areal InSAR deformation data and point GNSS displacement data; the external driving data includes meteorological and hydrological data.

[0038] The disaster identification module is used to input the fused deformation observation field and external driving data into a pre-trained dual-branch deep neural network model to obtain the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field; the dual-branch deep neural network model includes a data-driven branch and a physical mechanism branch based on the spatialized physical parameter field;

[0039] The physical correction module is used to perform bidirectional coupling calibration of the dual-branch deep neural network model based on the composite loss function by fusing the deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field, so as to obtain the calibrated spatialized physical parameter field.

[0040] The risk quantification module is used to generate a physical multi-quantity field based on the calibrated spatialized physical parameter field, using the physical mechanism branch to generate a physical multi-quantity field based on the fused deformation observation field and external driving data, and calculate the spatialized safety coefficient field based on the physical multi-quantity field to obtain a geological hazard risk zoning map of the target area.

[0041] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the refined identification method for any of the above-mentioned geological hazard risk hazards.

[0042] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the refined identification method for any of the above-mentioned geological hazard risks.

[0043] The aforementioned refined identification methods, devices, equipment, and media for geological hazard risks utilize a highly reliable fused deformation observation field formed by integrating InSAR and GNSS multi-source observation data. Simultaneously, external driving data such as meteorological and hydrological data are incorporated to construct a complete model input. The fused field and driving data are input into a pre-trained dual-branch deep neural network model. The data-driven branch rapidly generates the first displacement prediction field, while the physical mechanism branch, based on spatialized physical parameter fields, outputs a second displacement prediction field, stress tensor field, and pore water pressure field that satisfy physical laws. A composite loss function is used to jointly optimize the data fitting error, physical equation residuals, and the consistency between the two branch predictions. This collaborative training mechanism drives the parameter calibration module to dynamically adjust the spatialized physical parameter field in the physical mechanism branch, enabling real-time reverse calibration of the physical model parameters using observational data. Based on the calibrated high-confidence physical parameter field and physical mechanism branch, a spatialized safety factor field conforming to mechanical principles is calculated, and a geological hazard risk zoning map is automatically generated. This achieves deep bidirectional coupling and real-time feedback between physical mechanisms and data-driven approaches. Under the constraints of stringent physical laws, this mechanism fully utilizes observational data to improve model accuracy and generalization ability, while simultaneously using data feedback to correct physical understanding in real time. This significantly enhances the reliability, interpretability, and timeliness of dynamic risk identification results. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating the refined identification method for geological hazard risks according to the present invention.

[0046] Figure 2 This is a flowchart illustrating the steps of step S103.

[0047] Figure 3 This is a flowchart illustrating the steps of step S203.

[0048] Figure 4 This is a structural diagram of the geological disaster hazard risk identification device of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] In one embodiment, such as Figure 1 As shown, a refined method for identifying potential geological disaster risks is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S101. Acquire multi-source monitoring data and external driving data of the target area, and perform spatiotemporal alignment and fusion processing on the multi-source monitoring data to obtain a fused deformation observation field; the multi-source monitoring data includes area InSAR deformation data and point GNSS displacement data; the external driving data includes meteorological and hydrological data.

[0052] To illustrate, before conducting geological hazard risk identification work, a systematic collection of multi-source monitoring data and external driving data is completed for the target area. Among them, the planar InSAR (Synthetic Aperture Radar) deformation data of the multi-source monitoring data is obtained by relying on time-series synthetic aperture radar interferometry technology. By interferometric processing of radar images acquired at different times in the same area, the subtle deformation information of the surface of the target area can be inverted. In the data acquisition process, a series of operations such as image registration, interferogram generation, flat ground effect removal, phase unwrapping, and orbit refinement need to be performed on the original radar images in sequence to reduce the interference of factors such as terrain undulation and orbit error on the deformation inversion results. Finally, planar deformation data in the form of multi-dimensional arrays are obtained. This data can characterize the line-of-sight displacement of each pixel in the target area in the time series and can be further decomposed into vertical and east-west displacement components, thus fully reflecting the three-dimensional deformation characteristics of the surface. Point-based GNSS (Global Navigation Satellite System) displacement data is acquired through GNSS monitoring stations deployed within the target area. These stations, relying on satellite positioning technology, can collect real-time three-dimensional coordinate information of their locations. By performing differential calculations on coordinate information collected at different times, the three-dimensional displacement data of the stations over time can be obtained. This data possesses extremely high measurement accuracy and can provide precise control point references for areal deformation data. External driving data includes meteorological and hydrological data, primarily including parameters such as rainfall intensity, soil moisture content, and pore water pressure in the target area—key factors affecting the stability of soil and rock masses. Rainfall intensity data can be automatically collected by meteorological stations deployed within the area, while soil moisture content and pore water pressure data are acquired using multi-functional sensors deployed at typical geological hazard sites, ensuring that the data accurately reflects the hydrological and meteorological conditions of the target area.

[0053] Furthermore, after data acquisition, spatiotemporal alignment processing needs to be performed on multi-source monitoring data. This involves constructing a unified spatiotemporal reference, using the universal transverse Mercator projection coordinate system for the spatial reference and Coordinated Universal Time (UTC) for the time reference, thereby eliminating differences in spatiotemporal references between different data sources. For areal InSAR deformation data, atmospheric delay correction and phase unwrapping error correction operations are required. Atmospheric delay correction can be achieved through meteorological model simulation or coherent target selection, thereby reducing deformation measurement errors caused by atmospheric refraction effects. Phase unwrapping error correction is achieved by introducing external high-precision topographic data or using multi-path unwrapping algorithms to solve the phase jump problem that occurs during phase unwrapping, thus improving the reliability of deformation data. For point-based GNSS displacement data, since its data format is discrete point data, it cannot be directly matched and calculated with area-based InSAR deformation data. Therefore, Kriging interpolation is used for spatial upsampling. Based on the principle of spatial autocorrelation, Kriging interpolation constructs a variogram model to describe the spatial distribution characteristics of the GNSS displacement data, thereby interpolating the discrete point data into area-based data with the same spatial resolution as the InSAR deformation data, generating the interpolated GNSS displacement field. Optionally, a residual field is calculated between the area-based InSAR deformation data and the interpolated GNSS displacement field. The residual field is calculated as the displacement difference between the area-based InSAR deformation data and the interpolated GNSS displacement field at corresponding pixels and corresponding times. This residual field can intuitively reflect the consistency between the two types of deformation data, thereby assessing the measurement uncertainty of different data sources. Based on the distribution characteristics of the residual field, a weighted fusion algorithm is used to fuse the two types of deformation data. The weight coefficients of the weighted fusion algorithm are determined by the size of the residual. Regions with smaller residuals are given higher weights, thereby highlighting the reference value of high-precision data. After fusion processing, a fused deformation observation field with both areal coverage characteristics and high-precision characteristics can be obtained. This observation field can provide high-quality data support for subsequent model training and risk identification.

[0054] S102. Input the fused deformation observation field and external driving data into the pre-trained dual-branch deep neural network model to obtain the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field; the dual-branch deep neural network model includes a data-driven branch and a physical mechanism branch based on the spatialized physical parameter field.

[0055] Indicatively, the dual-branch deep neural network model consists of a data-driven branch and a physical mechanism branch based on a spatialized physical parameter field. The two branches share the same input data interface, and the input data is a fusion of deformation observation field, spatial coordinate information, time information, and external driving data, thereby realizing the collaborative input of multi-source information. The data-driven branch employs a hybrid network structure consisting of an attention-gated recurrent unit and a convolutional long short-term memory network. The gated recurrent unit effectively captures the temporal dependency features of the data, achieving selective memorization and forgetting of historical temporal information through dynamic adjustment of reset and update gates, thus avoiding the gradient vanishing or gradient explosion problems present in traditional recurrent neural networks. The introduction of the attention mechanism allows the model to automatically focus on key spatiotemporal features that contribute significantly to displacement prediction, enhancing the model's ability to capture important information. The convolutional long short-term memory network has the ability to extract spatial features, effectively mining and fusing the spatial distribution patterns contained in the deformation observation field by replacing the fully connected operations in the traditional long short-term memory network with convolutional operations. The hybrid network structure achieves deep fusion of spatiotemporal features by using the output of the attention-gated recurrent unit as the input of the convolutional long short-term memory network. The core function of this branch is to learn the nonlinear mapping relationship between the input data and the displacement field, and its output is the first displacement prediction field, which is based on the three-dimensional displacement distribution of the Earth's surface obtained through the data-driven approach.

[0056] The model's pre-training process involves constructing a large-scale historical geological disaster monitoring dataset. This dataset encompasses multi-source monitoring data and externally driven data across different geological types and disaster development stages. Specifically, the data-driven branch is pre-trained separately, using a mean squared error loss function to measure the difference between the predicted and actual displacement fields. The network weights are iteratively updated using a backpropagation algorithm until the model converges, thus enabling the data-driven branch to possess preliminary displacement prediction capabilities. The physical mechanism branch employs a multilayer perceptron as its basic network structure. In addition to fusing deformation observation fields, spatial coordinate information, temporal information, and externally driven data, this branch also introduces a spatialized physical parameter field. This spatialized physical parameter field consists of trainable parameter vectors containing key parameters characterizing the physical and mechanical properties of soil and rock, such as permeability coefficient, elastic modulus, cohesion, and internal friction angle. The initial values ​​of the parameters are determined based on geological surveys and laboratory test data of the target area, ensuring that the initial distribution of the parameters conforms to the actual geological conditions of the target area. The output of the physical mechanism branch includes not only the second displacement prediction field, but also the stress tensor field and the pore water pressure field. There are clear physical relationships between the physical field quantities. The stress tensor field and the second displacement prediction field are linked through the elastoplastic constitutive equation, while the pore water pressure field is linked with parameters such as rainfall intensity in the external driving data through the seepage equation. The core design concept of this branch is to ensure that the output physical field quantities satisfy the basic physical laws of geological disaster evolution.

[0057] Optionally, during the pre-training phase, a physical loss function is constructed based on existing physical equations. The output of the model is constrained by the physical loss function, ensuring that the physical mechanism branch always follows physical laws during training. The physical mechanism branch is pre-trained separately using historical datasets. After both branches have completed their individual pre-training, they are then jointly pre-trained. A consistency loss function is constructed to constrain the displacement fields output by the two branches to remain consistent, thereby achieving collaborative optimization of the two branches.

[0058] S103. Based on the composite loss function, the bidirectional coupling calibration of the dual-branch deep neural network model is performed by fusing the deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field to obtain the calibrated spatialized physical parameter field.

[0059] Furthermore, the expression for the composite loss function is as follows: ,in, , , These are the weighting coefficients for data loss, physical loss, and consistency loss, respectively. The values ​​of these weighting coefficients can be dynamically adjusted according to the changes in loss during model training to balance the contribution of different loss terms to model optimization. The specific process of bidirectional coupling calibration is divided into two stages: forward propagation and backward propagation. In the forward propagation stage, the fused deformation observation field and external driving data are simultaneously input into the two branches of the bi-branch deep neural network model to obtain the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field. The backpropagation phase calculates the composite loss function and then divides into two parts: regular gradient update and parameter calibration module update. Regular gradient update calculates the partial derivatives of the composite loss function with respect to the weights of the two branches and uses an Adam optimizer to iteratively update the network weights. The parameter calibration module is a lightweight network specifically designed to optimize the spatialized physical parameter field. This module takes the residuals between the first and second displacement prediction fields as input and outputs the increment of the spatialized physical parameter field. During the training of the parameter calibration module, the weights of the main network of the dual-branch deep neural network model are fixed, and the weights of the parameter calibration module are iteratively optimized using a meta-learning algorithm. This ensures that the updated spatialized physical parameter field effectively reduces consistency loss while maintaining the physical loss at a low level. The final update formula for the spatialized physical parameter field is as follows: ,in, For the spatialized physical parameter field before the update, For the updated spatialized physical parameter field, and These are the learning rates corresponding to the physical loss and the updated consistency loss, respectively. The partial derivative of the physical loss with respect to the spatialized physical parameter field is given. This is the partial derivative of the updated consistency loss with respect to the spatialized physical parameter field. By iteratively proceeding through forward and backward propagation until the composite loss function converges, the calibrated spatialized physical parameter field can be obtained. This parameter field not only conforms to the measured data characteristics of the target area but also satisfies the physical laws governing the evolution of geological hazards.

[0060] S104. Based on the calibrated spatialized physical parameter field, a physical multi-quantity field is generated using the physical mechanism branch according to the fused deformation observation field and external driving data. Based on the physical multi-quantity field, the spatialized safety factor field is calculated to obtain the geological hazard risk zoning map of the target area.

[0061] Based on the calibrated spatialized physical parameter field, a physical multi-quantity field is generated using the physical mechanism branch based on the fused deformation observation field and external driving data. A spatialized safety factor field is then calculated based on this physical multi-quantity field, resulting in a geological hazard risk zoning map of the target area. After completing the bidirectional coupling calibration of the dual-branch deep neural network model, the calibrated spatialized physical parameter field can accurately characterize the spatial distribution of the physical and mechanical properties of the soil and rock mass in the target area. At this point, the physical mechanism branch needs to generate a physical multi-quantity field that combines data accuracy and physical rationality. Specifically, the fused deformation observation field and external driving data of the target area are input into the physical mechanism branch carrying the calibrated spatialized physical parameter field. Based on the calibrated parameter field and combined with the fluid-structure interaction control equations, the physical mechanism branch outputs a physical multi-quantity field including the displacement field, stress tensor field, and pore water pressure field. Compared to the physical field before calibration, this physical multi-quantity field has higher consistency with the measured data and fully follows the physical laws of geological hazard evolution, providing reliable physical field support for geological hazard risk identification. The process of calculating the spatialized safety factor field based on physical multi-quantity fields employs the limit equilibrium method. By comparing the shear strength and shear stress on the shear surface of the soil and rock mass, the stability state of the soil and rock mass is determined. The calculation of the spatialized safety factor field requires discretizing the target region into several calculation units. For each calculation unit, the stress tensor field and pore water pressure field data from the physical multi-quantity fields are extracted. Specifically, the stress tensor field is decomposed into normal stress and shear stress. To calculate the normal stress components on the shear plane of the element, the shear stress... To calculate the shear stress components on the shear surface of the calculation element, the local safety factor of the element is then calculated based on the limit equilibrium method. The formula for calculating the local safety factor is as follows: ,in, This is to determine the cohesion in the calibrated spatialized physical parameter field, which reflects the bonding ability between soil and rock particles. The internal friction angle in the calibrated spatialized physical parameter field, which reflects the frictional characteristics between soil and rock particles, The pore water pressure in the physical multi-field can partially offset the normal stress, reducing the shear strength of the soil and rock mass. In the calculation of the local safety factor, if... This indicates that the shear stress of the soil and rock mass within the calculation unit exceeds its shear strength, indicating that the soil and rock mass is in an unstable state and belongs to a high-risk hazard zone; if This indicates that the shear stress of the soil and rock mass within the calculation unit is close to its shear strength, and the soil and rock mass is in a critical stability state, belonging to a medium-risk hidden danger zone; if This indicates that the shear strength of the soil and rock mass within the calculation unit is much greater than the shear stress, indicating that the soil and rock mass is in a stable state and belongs to a low-risk hazard zone. After completing the local safety factor calculation for all calculation units, the spatial safety factor field of the target area can be obtained. This safety factor field can intuitively reflect the spatial distribution characteristics of geological hazard risks in the target area. To present the risk distribution of the target area more intuitively, the spatial safety factor field needs to be overlaid with the geographic information system base map. The geographic information system base map contains basic geographic information such as topography, strata lithology, and distribution of human engineering activities in the target area. Through overlay processing, the distribution characteristics of the safety factor field can be correlated with the actual geographic environment. Based on this, a geological hazard risk zoning map is generated using visualization technology. During the visualization process, color gradients are used to mark different risk levels, with high-risk hazard areas marked in red, medium-risk hazard areas marked in yellow, and low-risk hazard areas marked in green. At the same time, the specific location, development range, and risk level of the hazard points are marked on the risk zoning map. Combined with the evolution trend of physical multi-quantity fields, the development and change trend of geological hazard risks in the target area in the future period is predicted.

[0062] In the aforementioned refined identification method for geological hazard risks, multi-source monitoring data, including isometric InSAR deformation data and point GNSS displacement data of the target area, as well as external driving data such as meteorological and hydrological data, are first acquired. The multi-source monitoring data undergoes spatiotemporal alignment and fusion processing to obtain a fused deformation observation field. Then, the fused deformation observation field and the external driving data are input into a pre-trained dual-branch deep neural network model containing a data-driven branch and a physical mechanism branch based on spatialized physical parameter fields. This yields a first displacement prediction field, a second displacement prediction field, a stress tensor field, and a pore water pressure field. Subsequently, based on a composite loss function, the aforementioned fused deformation observation field, prediction fields, and physical fields are used to... A bidirectional coupling calibration is performed on a dual-branch deep neural network model to obtain the calibrated spatialized physical parameter field. Finally, based on this calibrated spatialized physical parameter field, a physical multi-quantity field is generated by combining the deformation observation field and external driving data through the physical mechanism branch. The spatialized safety factor field is then calculated based on the physical multi-quantity field, ultimately obtaining a geological hazard risk zoning map of the target area. This achieves deep coupling between the model and the data, improves the accuracy and reliability of geological hazard risk identification, solves the problem of fusing multi-source heterogeneous data and multi-physics field models, balances the computational efficiency of the identification process with the physical interpretability of the results, and enhances the model's adaptability and generalization ability.

[0063] In one embodiment, spatiotemporal alignment and fusion processing is performed on multi-source monitoring data to obtain a fused deformation observation field, including:

[0064] S11. Using a regression model based on Gaussian processes, the point-like GNSS displacement data is spatially interpolated onto the same spatial grid as the area-like InSAR deformation data to obtain the GNSS interpolated displacement field.

[0065] Indicatively, the Gaussian process-based regression model transforms the interpolation problem into a function prediction problem based on probability distribution. This is achieved by assuming that the spatial distribution of displacement data follows a Gaussian process, meaning that the displacement observations at any finite number of spatial points jointly follow a multivariate normal distribution. The model's input variables are the spatial coordinates of point GNSS monitoring stations, and the output variables are the three-dimensional displacement data at those coordinates. The model's prior distribution is defined by the mean function and the covariance function. A zero-mean function is typically used because the spatial trend of point GNSS displacement data can be indirectly captured by the covariance function's characterization of spatial correlation. The covariance function chosen is the squared exponential covariance function, which effectively describes the smooth variation characteristics of spatial data; its expression is: ,in and The first The and the first Spatial coordinate vectors of GNSS monitoring stations The signal variance is used to characterize the fluctuation amplitude of the displacement data itself. This is a length scale parameter used to describe the attenuation range of spatial correlation in displacement data. A larger value indicates that spatial correlation decays more slowly. This is the noise variance, used to characterize the measurement noise in GNSS displacement observation data. For the Kronecker delta function, when hour ,otherwise The function introduces noise terms with diagonal elements into the covariance matrix to distinguish between signal and noise.

[0066] Model training solves for the hyperparameters in the covariance function through maximum likelihood estimation, i.e. , and Specifically, a log-likelihood function is constructed based on all point-based GNSS displacement observation data. This function, with hyperparameters as independent variables, reflects the likelihood of the observed data under the current hyperparameters. The log-likelihood function is iteratively optimized using a gradient descent algorithm to find the hyperparameter combination that maximizes the likelihood function value, thereby determining the prior distribution of the model. After model training is completed, displacement prediction can be performed for each spatial grid point in the orthographic InSAR deformation data. For any spatial grid point to be interpolated... The posterior distribution of the predicted displacement values ​​also follows a normal distribution. The posterior mean is the GNSS interpolated displacement value for that grid point, and the posterior variance can be used to assess the uncertainty of the interpolation results. The formula for calculating the posterior mean is... ,in Points to be interpolated With all GNSS monitoring stations The covariance vector between them, whose elements are , The covariance matrix between GNSS monitoring stations has the following elements: , This represents the point-based GNSS displacement observation data vector. By performing the above prediction calculations on all spatial grid points of the planar InSAR deformation data one by one, a GNSS interpolated displacement field with the same spatial resolution as the planar InSAR deformation data can be obtained. This displacement field retains the high-precision characteristics of GNSS data and also has planar coverage capability.

[0067] S12. Calculate the residuals of the planar InSAR deformation data and the GNSS interpolated displacement field at the overlapping spatial locations, and generate a spatialized data reliability weight map based on the spatial statistical characteristics of the residuals.

[0068] For example, the formula for calculating the residual is: ,in Spatial location The residual value at that point, Spatial location The area of ​​InSAR deformation data values, Spatial location The GNSS interpolated displacement field value at the location. During the calculation, it is necessary to ensure that the residual is solved only for the overlapping spatial location of the planar InSAR deformation data and the GNSS interpolated displacement field to avoid calculation errors caused by spatial mismatch.

[0069] Furthermore, after obtaining the residuals for all overlapping spatial locations, a spatialized data reliability weight map is generated by analyzing the spatial statistical characteristics of the residuals. The spatial statistical characteristics of the residuals mainly include the spatial variance, spatial autocorrelation coefficient, and variogram, reflecting the spatial distribution pattern of the residuals and thus revealing the reliability differences between the two types of data in different spatial regions. The formula for calculating the spatial variance is as follows: ,in Let V be the spatial variance of the residuals. This represents the total number of overlapping spatial locations. For the first The residual values ​​at each overlapping position. The spatial variance represents the mean of all residual values. A larger spatial variance indicates poorer consistency between the two types of data in that region, and lower data reliability. The spatial autocorrelation coefficient describes the spatial correlation of the residuals, and its calculation formula is... ,in, The distance is Spatial autocorrelation coefficient over time Spatial location and The covariance of the residuals and the magnitude of the spatial autocorrelation coefficient can reflect the spatial clustering characteristics of the residuals. If the spatial autocorrelation coefficient of the residuals in a certain region is large and positive, it indicates that the residuals in that region have significant spatial clustering and may have systematic errors.

[0070] The variogram is used to describe how the residuals change with spatial distance, and its calculation formula is: ,in The distance is The value of the variogram at time, The distance is The number of residual pairs. Based on the above spatial statistical characteristics, an exponential function is used to map the spatial statistical characteristics of the residuals into data reliability weights. The formula for calculating the weights is: ,in, Spatial location Reliability weight at the location, Spatial location Local spatial variance of the residuals The formula uses the ratio of local spatial variance to global spatial variance as the exponent to map the weights from 0 to 1. The smaller the local spatial variance of the residuals, the closer the weight is to 1, indicating higher data reliability in that region; conversely, the closer the weight is to 0, the lower the data reliability. Optionally, to obtain a continuous spatialized data reliability weight map, spatial interpolation is needed to interpolate the weight values ​​at discrete overlapping locations onto the entire spatial grid of the areal InSAR deformation data. The interpolation process can also use a Gaussian regression model to ensure the spatial smoothness and rationality of the weight map. The resulting spatialized data reliability weight map can accurately depict the data reliability distribution in different spatial regions.

[0071] S13. Displacement prediction field based on physical mechanism branch: The GNSS interpolated displacement field and the planar InSAR deformation data are weighted and fused using the spatialized data reliability weight map to obtain the fused deformation observation field.

[0072] By introducing the displacement prediction field from the physical mechanism branch as a constraint, the fusion result is ensured to possess both data accuracy and conformity to the physical laws of geological hazard evolution. The displacement prediction field from the physical mechanism branch is derived from the physical and mechanical parameters of the soil and rock mass and the fluid-structure interaction control equations. It can reflect the physical evolution trend of surface displacement and avoid outliers in the fusion result that violate physical laws. The calculation formula for the fusion process is as follows: ,in Spatial location The fusion deformation observations at the location, Spatial location Reliability weight at the location, Spatial location GNSS interpolated displacement field values ​​at the location, Spatial location The area of ​​InSAR deformation data values, These are the physical constraint weighting coefficients, used to adjust the influence of the physical mechanism branch displacement prediction field on the fusion result. Spatial location Displacement prediction values ​​at the branch of physical mechanism. Spatial location The mean of the GNSS interpolated displacement field value and the area InSAR deformation data value.

[0073] Physical constraint weighting coefficient The determination of the result needs to be achieved through cross-validation. Specifically, the measured data from some overlapping spatial locations are selected as the validation set, and the remaining data are used as the training set. This is achieved by adjusting... The value of is used to calculate different The mean square error between the fusion result and the measured data of the validation set is selected, and the result corresponding to the minimum mean square error is chosen. As the optimal physical constraint weighting coefficient, it balances the contributions of data-driven and physical constraints. During the fusion calculation, the above formula needs to be calculated for each spatial grid point to ensure the spatial continuity of the fusion result. For boundary regions or regions with extremely low data reliability, the role of the physical constraint term is particularly important, as it can effectively suppress the interference of noisy data on the fusion result and avoid the occurrence of local outliers.

[0074] After completing the fusion calculation of all spatial grid points, the fused deformation observation field needs to be spatially smoothed to eliminate local fluctuations caused by data heterogeneity. The spatial smoothing process employs a Gaussian filtering method, the core of which is to correct the fused value of each grid point through a weighted average. The kernel function expression for the Gaussian filter is as follows: ,in The Gaussian filter kernel function is located in space. The value at that location, This is the standard deviation of the filter, used to control the smoothness of the filter. These are the coordinates of the grid points to be smoothed. Through Gaussian filtering, the spatial distribution of the fused deformation observation field is smoother, more accurately reflecting the overall trend of surface deformation.

[0075] In one embodiment, such as Figure 2 As shown, based on the composite loss function, the bidirectional coupling calibration of the dual-branch deep neural network model is performed by fusing the deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field, resulting in the calibrated spatialized physical parameter field, including:

[0076] S201. Based on the fusion deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field, a composite loss function is obtained. The composite loss function includes a data fitting loss term, a physical equation residual loss term, and an inter-branch consistency loss term. The data fitting loss term represents the difference between the first displacement prediction field and the fusion deformation observation field. The physical equation residual loss term represents the degree to which the stress tensor field and the pore water pressure field violate the preset seepage-stress coupling control equation. The inter-branch consistency loss term represents the difference between the first displacement prediction field and the second displacement prediction field.

[0077] Data loss The calculation formula is: In the formula The first displacement prediction field output by the data-driven branch. To fuse the deformation observation fields, the essence of data loss is the mean square error between the first displacement prediction field and the fused deformation observation field. Its function is to constrain the output of the data-driven branch to be consistent with the measured data, ensuring that the model's predictions are supported by data. Physical loss. It is constructed based on the fluid-structure interaction control equations of geological hazard evolution. The fluid-structure interaction control equations adopt the Biot consolidation theory framework and include two core equations: the equilibrium equation and the seepage continuity equation. The expression of the equilibrium equation is as follows: ,in, The effective stress tensor is related to the strain of the soil mass through the elastoplastic constitutive equation. Biot coefficient, Pore ​​water pressure, The density of the rock and soil mass Let be the gravitational acceleration vector. The equilibrium equations reflect the equilibrium state of the stress field in the soil and rock mass; the expression for the seepage continuity equation is: ,in, Let be the permeability tensor. The dynamic viscosity of water, The water storage coefficient, The second displacement prediction field is output from the physical mechanism branch. The source and sink terms are related to rainfall intensity in the external driving data. The seepage continuity equation reflects the mass conservation law of the seepage field in the soil and rock mass. The formula for calculating physical losses is as follows: ,in, and To adapt weighting coefficients and balance the magnitude difference between the equilibrium equation and the seepage continuity equation, the calculation of physical losses requires the use of automatic differentiation techniques. This involves calculating the spatial and temporal partial derivatives of the physical field quantities output by the physical mechanism branch, substituting the results into the governing equations to obtain the equation residuals. These residuals then measure the degree to which the outputs of the physical mechanism branch satisfy the physical laws. Consistency loss. The calculation formula is: ,in, The second displacement prediction field output by the physical mechanism branch is used to force the displacement field output by the data-driven branch to be consistent with that output by the physical mechanism branch, thus building a two-way feedback bridge between data information and physical information.

[0078] S202. Freeze the weights and spatialized physical parameter fields of the physical mechanism branch, and perform a predetermined number of gradient descent iterations with the goal of minimizing the composite loss function to obtain the updated weights of the data-driven branch.

[0079] The weights and spatialized physical parameter field of the physical mechanism branch are frozen, and gradient descent iterations are performed a predetermined number of times with the goal of minimizing the composite loss function, thereby achieving directional updates of the data-driven branch weights. The "freeze" operation fixes the initial values ​​of the network connection weights and spatialized physical parameter field of the physical mechanism branch, preventing them from participating in gradient calculations and parameter updates during this iteration phase. They serve only as fixed physical constraints, ensuring that the optimization of the data-driven branch always operates within the framework of physical laws. The specific process of gradient descent iteration is as follows: the fused deformation observation field and external driving data are input into the two-branch model; the values ​​of each loss term and the total value of the composite loss function are obtained through forward propagation; based on automatic differentiation, the gradient of the composite loss function with respect to all trainable weights of the data-driven branch is calculated, and the gradient direction reflects the trend of weight adjustment's contribution to reducing the loss; an adaptive optimizer is used to update the weights of the data-driven branch based on the gradient information. The weight update amount in each iteration is determined by both the learning rate and the gradient value. The learning rate can be set to a dynamic decay mode to avoid gradient oscillations in the early stages of iteration and slow convergence in later stages. The predetermined number of iterations should be reasonably set according to the model complexity and data scale, so that the data-driven branch can fully learn the spatiotemporal characteristics of the measured data under the current physical constraints, so that the first displacement prediction field can not only fit the fused deformation observation field, but also move closer to the output of the physical mechanism branch through the consistency loss term.

[0080] S203. Freeze the weights of the updated data-driven branch, unlock the spatialized physical parameter field, and perform a predetermined number of gradient descent iterations based on the composite loss function to obtain the calibrated spatialized physical parameter field.

[0081] Furthermore, the reverse switching of "freeze-unfreeze" enables the directional calibration of physical parameters using data information, which is the core of bidirectional coupled calibration. "Freezing" the updated data-driven branch weights means they remain stable during this iteration phase, serving as a reliable carrier of data features and providing a fixed reference standard for physical parameter optimization. "Unfreezing" the spatialized physical parameter field allows it to adjust its values ​​based on the feedback of the loss function during gradient descent, achieving adaptation between physical parameters and data features. During the iteration process, the forward propagation phase still calculates each loss term based on the fused deformation observation field and external driving data, focusing on the changes in the physical equation residual loss term and the inter-branch consistency loss term. The backpropagation phase only calculates the gradient of the composite loss function with respect to the spatialized physical parameter field. This gradient simultaneously contains constraint information from the physical equation residuals and calibration information from the inter-branch prediction differences, ensuring that parameter adjustments do not violate physical laws while allowing the parameter field to adapt to the high-precision predictions of the data-driven branches. The update formula for the spatialized physical parameter field is as follows: ,in For the parameter field before the update, For the updated parameter field, This represents the learning rate for this stage. This represents the gradient of the composite loss function with respect to the parameter field. The predetermined number of iterations must be judged based on the convergence of the composite loss function to ensure that the calibrated spatialized physical parameter field not only conforms to the actual physical and mechanical properties of the soil and rock mass, but also generates a second displacement prediction field consistent with the data-driven branch prediction through the operation of the physical mechanism branch, thus achieving two-way collaboration between data and physics.

[0082] In one embodiment, such as Figure 3 As shown, the weights of the updated data-driven branch are frozen, the weights of the physical mechanism branch and the spatialized physical parameter field are unlocked, and a predetermined number of gradient descent iterations are performed based on the composite loss function to obtain the calibrated spatialized physical parameter field, including:

[0083] S301. Based on the inter-branch consistency loss term, extract the difference between the displacement components of the first displacement prediction field and the second displacement prediction field at each spatial location to obtain the prediction difference field tensor.

[0084] Specifically, for each spatial grid point within the target area, the displacement component differences between the first and second displacement prediction fields in three-dimensional space (x, y, z directions) are calculated. Each spatial location corresponds to a set of three-dimensional vectors containing differences in the x, y, and z directions. These dispersed spatial location difference vectors are organized according to their corresponding spatial coordinates (x, y) or (x, y, z) to form a prediction difference field tensor whose dimensions match the spatialized physical parameter field. Each element of this tensor contains both spatial location information and clearly corresponds to a specific direction of displacement difference, accurately characterizing the prediction deviation distribution of the two branches in different spatial regions, providing a direct basis for subsequent parameter calibration.

[0085] S302. Input the predicted difference field tensor and the spatialized physical parameter field into the pre-trained parameter calibration model to obtain the parameter adjustment field.

[0086] The parameter calibration model employs a lightweight multilayer perceptron (MLP) structure. Its core function is to learn the mapping relationship between "predicted difference" and "parameter adjustment." The model's pre-training process is completed based on a sample set constructed from historical monitoring data. The input to the sample set is the simulated predicted difference field tensor and the corresponding initial spatialized physical parameter field. The output is the validated optimal parameter adjustment amount. The model is iteratively trained using a gradient descent algorithm until convergence, ensuring its ability to accurately output the direction and magnitude of parameter adjustment based on the difference distribution. During the input process, the predicted difference field tensor and the spatialized physical parameter field must maintain strict alignment in the spatiotemporal dimensions, meaning that the difference vector at each spatial location is matched one-to-one with the corresponding physical parameter set. The model outputs a parameter adjustment field through feature fusion and nonlinear mapping of the input data. Each element at each spatial location in this adjustment field corresponds to a specific adjustment value for each physical parameter in the spatialized physical parameter field, such as cohesion, internal friction angle, and permeability coefficient.

[0087] S303. Based on the parameter adjustment field and the gradients calculated for the spatialized physical parameter field by combining the residual loss term and the inter-branch consistency loss term of the physical equation, the spatialized physical parameter field is updated according to the preset mixing ratio to obtain the calibrated spatialized physical parameter field.

[0088] Indicative, the core update formula is: ,in For the calibrated spatialized physical parameter field, For the spatialized physical parameter field before the update, The preset mixing ratio coefficient, To adjust the parameter field output by the parameter calibration model. This represents the gradient of the residual loss term in the physical equations with respect to the spatialized physical parameter field. This represents the gradient of the inter-branch consistency loss term with respect to the spatialized physical parameter field. and This is the balance coefficient between the two gradient terms. It is a preset mixing ratio coefficient. The contribution weights used to adjust the direct adjustment of the parameter calibration model and the gradient-guided adjustment of the loss function are in the range of [0,1]. These weights can be determined through cross-validation with a small number of samples, ensuring that the data calibration needs reflected by the predicted difference field are fully utilized without deviating from the constraints of physical laws. During the update process, the physical parameters at each spatial location are calculated and adjusted synchronously according to the above formula, ensuring the spatial continuity and overall rationality of the parameter field. The resulting calibrated spatialized physical parameter field accurately adapts to the data prediction differences while maintaining consistency with the seepage-stress coupling control equations.

[0089] In one embodiment, the physical equation residual loss term represents the degree to which the stress tensor field and pore water pressure field violate the preset seepage-stress coupling control equations, including:

[0090] The seepage-stress coupling control equation is obtained through the following formula:

[0091]

[0092]

[0093] in, The effective stress tensor is calculated by the elastoplastic constitutive model from the stress tensor field output by the physical mechanism branch. Biot coefficient; The pore water pressure field is output from the physical mechanism branch; Density; It is the acceleration due to gravity; For the permeability tensor of the spatialized physical parameter field; The dynamic viscosity of water; This is the water storage coefficient; This is the second displacement prediction field output by the physical mechanism branch; For externally driven data source and sink items.

[0094] Indicative, The effective stress tensor is not directly output, but is calculated by the stress tensor field output from the physical mechanism branch through the elastoplastic constitutive model. The effective stress tensor is a key physical quantity that reflects the actual stress state between soil particles and directly determines the deformation and strength characteristics of soil. Biot coefficient is a core parameter characterizing the coupling effect between pore pressure and effective stress in porous media. Its value is related to the porosity, particle compression modulus, and soil compression modulus of the rock and soil, and is used to quantify the degree of influence of changes in pore water pressure on effective stress. The pore water pressure field is a direct output of the physical mechanism branch, reflecting the pressure state of groundwater in the pores of rock and soil. Its distribution and changes directly affect the effective stress of the rock and soil. Density is the inherent physical property parameter of rock and soil, representing the mass of rock and soil per unit volume. This is the gravitational acceleration vector, directed vertically downwards, and is the fundamental cause of the self-weight stress in soil and rock masses; It represents the divergence of the effective stress tensor, describing the rate of change of the distribution of effective stress in the spatial coordinate system; This represents the gradient of the pore water pressure field, reflecting the spatial distribution differences of pore water pressure. The physical essence of this equation is the static equilibrium condition of the soil and rock mass, namely, the spatial variation effect of the effective stress, the force generated by the pore water pressure gradient, and the self-weight of the soil and rock mass cancel each other out, ensuring that the stress field satisfies the static equilibrium law.

[0095] The permeability coefficient tensor is a key component of the spatialized physical parameter field, characterizing the ability of soil and rock to allow groundwater to seep through. The tensor form can reflect the anisotropic characteristics of the permeability coefficient, that is, the difference in permeability in different spatial directions. The dynamic viscosity of water is a physical property parameter of groundwater, which characterizes the viscous resistance between water molecules. Its value is adjusted with changes in water temperature. The water storage coefficient describes the ability of a porous medium to store or release groundwater when the pore water pressure changes, reflecting the elastic water storage characteristics of the rock and soil mass. The second displacement prediction field is the three-dimensional displacement distribution of the soil and rock mass output by the physical mechanism branch, which directly reflects the deformation state of the soil and rock mass. As source and sink terms, they are closely related to external driving data and mainly cover external factors such as rainfall infiltration recharge, artificial groundwater extraction, and surface runoff discharge. Positive values ​​indicate groundwater recharge, and negative values ​​indicate groundwater discharge. Divergence, representing the seepage velocity, describes the degree of convergence or divergence of groundwater seepage in space; This represents the rate of change of the seepage velocity divergence over time; The time rate of change of the pore water pressure field reflects the dynamic evolution characteristics of the pore water pressure. The divergence of the second displacement prediction field corresponds to the volumetric strain change of the soil and rock mass. The time rate of change of volumetric strain characterizes the dynamic process of volumetric deformation of soil and rock. The physical essence of this equation is the mass conservation condition of groundwater, that is, the balance between the changes in water volume caused by seepage, the changes in water storage caused by changes in pore water pressure, and the changes in pore volume caused by the volumetric deformation of soil and rock, ensuring that the seepage field conforms to the law of mass conservation.

[0096] In one embodiment, based on the calibrated spatialized physical parameter field, a physical multi-quantity field is generated using a physical mechanism branch based on the fused deformation observation field and external driving data. A spatialized safety factor field is then calculated based on the physical multi-quantity field to obtain a geological hazard risk zoning map of the target area, including:

[0097] S21. Input the fused deformation observation field and external driving data into the physical mechanism branch based on the calibrated spatialized physical parameter field to obtain the corrected stress tensor field and the corrected pore water pressure field.

[0098] The calibrated spatialized physical parameter field fully integrates data characteristics and physical laws, accurately characterizing the actual physical and mechanical properties of the soil and rock mass in the target area, providing precise parameter support for the calculation of the physical mechanism branch. The fusion of deformation observation fields as measured data provides realistic feedback on surface deformation for model calculations, while external driving data reflects external environmental factors influencing the evolution of geological hazards. The synergistic input of these three ensures that the calculation of the physical mechanism branch not only closely matches actual monitoring conditions but also follows the physical laws of fluid-structure interaction. The physical mechanism branch substitutes the input data and the calibrated parameter field into the seepage-stress coupling control equations. Through automatic differentiation and numerical solution, it outputs a corrected stress tensor field and a corrected pore water pressure field. Compared to the uncalibrated fields, these two corrected physical fields show higher consistency with the measured data and stronger physical rationality, laying a reliable foundation for subsequent calculations of local safety factors.

[0099] S22. Read the strength parameters corresponding to each spatial location in the target area from the calibrated spatialized physical parameter field, and calculate the local safety factor of each spatial location by combining the corrected stress tensor field and the corrected pore water pressure field to obtain the spatialized safety factor field; the strength parameters include cohesion and friction angle.

[0100] From the calibrated spatialized physical parameter field, the corresponding strength parameters are read according to the coordinate index of each spatial location. These strength parameters specifically include cohesion and internal friction angle, which are the core indicators determining the shear strength of soil and rock masses. Their spatial distribution has been precisely optimized through two-way coupling calibration. Simultaneously, the normal stress and shear stress at each spatial location are extracted from the corrected stress tensor field. The normal stress is the component of the stress tensor in the direction normal to the shear plane, and the shear stress is the component of the stress tensor in the direction tangent to the shear plane. The pore water pressure values ​​at the corresponding spatial locations are extracted from the corrected pore water pressure field. Based on the principle of limit equilibrium, the local safety factor is calculated using the formula... Complete the calculation of the local safety factor for each spatial location, where For cohesion, It is the internal friction angle. For normal stress, Pore ​​water pressure, This represents shear stress. During the calculation, it is necessary to ensure that the strength parameters, stress components, and pore water pressure values ​​at each spatial location strictly correspond. By performing point-by-point calculations, a set of local safety factors covering the entire spatial range of the target area is obtained, which is then organized into a spatialized safety factor field. This field intuitively reflects the stability differences of the soil and rock mass at different locations in the target area.

[0101] S23. Based on the preset threshold mapping rules, the spatial safety coefficient field is spatially statistically analyzed and mapped to obtain a geological hazard risk zoning map.

[0102] For example, a local safety factor less than 1 corresponds to a high-risk level, between 1 and 1.2 corresponds to a medium-risk level, and greater than 1.2 corresponds to a low-risk level. The spatial statistical analysis process mainly includes statistically analyzing the spatial range, distribution density, and continuous distribution areas corresponding to each risk level to ensure the scientific validity and rationality of the risk zoning. The mapping transformation converts the safety factor value of each spatial location in the spatialized safety factor field into a corresponding risk level identifier according to preset rules. Then, using spatial visualization technology, the risk level identifiers are overlaid on the geographic information system base map in different colors or textures to form a geological hazard risk zoning map. This zoning map clearly presents the spatial distribution pattern of high, medium, and low-risk hazard areas within the target area, clearly defining the specific location and extent of the hazards, providing intuitive and accurate technical support for disaster prevention and mitigation decision-making.

[0103] In one embodiment, the method further includes:

[0104] S31. Obtain the scenario-driven data sequence for a future preset time period; the scenario-driven data includes rainfall forecast data for the future preset time period.

[0105] Indicatively, the core component of scenario-driven data is rainfall forecast data for a predetermined future period. Supplementary driving data, such as future temperature changes and groundwater extraction plans, which may influence the evolution of geological hazards, can be added as needed. The acquired data must be organized according to a unified time base and spatial coordinate system to form a time-series scenario-driven data sequence. Each data point in the sequence includes a corresponding timestamp, spatial coverage, and specific driving parameter values, such as rainfall intensity and cumulative rainfall.

[0106] S32. Using the physical multi-quantity field as the initial condition, the future scenario-driven data sequence is input as the source and sink term into the physical mechanism branch based on the calibrated spatialized physical parameter field to obtain the displacement prediction field, stress tensor field and pore water pressure field for the future preset time period.

[0107] Furthermore, the physical multi-quantum field contains complete physical state information such as the displacement field, stress tensor field, and pore water pressure field at the current moment, which can truly reflect the current mechanical and hydrological state of the soil and rock mass in the target area, providing an accurate initial state benchmark for future scenario prediction. The processed future scenario-driven data sequence is substituted as the source-sink term Q in the seepage-stress coupling control equation, where rainfall forecast data is transformed into specific values ​​of the source-sink term through the rainfall infiltration model, positively representing the groundwater recharge intensity. Subsequently, the initial boundary conditions and scenario-driven data sequence are jointly input into the physical mechanism branch equipped with the calibrated spatialized physical parameter field. Based on the calibrated physical and mechanical parameters of the soil and rock mass, the model performs time-series numerical extrapolation by solving the fluid-structure interaction control equation. Each time step uses the output result of the previous time step as the input condition for the next time step, iteratively calculating the displacement prediction field, stress tensor field, and pore water pressure field corresponding to each time node within the preset future period, forming a complete future physical field time series, ensuring that the prediction results not only conform to the laws of physical evolution but also respond to changes driven by future scenarios.

[0108] S33. Calculate the spatialized safety factor field corresponding to the future time based on the displacement prediction field, stress tensor field and pore water pressure field of the future preset time period, and obtain the safety factor time series transformation field by combining the spatialized safety factor field of the current time period.

[0109] For the displacement prediction field, stress tensor field, and pore water pressure field at each time node within a preset future time period, a local safety factor calculation method consistent with that for the current time period is adopted. This method is based on the limit equilibrium principle, combined with strength parameters such as cohesion and internal friction angle in the calibrated spatialized physical parameter field, and calculated using the formula... The future local safety factor is calculated point by point at each spatial location, and then a spatialized safety factor field corresponding to each future moment is formed. The spatialized safety factor fields of each future moment are spatiotemporally aligned with the spatialized safety factor fields acquired in the current time period to construct a temporal transformation field of safety factors. This temporal transformation field not only includes the spatial distribution of safety factors at each moment, but also intuitively presents the dynamic evolution trend of geological disaster risks at different spatial locations in the target area by calculating indicators such as the difference and rate of change of safety factors between adjacent moments. For example, it shows the expansion or contraction of high-risk areas, the continuous decrease or rebound of safety factors, etc., providing dynamic technical support for early warning and prevention decisions of future geological disaster hazards.

[0110] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0111] Based on the same inventive concept, this application also provides a device for finely identifying geological hazard risks, which is used to implement the fine-grained identification method for geological hazard risk as described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the device for finely identifying geological hazard risks provided below can be found in the limitations of the fine-grained identification method for geological hazard risks described above, and will not be repeated here.

[0112] In one exemplary embodiment, such as Figure 4 As shown, a sophisticated device for identifying potential geological hazards is provided, comprising:

[0113] The data monitoring module 401 is used to acquire multi-source monitoring data and external driving data of the target area, and to perform spatiotemporal alignment and fusion processing on the multi-source monitoring data to obtain a fused deformation observation field; the multi-source monitoring data includes areal InSAR deformation data and point GNSS displacement data; the external driving data includes meteorological and hydrological data.

[0114] The disaster identification module 402 is used to input the fused deformation observation field and external driving data into a pre-trained dual-branch deep neural network model to obtain the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field; the dual-branch deep neural network model includes a data-driven branch and a physical mechanism branch based on the spatialized physical parameter field;

[0115] The physical correction module 403 is used to perform bidirectional coupling calibration of the dual-branch deep neural network model based on the composite loss function by fusing the deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field, so as to obtain the calibrated spatialized physical parameter field.

[0116] The risk quantification module 404 is used to generate a physical multi-quantity field based on the calibrated spatialized physical parameter field, using the physical mechanism branch to generate a physical multi-quantity field according to the fused deformation observation field and external driving data, and calculate the spatialized safety coefficient field based on the physical multi-quantity field to obtain the geological hazard risk zoning map of the target area.

[0117] In one embodiment, the data monitoring module 401 is further configured to:

[0118] By using a regression model based on Gaussian processes, point-based GNSS displacement data is spatially interpolated onto the same spatial grid as the area-based InSAR deformation data to obtain the GNSS interpolated displacement field.

[0119] The residuals of the planar InSAR deformation data and the GNSS interpolated displacement field at the overlapping spatial locations are calculated, and a spatialized data reliability weight map is generated based on the spatial statistical characteristics of the residuals.

[0120] The displacement prediction field based on the physical mechanism branch uses a spatialized data reliability weight map to weight and fuse the GNSS interpolated displacement field and the planar InSAR deformation data to obtain the fused deformation observation field.

[0121] In one embodiment, the physical correction module 403 is further configured to:

[0122] Based on the fusion of the deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field, a composite loss function is obtained. The composite loss function includes a data fitting loss term, a physical equation residual loss term, and an inter-branch consistency loss term. The data fitting loss term represents the difference between the first displacement prediction field and the fusion deformation observation field. The physical equation residual loss term represents the degree to which the stress tensor field and the pore water pressure field violate the preset seepage-stress coupling control equation. The inter-branch consistency loss term represents the difference between the first displacement prediction field and the second displacement prediction field.

[0123] The weights and spatialized physical parameter fields of the physical mechanism branch are frozen, and gradient descent iterations are performed a predetermined number of times with the goal of minimizing the composite loss function to obtain the updated weights of the data-driven branch.

[0124] Freeze the weights of the updated data-driven branch, unlock the spatialized physical parameter field, and perform a predetermined number of gradient descent iterations based on the composite loss function to obtain the calibrated spatialized physical parameter field.

[0125] In one embodiment, a parameter correction module is also included, for:

[0126] Based on the inter-branch consistency loss term, the difference between the displacement components of the first displacement prediction field and the second displacement prediction field at each spatial location is extracted to obtain the prediction difference field tensor.

[0127] The predicted difference field tensor and the spatialized physical parameter field are input into the pre-trained parameter calibration model to obtain the parameter adjustment field.

[0128] Based on the parameter adjustment field, and the gradients calculated for the spatialized physical parameter field by combining the residual loss term and the inter-branch consistency loss term of the physical equation, the spatialized physical parameter field is updated according to the preset mixing ratio to obtain the calibrated spatialized physical parameter field.

[0129] In one embodiment, the risk quantification module 404 is further configured to:

[0130] By inputting the fused deformation observation field and external driving data into the physical mechanism branch based on the calibrated spatialized physical parameter field, the corrected stress tensor field and the corrected pore water pressure field are obtained.

[0131] The strength parameters corresponding to each spatial location in the target region are read from the calibrated spatialized physical parameter field. Combined with the corrected stress tensor field and the corrected pore water pressure field, the local safety factor of each spatial location is calculated to obtain the spatialized safety factor field. The strength parameters include cohesion and friction angle.

[0132] Based on preset threshold mapping rules, spatial statistics are performed on the spatialized safety coefficient field and mapped and transformed to obtain a geological hazard risk zoning map.

[0133] In one embodiment, an evolution prediction module is also included, for:

[0134] Obtain scenario-driven data sequences for a future preset time period; the scenario-driven data includes rainfall forecast data for the future preset time period;

[0135] Using physical multi-quantum fields as initial conditions, the future scenario-driven data sequence is input as a source and sink term into the physical mechanism branch based on the calibrated spatialized physical parameter field to obtain the displacement prediction field, stress tensor field and pore water pressure field for a future preset time period.

[0136] The spatialized safety factor field corresponding to the future time period is calculated based on the displacement prediction field, stress tensor field and pore water pressure field of the future preset time period, and the safety factor time series transformation field is obtained by combining the spatialized safety factor field of the current time period.

[0137] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0138] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0139] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0140] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for fine identification of geological disaster hidden danger risk, characterized in that, The method includes: The system acquires multi-source monitoring data and external driving data for the target area, and performs spatiotemporal alignment and fusion processing on the multi-source monitoring data to obtain a fused deformation observation field. The multi-source monitoring data includes area InSAR deformation data and point GNSS displacement data. The external driving data includes meteorological and hydrological data. The fused deformation observation field and the external driving data are input into a pre-trained dual-branch deep neural network model to obtain a first displacement prediction field, a second displacement prediction field, a stress tensor field, and a pore water pressure field; the dual-branch deep neural network model includes a data-driven branch and a physical mechanism branch based on a spatialized physical parameter field; Based on the fused deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field, a composite loss function is obtained. The composite loss function includes a data fitting loss term, a physical equation residual loss term, and an inter-branch consistency loss term. The data fitting loss term represents the difference between the first displacement prediction field and the fused deformation observation field. The physical equation residual loss term represents the degree to which the stress tensor field and the pore water pressure field violate the preset seepage-stress coupling control equation. The inter-branch consistency loss term represents the difference between the first displacement prediction field and the second displacement prediction field. Freeze the weights of the physical mechanism branch and the spatialized physical parameter field, and perform a predetermined number of gradient descent iterations with the goal of minimizing the composite loss function to obtain the updated weights of the data-driven branch; Based on the inter-branch consistency loss term, the difference between the displacement components of the first displacement prediction field and the second displacement prediction field at each spatial location is extracted to obtain the prediction difference field tensor. The predicted difference field tensor and the spatialized physical parameter field are input into a pre-trained parameter calibration model to obtain the parameter adjustment field. Based on the parameter adjustment field, and the gradients calculated on the spatialized physical parameter field by combining the residual loss term of the physical equation and the consistency loss term between branches, the spatialized physical parameter field is updated according to a preset mixing ratio to obtain the calibrated spatialized physical parameter field. Based on the calibrated spatialized physical parameter field, a physical multi-quantity field is generated using the physical mechanism branch according to the fused deformation observation field and the external driving data. Based on the physical multi-quantity field, a spatialized safety factor field is calculated to obtain a geological hazard risk zoning map of the target area.

2. The method of claim 1, wherein, The process of spatiotemporal alignment and fusion of the multi-source monitoring data to obtain a fused deformation observation field includes: The point GNSS displacement data is spatially interpolated onto the same spatial grid as the area InSAR deformation data using a regression model based on Gaussian processes to obtain the GNSS interpolated displacement field. Calculate the residual between the planar InSAR deformation data and the GNSS interpolated displacement field at the overlapping spatial location, and generate a spatialized data reliability weight map based on the spatial statistical characteristics of the residual; Based on the displacement prediction field of the physical mechanism branch, the GNSS interpolated displacement field and the planar InSAR deformation data are weighted and fused using the spatialized data reliability weight map to obtain the fused deformation observation field.

3. The method of claim 1, wherein, The residual loss term of the physical equation represents the degree to which the stress tensor field and the pore water pressure field violate the preset seepage-stress coupling control equation, including: The seepage-stress coupling control equation is obtained through the following formula: ； ； in, The effective stress tensor is calculated by the elastoplastic constitutive model for the stress tensor field output by the physical mechanism branch. Biot coefficient; The pore water pressure field output by the branch of the physical mechanism; Density; It is the acceleration due to gravity; Let be the permeability coefficient tensor of the spatialized physical parameter field; The dynamic viscosity of water; This is the water storage coefficient; This is the second displacement prediction field output by the physical mechanism branch; For externally driven data source and sink items.

4. The method according to claim 1, characterized in that, The process involves generating a physical multi-quantity field based on the calibrated spatialized physical parameter field, utilizing the physical mechanism branch according to the fused deformation observation field and the external driving data, and calculating a spatialized safety factor field based on the physical multi-quantity field to obtain a geological hazard risk zoning map of the target area, including: The fused deformation observation field and the external driving data are input into the physical mechanism branch based on the calibrated spatialized physical parameter field to obtain the corrected stress tensor field and the corrected pore water pressure field. The strength parameters corresponding to each spatial location in the target region are read from the calibrated spatialized physical parameter field. Combined with the corrected stress tensor field and the corrected pore water pressure field, the local safety factor of each spatial location is calculated to obtain the spatialized safety factor field. The strength parameters include cohesion and friction angle. Based on preset threshold mapping rules, the spatialized safety coefficient field is spatially statistically analyzed and mapped to obtain the geological hazard risk zoning map.

5. The method according to claim 1, characterized in that, The method further includes: Obtain a scenario-driven data sequence for a future preset time period; the scenario-driven data includes rainfall forecast data for the future preset time period; Using the physical multi-quantum field as the initial condition, the scenario-driven data sequence is input as the source and sink term into the physical mechanism branch based on the calibrated spatialized physical parameter field to obtain the displacement prediction field, stress tensor field and pore water pressure field for a future preset time period. The spatialized safety factor field corresponding to the future time period is calculated based on the displacement prediction field, stress tensor field and pore water pressure field of the future preset time period, and the safety factor time sequence transformation field is obtained by combining the spatialized safety factor field of the current time period.

6. A sophisticated device for identifying potential geological disaster risks, characterized in that, The device includes: The data monitoring module is used to acquire multi-source monitoring data and external driving data of the target area, and to perform spatiotemporal alignment and fusion processing on the multi-source monitoring data to obtain a fused deformation observation field; the multi-source monitoring data includes area InSAR deformation data and point GNSS displacement data; the external driving data includes meteorological and hydrological data. The disaster identification module is used to input the fused deformation observation field and the external driving data into a pre-trained dual-branch deep neural network model to obtain a first displacement prediction field, a second displacement prediction field, a stress tensor field, and a pore water pressure field; the dual-branch deep neural network model includes a data-driven branch and a physical mechanism branch based on a spatialized physical parameter field; The physical correction module is used to obtain a composite loss function based on the fused deformation observation field, the first displacement prediction field, the second displacement prediction field, the stress tensor field, and the pore water pressure field. The composite loss function includes a data fitting loss term, a physical equation residual loss term, and an inter-branch consistency loss term. The data fitting loss term represents the difference between the first displacement prediction field and the fused deformation observation field. The physical equation residual loss term represents the degree to which the stress tensor field and the pore water pressure field violate a preset seepage-stress coupling control equation. The inter-branch consistency loss term represents the difference between the first displacement prediction field and the second displacement prediction field. The physics correction module is also used to freeze the weights of the physical mechanism branch and the spatialized physical parameter field, and perform a predetermined number of gradient descent iterations with the goal of minimizing the composite loss function to obtain the updated weights of the data-driven branch. The parameter correction module is used to extract the difference between the displacement components of the first displacement prediction field and the second displacement prediction field at each spatial location based on the inter-branch consistency loss term, and obtain the prediction difference field tensor. The parameter correction module is also used to input the predicted difference field tensor and the spatialized physical parameter field into the pre-trained parameter calibration model to obtain the parameter adjustment field; The parameter correction module is also used to adjust the quantity field based on the parameters, and to update the spatialized physical parameter field according to a preset mixing ratio by combining the gradients calculated by the physical equation residual loss term and the inter-branch consistency loss term for the spatialized physical parameter field, so as to obtain the calibrated spatialized physical parameter field. The risk quantification module is used to generate a physical multi-quantity field based on the calibrated spatialized physical parameter field, using the physical mechanism branch according to the fused deformation observation field and the external driving data, and calculate the spatialized safety coefficient field based on the physical multi-quantity field to obtain a geological hazard risk zoning map of the target area.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

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

Images