A tailings dam load-mechanical parameter collaborative inversion method and system based on CR-InSAR time series and GA-RBF

Abstract

The application provides a tailings dam load-mechanical parameter collaborative inversion method and system based on CR-InSAR time series and GA-RBF, and is applied to the technical field of data processing. Relying on unmanned aerial vehicle LiDAR and drilling data, a three-dimensional geological model is constructed by adopting NURBS, the dam area grid is encrypted, the initial stress field is generated through the Mohr-Coulomb constitutive model and layered stacking, and invalid parameters are screened out in combination with physical constraints. Latin hypercube sampling is used to make samples, and after batch simulation, the LOS displacement is converted through the projection matrix, and the GA optimized RBF proxy model is built. During operation, the load is updated according to the dam height increment, the dam body time series deformation is extracted through CR-InSAR, and the soil mechanical parameters are input into the model to invert. The load and parameter are returned to the numerical model, the safety factor is solved by the strength reduction method, and the national standard is warned; the model is updated by the new data increment, the parameter interval is contracted, and relying on the physical prior-data posterior dual driving system, the dynamic safety research and judgment of the tailings dam is completed.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for the collaborative inversion of load-mechanical parameters of tailings dam body based on CR-InSAR time series and GA-RBF. Background Technology

[0002] Tailings dams are major hazard sources used to store tailings during mining operations. The stability of their dams directly affects the safety of life and property of people downstream and the ecological environment. As mining operations continue, upstream tailings dams typically rise several meters per year. During operation, the load conditions and material mechanical properties of the dam are in a dynamic evolution process, placing extremely high demands on the timeliness and accuracy of stability analysis methods.

[0003] During the construction of tailings dam surcharge, the physical and mechanical parameters of tailings (such as elastic modulus and unit weight) often evolve over time. At the same time, the surcharge load usually exhibits dynamic changes over time. These time-varying characteristics directly affect the soil displacement field, stress field distribution, and engineering safety during the surcharge process.

[0004] With the development of remote sensing monitoring technology, CR-InSAR (corner reflector synthetic aperture radar interferometry) technology has been widely used for surface displacement monitoring of tailings dam bodies, enabling the acquisition of large-scale, high-precision, and long-term dam deformation data. However, most existing technologies only use this type of displacement data for deformation anomaly identification, failing to fully explore the evolution information of dam mechanical parameters contained in the time-series displacement field. Meanwhile, traditional parameter inversion methods rely on iterative numerical simulation calculations, resulting in high computational costs and low efficiency, making it difficult to support the dynamic update requirements during operation. Although some studies have attempted to introduce surrogate models to improve inversion efficiency, traditional radial basis function (RBF) neural networks suffer from limitations such as parameter selection dependence on experience and susceptibility to local optima. Furthermore, while existing research has achieved tailings dam parameter inversion and load updates, most studies conduct these two processes independently, failing to achieve coordinated dynamic updates during operation. Additionally, traditional inversion methods have low computational efficiency, making it difficult to meet the needs of real-time analysis and failing to fully utilize CR-InSAR time-series monitoring data for monitoring-driven parameter correction.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of this application, a method for the joint inversion of load-mechanical parameters of tailings dam bodies based on CR-InSAR time series and GA-RBF is provided, including: collecting UAV LiDAR topographic and borehole stratigraphic data; constructing a three-dimensional geological model of the dam area based on NURBS surfaces; refining the mesh in key areas of the dam body; generating an initial stress field using a Mohr-Coulomb constitutive model and a layered stacking equilibrium method; delineating the joint feasible region of load and mechanical parameters; adding physical constraints to eliminate invalid solutions; generating load-parameter samples using Latin hypercube sampling; obtaining three-dimensional displacements through batch numerical simulation; converting the displacements to radar LOS displacements using a projection matrix; and constructing a Gaussian kernel RBF network, using the stacking height and LOS displacements as inputs, and uniformly encoding the entire network using a genetic algorithm. The hyperparameters are optimized globally using leave-one-out cross-validation error as the fitness, and adaptive mutation is employed to train the GA-RBF surrogate model. The measured dam height is used to dynamically update the model's load boundary. CR-InSAR is used to process satellite imagery, filtering out noise and extracting the time-series LOS displacement of the dam's corner reflectors. Real-time dam height and measured displacement are input into the GA-RBF surrogate model to invert current soil mechanical parameters. Updated loads and inverted parameters are fed back to the numerical model, and the safety factor is calculated using the strength reduction method, with early warning levels based on national standard thresholds. New monitoring data is used to update the surrogate model using incremental learning, narrowing the parameter search range for the next period to ±10%. A dual-drive framework based on physical priors and data posteriors is used to complete the working condition verification and optimization, generating dynamic monitoring and early warning information for the tailings dam.

[0008] Another aspect of this application is a tailings dam load-mechanical parameter co-inversion system based on CR-InSAR time series and GA-RBF, used to execute executable instructions to implement the above-mentioned tailings dam load-mechanical parameter co-inversion method based on CR-InSAR time series and GA-RBF.

[0009] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute, by executing the executable instructions, a method for the collaborative inversion of load-mechanical parameters of tailings dam body based on CR-InSAR time series and GA-RBF as described above.

[0010] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for the coordinated inversion of load-mechanical parameters of tailings dam body based on CR-InSAR time series and GA-RBF.

[0011] The beneficial technical effects of this application are as follows: It integrates LiDAR topographic and borehole data, constructs a dense three-dimensional geological model using NURBS, generates an initial stress field using a Mohr-Coulomb constitutive model and a layered stacking method, and eliminates invalid parameters using physical constraints; it generates samples using Latin hypercube sampling, transforms the LOS displacement through a projection matrix after batch simulation, and globally optimizes the Gaussian kernel RBF using a genetic algorithm to construct a high-precision surrogate model. During runtime, it updates the load boundary based on real-time stacking height, extracts the temporal deformation of corner reflectors using CR-InSAR noise reduction, and synchronously inverts a complete set of soil mechanical parameters into the input model; it transmits the parameters and loads back to the numerical model, calculates the safety factor using the strength reduction method, and implements graded early warning according to national standards; it incrementally updates the model with newly added monitoring data, shrinks the parameter search range to ±10%, and achieves dynamic stability assessment and long-term monitoring of tailings dams based on both physical priors and measured data.

[0012] This application aims to employ NURBS modeling and differentiated mesh refinement to recreate the true stress field through layered stacking, and to eliminate unreasonable parameters through physical constraints, thereby improving the realism of numerical simulations. It achieves coordinated dynamic inversion of dam load and soil parameters, overcoming the shortcomings of separate calculations that cannot match real-time dam conditions. Furthermore, incremental learning and interval adaptive shrinkage reduce repetitive simulation calculations, significantly lowering iteration costs while maintaining update efficiency. The integration of CR-InSAR's large-scale temporal deformation with strength reduction stability calculations forms a complete closed loop from monitoring and inversion to graded early warning, improving the timeliness of tailings dam risk prediction.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0014] Figure 1 This document shows a flowchart of a tailings dam load-mechanical parameter co-inversion method based on CR-InSAR time series and GA-RBF provided in an embodiment of this application. Figure 2 A flowchart of the overall process for collaborative inversion of soil mechanical parameters and stability early warning of tailings dam body based on CR-InSAR time series monitoring and GA-RBF algorithm; Figure 3 This is a schematic diagram of a three-layer RBF neural network topology; Figure 4 A complete iterative flowchart of a genetic algorithm for optimizing hyperparameters of RBF networks; Figure 5 A flowchart illustrating the complete construction and testing process of a GA-RBF joint optimization proxy model; Figure 6 This is a drone-based optical aerial photographic base map of a tailings dam site; Figure 7 A schematic diagram of the elevation of a three-dimensional numerical mesh model of a tailings dam; Figure 8 This is a line graph comparing the settlement values ​​along the LOS axis at various effective monitoring points. Detailed Implementation

[0015] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] In one implementation, Figure 1 The schematic diagram illustrates a process flow diagram of a tailings dam load-mechanical parameter co-inversion method based on CR-InSAR time series and GA-RBF according to an embodiment of this application.

[0017] S101: Collect LiDAR terrain and borehole stratigraphic data from UAVs; construct a three-dimensional geological model of the reservoir area based on NURBS surfaces; refine the mesh in the key areas of the dam body; generate the initial stress field using the Mohr-Coulomb constitutive model and the layered stacking equilibrium method; delineate the joint feasible domain of load and mechanical parameters; and add physical constraints to eliminate invalid solutions.

[0018] In one implementation, the system combines the overall application scenario of three-dimensional numerical simulation for tailings engineering, synchronously matching the fusion requirements of topographic and borehole exploration data, the NURBS surface reconstruction technology of the strata, the differentiated grid partitioning rules of the dam body, the method for generating initial stress of phased surcharge, and the joint screening and constraint of load and soil mechanical parameters. It determines the collection of multi-source exploration data, the construction of three-dimensional geometric entities, the solution of initial stress of surcharge, and the feasible domain of parameters. The four core processing procedures of constraint control are sequentially linked to form a complete set of execution systems for standardized three-dimensional geological models of the reservoir area. For example, at the construction site of a Level 4 tailings dam project, a UAV LiDAR full-domain surface point cloud and 12 sets of vertical geological borehole layer records were simultaneously acquired. The point cloud coordinates used the WGS84 geodetic coordinate system, while the borehole data used the local engineering independent coordinate system. The two types of data references and layer storage formats were completely inconsistent. Relying on the multi-source data collection and processing process built into this system, the coordinates were uniformly registered and the stratigraphic layer elevations were aligned. After the coordinate transformation was completed, the NURBS stratigraphic surface reconstruction, the differentiated structured mesh generation of the dam body, the iterative balancing of self-weight stress in the layered construction, and the selection of multiple constraint parameter intervals for soil and rock were performed in a continuous process. Each process was seamlessly connected, which can avoid distortion problems such as stratigraphic interface misalignment, dam slope mesh density imbalance, and mismatch between the initial stress field and the on-site construction history caused by segmented independent modeling.

[0019] In the NURBS surface reconstruction stage, a non-uniform rational B-spline surface algorithm was used to stitch together the layered contour lines of each borehole. The sampling interval of the surface control points was set to 2 meters, matching the reservoir area topographic data with a resolution of 3 meters. This can accurately restore the natural undulation of the three strata of tailings sand, silty clay, and bedrock. In the meshing stage, the size was set in different regions. The side length of the mesh unit in the dam crest, upstream and downstream dam slopes, and dam foundation slip risk zone was 2 meters, while the side length of the mesh unit in the gentle base of the reservoir area was 8 meters. All of them adopted hexahedral dominant structured meshes to meet the requirements of FLAC3D finite difference continuous mechanics calculation.

[0020] The initial stress generation follows the phased damming equilibrium method, with the grid cells activated layer by layer according to the actual mine's 12-phase annual damming records. The height of each damming layer is 2.0m. After each layer is activated, a self-weight stress equilibrium iteration is performed. The criterion for convergence of the iteration is that the unbalanced force ratio is less than 10. -5 The model uniformly adopts the Mohr-Coulomb constitutive model to describe the shear failure characteristics of soil. After obtaining the basic values ​​of soil weight, elastic modulus, cohesion, and internal friction angle from geotechnical tests, the load-mechanical joint feasible domain is delineated by incorporating geotechnical constraints, time evolution constraints, and parameter correlation constraints. The soil weight range is limited to 16~22kN / m³, and the elastic modulus range is 20~80MPa. Parameter combinations without physical meaning, such as inverse pairing of weight and modulus and strength parameters that violate the conventional distribution of soil, are eliminated, thus completing the pre-modeling parameter screening work.

[0021] The entire execution system outputs a complete 3D geological mesh model. The model includes an initial stress field obtained from layered equilibrium and parameter value ranges that have been constrained and filtered. It can be directly connected to the Latin hypercube sampling process to perform batch numerical calculations of load-mechanical parameters. The output 3D displacement vector can be substituted into the Hilbert space linear inner product mapping formula. , , ,in, The displacement of spatial point x along the radar line of sight (LOS) at time t is expressed in mm. Negative values ​​indicate subsidence away from the satellite. Let x be the three-dimensional displacement column vector of point x; The unit direction cosine vector of the radar line of sight; The unit cosine of the eastward direction; The cosine of the northward direction; The vertical / zenith (Up) unit direction cosine; This is an eastward displacement; This is a northward displacement; For celestial displacement; Complete the conversion of the global 3D displacement to the radar LOS line-of-sight matrix, achieving seamless integration between the modeling stage and the subsequent GA-RBF proxy model training process.

[0022] Following four fundamental specifications—matching the stratigraphic contour to the measured on-site topography, adapting the grid size to finite difference numerical calculations, matching stress evolution to the actual phased construction, and ensuring parameter values ​​conform to geotechnical testing—the system unifies and matches the coordinates, dimensions, and elevations of three data sources: UAV LiDAR full-domain topographic mapping results, engineering geological borehole stratification data, and soil foundation mechanical parameters measured by laboratory geotechnical tests. This standardizes the entire process of modeling the reservoir area's 3D numerical model. For example, reasonable value ranges are defined for the cohesion, internal friction angle, unit weight, and elastic modulus of the tailings and dam foundation soil determined by laboratory geotechnical tests. Taking a Class 4 tailings project as an example, the soil weight range is 16~22kN / m³, and the elastic modulus is 20~80MPa. Parameter combinations outside the range are directly eliminated. The layer thickness of the dam is completely matched with the mine's annual dam construction log. The single dam layer height is fixed at 2.0m, and the dam construction simulation is completed in 12 phases. The grid size is set according to the different deformation risks of the dam body. The grid unit side length is 2m in the dam top, upstream and downstream dam slopes, and dam foundation sliding high-risk areas, and the grid unit side length is 8m in the gentle base of the reservoir area. A hexahedral structured grid is used uniformly. Geometric modeling is carried out after all data is aligned with the benchmark to prevent simulation distortion problems such as stratigraphic interface misalignment, grid adaptation imbalance, and stress field mismatch with actual working conditions.

[0023] In the geometric modeling stage, the borehole layer contour lines were stitched together using the NURBS surface algorithm, with a sampling interval of 2m for the surface control points. This was matched with a 3m resolution LiDAR terrain point cloud to reconstruct the natural undulations of the multi-layered strata. After mesh generation, the data was imported into numerical software, and the Mohr-Coulomb constitutive model was used to characterize the soil shear mechanics. The mesh was activated layer by layer using the layered stacking equilibrium method, and gravity loads were applied. After each layer was loaded, iterative equilibrium of self-weight stress was performed, with the unbalanced force ratio being less than [a certain value]. Convergence is determined, and the initial stress field that fits the field is obtained.

[0024] Based on the execution requirements of achieving model accuracy, realistic stress, and no physical contradictions in parameters, NURBS surface modeling rules, mesh refinement standards for key areas of the dam body, layered embankment stress calculation procedures, and geotechnical physical constraint judgment rules were established to complete the entire three-dimensional numerical modeling operation. Using three key control requirements as benchmarks—matching model geometric accuracy to the stratigraphic outline marked on the survey drawings, completely reproducing the stress field of the entire phased loading process of the mine, and ensuring that the parameter combinations do not contain logical contradictions in geotechnical mechanics—NURBS surface stratigraphic construction operation standards, differentiated mesh refinement requirements for high-slip risk areas of the dam body, complete calculation procedures for the layered embankment's progressive self-weight stress balance, and multi-dimensional geotechnical parameter constraint judgment criteria were determined. The entire three-dimensional numerical modeling operation was executed in the order of stratigraphic modeling, mesh generation, stress iteration, and parameter selection.

[0025] For example, when modeling a level 4 tailings dam, the sampling interval of control points in the stratigraphic surface construction stage is set to 2m, matching the 3m resolution UAV LiDAR terrain point cloud. The surface contour is aligned with the measured elevation of each borehole layer to prevent deviations between stratigraphic undulations and exploration data. In the mesh generation stage, 2m hexahedral meshes are used for the dam crest, upstream and downstream dam slopes, and high-risk areas of dam foundation slippage. A standard 8m mesh is used for the gentle base surrounding the dam area. All meshes meet the hexahedral dominant mesh requirements for FLAC3D finite difference continuous medium mechanics calculations. Stress calculation follows the layered construction logic of the mine's 12-phase embankment ledger, with a single embankment layer height of 2.0m. After each layer is completed and activated, a self-weight iterative balancing process is performed, ensuring that the unbalanced force ratio is less than... As a convergence criterion, the Mohr-Coulomb constitutive model is uniformly used to characterize the soil shear properties across the entire domain. The parameter constraint process defines the soil weight range as 16~22kN / m³ and the elastic modulus range as 20~80MPa, and simultaneously constrains the upper and lower limits of cohesion and internal friction angle. Unreasonable combinations such as low unit weight with high elastic modulus and strength parameters exceeding the conventional distribution range of soil are directly eliminated, thus completing the joint load-mechanism feasible domain screening.

[0026] The raw LiDAR terrain and borehole strata data from UAVs are input into the execution system, which outputs a 3D numerical model of the tailings reservoir area. This model is refined by meshing, assigning values ​​to the Mohr-Coulomb constitutive model, generating an initial stress field through layered stacking equilibrium, and defining the load-mechanical joint feasible domain while eliminating invalid solutions. Unprocessed UAV full-domain LiDAR terrain point clouds and a complete set of engineering geological borehole layered raw data are directly imported into the established 3D geological modeling execution system for the reservoir area. Strictly following the aforementioned four established operational specifications—NURBS surface, differentiated meshing, layered stacking, and soil and rock constraints—the system completes four core processes step-by-step: strata surface splicing, dam area zoning and mesh generation, full-domain Mohr-Coulomb constitutive model material assignment, and phased stacking self-weight stress iterative equilibrium. Simultaneously, the joint value ranges for soil weight, elastic modulus, cohesion, internal friction angle, and surcharge height are defined. Triple constraints—soil and rock mechanics, time evolution, and parameter correlation—are superimposed to eliminate logically contradictory parameter combinations, ultimately generating a complete 3D numerical model of the tailings reservoir area suitable for subsequent batch sample simulations.

[0027] For example, taking the practical modeling of a level 4 tailings dam as an example, the original LiDAR point cloud in the WGS84 coordinate system and the borehole layer files in the local independent coordinate system are directly imported into the system without manual coordinate transformation. The system automatically lays out NURBS surface control points at 2m intervals to match 3m resolution terrain data. 2m side length hexahedral meshes are generated for the dam crest and upstream and downstream slip zones, and 8m conventional meshes are generated for the flat base of the reservoir area. The units are activated layer by layer according to the 12-phase tailings dam construction ledger, with a single construction layer height of 2.0m. The self-weight of each layer is iterated until the unbalanced force ratio is less than 1. Convergence is determined; the parameter screening process limits the soil density to the range of 16~22kN / m³ and the elastic modulus to the range of 20~80MPa, automatically eliminating combinations that violate the laws of soil and rock, such as low density combined with high modulus. The output model comes with a complete initial stress field and feasible domain of compliant parameters, which can be directly transferred to the Latin hypercube sampling process to generate load-mechanical parameter samples.

[0028] S102 uses Latin hypercube sampling to generate load-parameter samples, batch numerical simulations to obtain three-dimensional displacements, which are then converted into radar LOS displacements through projection matrices. A Gaussian kernel RBF network is built, with stack height and LOS displacement as inputs, and a genetic algorithm is used to uniformly encode all hyperparameters of the network. Leave-one-out cross-validation error is used as the fitness, and adaptive mutation is used to complete global optimization and train the GA-RBF surrogate model.

[0029] In one implementation, a load-mechanism augmented parameter space containing multiple constraints of soil, time, and parameter correlation is defined. Samples are generated using Latin hypercube sampling. Three-dimensional displacements are output by batch calling numerical models through scripts. A sparse block diagonal global topological projection matrix is ​​constructed based on the linear inner product of Hilbert space to achieve a unified conversion of global three-dimensional displacements into radar LOS displacements. First, three types of constraints are integrated: the boundary conditions of geotechnical mechanics values, the evolution law of the phased surcharge time of the dam body, and the inherent coupling relationship of different soil parameters. A load-mechanism joint augmented parameter space covering the surcharge height and a complete set of soil mechanical indices is delineated. Using the Latin hypercube sampling rule, sample points are evenly distributed within the complete parameter range. An automated, cyclical script continuously calls the existing 3D numerical model of the reservoir area to perform batch simulation calculations, simultaneously outputting the eastward, northward, and vertical 3D displacement components of all monitoring nodes in the model. Based on the linear inner product derivation of Hilbert space, a sparse block diagonal global topological projection matrix is ​​constructed. Matrix operations are used to perform a unified dimensional transformation on all 3D displacement fields across the entire domain, batch-computing a radar line-of-sight (LOS) displacement dataset adapted to the network input. The corresponding displacement transformation calculation formula is as follows: ,in, This is an eastward displacement; This is a northward displacement; For celestial displacement, Represents the three-dimensional displacement vector of the node. Representing the unit direction vector of the radar line of sight, the globally discrete system can be extended into matrix form. , where G is a sparse block diagonal matrix, whose diagonal blocks are composed of the cosine vectors of the local line-of-sight directions of each node.

[0030] For example, in the sample generation work for a level 4 tailings dam, geotechnical tests were used to define the soil weight range as 16~22kN / m³ and the elastic modulus as 20~80MPa. Cohesion and internal friction angle were matched with the range of conventional soil and rock values ​​for tailings and bedrock. The surcharge increment covered the entire operational surcharge height range of 0~3.5m. The Latin hypercube algorithm was used to evenly split the boundaries of each parameter range, generating 75 sets of independent sample combinations without repetition. The script automatically imported each set of parameters and surcharge combinations into the FLAC3D finite difference model to complete the layered self-weight balance simulation. It batch output the three-dimensional displacement components of all monitoring points and directly substituted them into the global projection matrix to complete the global synchronous conversion. There was no need to substitute each point into the formula for calculation. The standardized LOS displacement sequence was output at once and directly used as the input feature material for the subsequent Gaussian kernel RBF network, which greatly reduced the repetitive calculation overhead caused by single-point conversion. The entire sample generation and displacement conversion process takes over the output results of the three-dimensional geological model of the reservoir area. The processed displacement dataset can be directly fed into the subsequent GA-RBF network construction and iterative optimization process, and fully connects the entire dam body mechanical parameter inversion technology chain.

[0031] A radial basis function neural network with independent Gaussian kernels in each dimension was constructed. The input was the dam height and multiple sets of LOS displacements, and the output was a complete set of soil mechanical parameters. The network center, kernel width in each dimension, and regularization coefficient were all integrated into a single real chromosome. Leave-one-out crossover was used to verify the negative root mean square error as the fitness of the genetic algorithm. A Gaussian kernel radial basis function neural network architecture was constructed, consisting of three layers from top to bottom: an input layer, a hidden layer, and an output layer. The input layer transmits input features to the hidden layer, completing the nonlinear mapping through a Gaussian kernel. The connection between the hidden layer and the output layer is a linear weighted fully connected layer. The input layer is responsible for receiving two types of field-monitorable variables: phased dam height data and time-series displacement data of radar line-of-sight (LOS) from multiple measurement points. The hidden layer is equipped with a multi-dimensional independent Gaussian kernel operation unit, with a dedicated kernel width parameter configured for each input feature to differentiate the data scale differences of different loads and deformation characteristics. The output layer simultaneously outputs five sets of soil mechanical indicators to be inverted: soil weight, elastic modulus, Poisson's ratio, cohesion, and internal friction angle, fully covering all geotechnical parameters required for tailings dam stability calculation. All adjustable hyperparameters within the network are uniformly encoded as real-valued chromosomes. The encoding includes three types of variables: coordinates of all hidden layer nodes, Gaussian kernel widths for each input dimension, and regularization coefficients. The entire chromosome is used as a complete search individual for iterative optimization in the genetic algorithm. A fitness metric is constructed using the negative root mean square error obtained from leave-one-out cross-validation. A higher value indicates stronger predictive generalization ability on unknown samples, thus reducing overfitting at its root. The formula for Gaussian kernel operation of a single node in the hidden layer is as follows: In the formula, For the input feature vector, For input dimensions; For the first The center in Coordinates in dimensional space; For the first Independent Gaussian kernel width parameters for each input dimension.

[0032] The network output is given by a linear weighted combination of the basis functions: In the formula, The number of central nodes, For the first From the center to the first The connection weights of each output node are used. The coordinates of the center node, the independent width parameters of each dimension, and the regularization parameters are uniformly encoded to form a real-valued chromosome, serving as the complete search space for the genetic algorithm. The chromosome vector construction form is as follows: The formula for calculating the total length of chromosomes is: When constructing the fitness function, leave-one-out cross-validation with negative root mean square error is used, corresponding to the expression: In the formula, Indicates the removal of the first After training samples, by chromosome The first calculated from the network parameters obtained by decoding The predicted value for each sample.

[0033] For example, in actual network construction, the input dimension d is set to 9, consisting of one column of dam height values ​​superimposed with the LOS displacement sequences of eight corner reflector monitoring points; the number of hidden layer center nodes is set to 20, with independent width parameters configured for each of the nine input dimensions; the complete chromosome encoding sequence is composed of 20 sets of 9-dimensional center coordinates, 9 sets of dimension kernel widths, and one regularization coefficient, with a total chromosome length of 20×9+9+1=190. In each round of population fitness calculation, a leave-one-out cross-validation process is executed, sequentially removing single training samples and using the remaining samples to train the network and calculate the prediction error for that sample. The root mean square of all sample errors is calculated, and the negative value is taken as the fitness of the corresponding individual. Throughout the process, the training set fitting accuracy is not used as the optimization objective; priority is given to ensuring the network's predictive ability for unmonitored conditions. After completing chromosome encoding and fitness rule setting, genetic iteration operations such as spatial Gaussian mutation and tournament selection can be performed to finally solve for the network connection weights, obtaining a complete GA-RBF surrogate model for real-time inversion of mechanical parameters from subsequent CR-InSAR measured deformation data.

[0034] Differentiated Gaussian mutations are applied in both the original and logarithmic spaces. When evolution stalls, the mutation probability is automatically increased. This is combined with elite retention, tournament selection, and simulated binary crossover to complete iterative optimization of the population. Finally, regularized least squares are used to solve for the network weights, completing the full training of the GA-RBF surrogate model. After setting the chromosome encoding rules and fitness determination logic, differentiated Gaussian mutations are initiated to coordinate the entire genetic iteration process for global parameter optimization. For the network hidden layer center coordinate variable, Gaussian perturbation is applied in the standardized original feature space to achieve local fine-grained search. For hyperparameters with large numerical ranges, such as Gaussian kernel width and regularization coefficient, the parameters are transformed to the logarithmic value space for Gaussian mutation to balance the search coverage efficiency of parameters of different magnitudes. During the population iteration process, the fitness value changes of each generation are monitored in real time. When the fitness index does not show a significant improvement for ten consecutive generations, the program automatically increases the mutation probability to broaden the search range and escape the local optimum. Throughout the iteration, three types of standard genetic operators—elite retention, tournament individual selection, and simulated binary crossover—are executed synchronously until the preset convergence condition is met. After iteration, the globally optimal chromosome is extracted and decoded, and all network centers, dimensionality kernel width, and regularization hyperparameters are obtained. These are then substituted into the weight calculation formula to calculate all connection weights from the hidden layer to the output layer in one go, completing the entire training process of the GA-RBF surrogate model. The network weight calculation formula is as follows: In the formula, The kernel matrix has the following elements. ; It is the identity matrix; Output matrix for the training set. is the regularization coefficient, I is the identity matrix, and Y is the output matrix of the mechanical parameters of the training set.

[0035] For example, in the actual iterative calculation, the basic population size is set to 40 individuals. In each iteration, two individuals with the best fitness ranking are selected and retained to be directly inherited by the offspring population. In each parent selection phase, three individuals are randomly selected from the population to perform tournament selection, and the individual with the highest fitness is retained as the crossover parent. The simulated binary crossover operator distribution index is set to 20, and the basic mutation probability is fixed at 0.15. If the average negative root mean square error of the population does not significantly improve for 10 consecutive generations, the upper limit of the mutation probability is automatically increased to 0.5. The standard deviation of the Gaussian perturbation of the center coordinate is set to 0.05, and the kernel width and regularization coefficient are set only when... Perturbation samples are generated within the logarithmic domain to prevent probability imbalance in the search of minimum and maximum parameter intervals. A maximum of 100 generations is uniformly set as the maximum number of iterations; reaching this limit indicates algorithm convergence. After convergence, the optimal chromosome of length 190 is analyzed, and all hyperparameters are extracted and substituted into the regularized least squares formula to directly solve for the weight matrix. This eliminates the need for phased debugging of network structure parameters. The trained model can be directly imported into the real-time monitoring process, inputting CR-InSAR measured pile height and LOS displacement data to quickly invert the complete set of soil mechanics parameters for the dam body, thus supporting the subsequent calculation of the dam's stability safety factor. The entire genetic optimization process connects RBF network encoding and sample simulation datasets. Upstream, the training set is constructed based on load mechanics samples generated by Latin hypercube sampling; downstream, it connects to the tailings dam dynamic stability assessment module, completely linking the entire collaborative inversion technology chain.

[0036] S103, dynamically update the load boundary of the model by incremental measurement of dam height; process satellite images with CR-InSAR, filter out noise points and extract time-series LOS displacement of dam corner reflectors; input the real-time dam height and measured displacement into the GA-RBF proxy model to invert the current soil mechanical parameters.

[0037] In one implementation, a real-time inversion calculation framework for soil mechanics parameters is established based on the time-series data of the measured dam height increment, raw CR-InSAR satellite radar images, and coherent deformation datasets from dam corner reflectors. This framework establishes a complete calculation logic for dynamic load boundary updates, time-series deformation noise reduction and extraction, and GA-RBF proxy model inversion calculations. An integrated soil mechanics parameter extrapolation and calculation system is built by integrating three types of independent monitoring data sources: annual measured dam height increments from the mine site, raw CR-InSAR radar images from the satellite platform, and long-term coherent deformation monitoring datasets from corner reflectors on the dam slope. A coherent calculation logic chain is built based on the time-series matching relationship of these three types of data, fully covering three continuous business processes: dynamic load boundary updates, satellite time-series deformation noise reduction analysis, and training of GA-RBF neural network for inversion calculation and prediction. The entire process requires no manual segmentation or data format conversion, directly achieving fully automated flow and calculation of raw on-site monitoring data to a complete set of soil mechanics indicators. It also supports upstream batch simulation results of three-dimensional numerical models and connects with downstream dam stability and safety factor calculation processes.

[0038] For example, a Class 4 tailings dam in operation has been conducting long-term height measurement work, and has been recording the incremental height over time each year. Simultaneously, it has been collecting raw X-band COSMO-SkyMed satellite radar images from November 2024 to July 2025, with a single-period spatial resolution of 3 meters and a total of 14 images. Fifteen corner reflectors have been deployed throughout the dam slope to continuously collect surface coherent deformation data. All three types of monitoring data are integrated into the same computing framework. The time series of tailings height, satellite images, and corner reflector deformation data are automatically aligned on the time axis. There is no need to switch between multiple software programs to process load, deformation, and parameter inversion modules separately. The entire chain of load boundary refresh, image noise filtering, deformation sequence extraction, and GA-RBF network inversion calculation and prediction is fully connected, eliminating problems such as time series misalignment, coordinate mismatch, and data loss caused by the interaction of multiple software programs.

[0039] The entire computational system forms a complete closed loop with the pre-processing Latin hypercube sampling and 3D numerical simulation. The pile height-LOS displacement samples generated by the batch simulation in the early stage are used to complete the global optimization of hyperparameters of the GA-RBF network. The network input is fixed as a single set of dam pile height values ​​combined with multiple measurement points of LOS time-series displacement. The network hierarchy is divided into an input layer, a hidden layer with an independent Gaussian kernel, and an output layer of five types of mechanical parameters. Fully connected mapping is used between the layers. The three types of data sources measured in the field are directly used as network input materials after standardized preprocessing. There is no need to carry out numerical simulation again. A single inversion calculation can simultaneously output a complete set of soil mechanical parameters, including soil weight, elastic modulus, Poisson's ratio, cohesion, and internal friction angle. The output parameters are synchronously transmitted back to the FLAC numerical model to carry out strength reduction stability calculation. The entire data stream does not require manual secondary processing and conversion.

[0040] A standardized CR-InSAR analytical processing workflow is configured for raw satellite SAR image input, uniformly completing preprocessing steps such as filtering geometrically coupled noise on dam slopes, extracting time-series cumulative LOS displacement, and selecting effective monitoring points. The CR-InSAR standardized analytical workflow establishes fixed standardized processing steps for raw satellite radar images. The entire process is divided into three orderly preprocessing stages. The processing logic sequentially completes the filtering of topographically coupled noise on dam slopes, the extraction of full-time cumulative LOS deformation values, and the selection of effective corner reflector points, progressively eliminating invalid interference information within the image. The final output is a time-stable deformation dataset with satisfactory coherence. This dataset can be directly paired with contemporaneous dam height monitoring data as input feature material for the GA-RBF surrogate model, providing reliable deformation input for soil mechanics parameter inversion calculations. The entire satellite image preprocessing process relies on raw COSMO-SkyMedX band radar data. The spatial resolution, temporal period, and corner reflector deployment parameters of the images are matched to the on-site monitoring deployment plan for the reservoir area. The processed LOS displacement data has the same dimensions as the displacement output from numerical simulation. This allows for simultaneous substitution into the Hilbert spatial linear inner product mapping formula to verify the consistency between simulation and measured data. The formula is expressed as follows: In the formula, Represents the three-dimensional displacement vector of a spatial node. It represents the unit direction vector of the radar line of sight, and can simultaneously complete the unified comparison of the simulated output displacement and the actual displacement dimension measured by the satellite.

[0041] For example, 14 X-band satellite images covering the monitoring period from November 2024 to July 2025 were selected, with a single image spatial resolution of 3 meters. Standardized three-level preprocessing was performed in batches. The first step relied on the dam slope topographic geometry model to remove false deformation noise caused by topographic undulation coupling, eliminating false deformation increments caused by slope gradient and stratum elevation difference. The second step was to perform scene-by-scene time-series calculation and output the complete cumulative LOS displacement time-series curves of each corner reflector. The third step set a displacement fluctuation threshold to screen monitoring points and remove failed corner reflectors whose deformation fluctuations exceeded the reasonable range. After screening, 8 sets of coherent and stable points were retained from the original 15 slope corner reflectors. The retained displacement time series had no abrupt changes and could be directly matched with the measured dam height increment values ​​of the corresponding time period. After coupling, the features were input into the network to carry out mechanical parameter inversion, which greatly reduced the problem of inversion parameters deviating from the real soil and rock state caused by abnormal deformation data. The entire image preprocessing process is located in the middle of the load data acquisition and GA-RBF model inversion calculation. Upstream, it receives the on-site dam height time series measurement data, and downstream, it directly connects to the trained three-layer fully connected Gaussian kernel RBF network. The network input dimension consists of one dam height parameter superimposed with eight sets of filtered LOS displacements. No additional manual processing of coordinates and time series alignment is required, and the entire collaborative inversion calculation link is completely connected.

[0042] The real-time increase in dam height Δh is retrieved, and the measured LOS displacement sequence of the corner reflectors is simultaneously matched and screened. The load variables and deformation monitoring data are coupled as input to the GA-RBF surrogate model. The network inversion calculation process is traversed to complete the inversion calculation of the complete set of mechanical parameters of the tailings dam soil in the current period, including weight, elastic modulus, Poisson's ratio, cohesion, and internal friction angle. After completing the two preliminary steps of standardizing satellite imagery preprocessing and collecting the time series data of dam height, the measured increase in dam height Δh in the current period is retrieved. This increase corresponds to the current construction thickness of the tailings dam. It is then matched with the effective corner reflector LOS displacement sequence retained by the CR-InSAR process for dimensionality matching and time series alignment to form a standardized network input dataset. The input dataset dimensions strictly match the GA-RBF neural network input layer construction specifications. The network input layer is fixed to consist of a single stack height scalar superimposed with multiple sets of temporal displacement features. All input features are pre-normalized to the 0-1 interval using Min-Max normalization to eliminate inference bias caused by dimensional differences. The paired input dataset is directly fed into the trained GA-RBF surrogate model, which has undergone global optimization using a genetic algorithm and regularized least squares to solve for the weights. It performs non-iterative inversion calculations without the need for repeated numerical simulations or genetic population iterations. A single forward calculation synchronously outputs multi-dimensional soil mechanics prediction results. The output indicators cover five complete geotechnical parameters required for dam stability strength reduction calculations: soil weight, elastic modulus, Poisson's ratio, cohesion, and internal friction angle. Each parameter can be directly exported for the FLAC3D numerical model to read and update material constitutive assignments.

[0043] For example, the measured increase in pile height Δh of a single phase of a level 4 tailings dam is 3.5m. Combined with the time-series LOS displacements of 8 effective corner reflectors after CR-InSAR preprocessing, this constitutes a 9-dimensional input feature, perfectly matching the network's preset input dimension d=9. The network's hidden layer is configured with 20 independent Gaussian kernel centers, each input dimension having its own width parameter. The Gaussian kernel calculation expression is... After the input dataset is normalized, it is directly fed into the model to complete the inversion calculation. It outputs five sets of parameters at one time: the weight of the tailings dam soil, the elastic modulus, Poisson's ratio, cohesion, and internal friction angle. The output parameters do not need to be converted to dimensions twice. They can be directly fed back to the reservoir area three-dimensional numerical model to update the material parameters of the Mohr-Coulomb constitutive model in a closed loop. This connects the process of solving the dam safety factor and saves the repetitive calculation steps of deriving multiple sets of parameters in batches, which greatly shortens the time required for dynamic stability assessment.

[0044] S104 will update the load and inversion parameters back to the numerical model, calculate the safety factor using the strength reduction method, and classify and issue early warnings according to the national standard threshold. The newly added monitoring data will be updated using incremental learning to update the surrogate model, narrowing the parameter search range for the next period to ±10%. The working condition verification and optimization will be completed based on the dual-drive framework of physical prior and data posterior, forming dynamic monitoring and early warning information for tailings dams.

[0045] In one implementation, a closed-loop calculation framework for tailings dam stability evaluation is established based on real-time updated dam load boundaries, inverted soil mechanical parameters, and time-series SAR monitoring datasets. This framework establishes complete computational logic including safety factor calculation, graded early warning, incremental iteration of the surrogate model, and adaptive contraction of the parameter range. Using three types of fundamental data as underlying inputs—real-time updated dam load boundary conditions, a complete set of soil mechanical indices obtained from each inversion, and a long-term angular reflector SAR deformation monitoring dataset—a closed-loop calculation framework for tailings dam stability is constructed. Within this framework, four sequentially executed computational logics are fixed: iterative calculation of the dam stability safety factor, dam risk level determination and early warning output, incremental iterative update of the GA-RBF surrogate model, and adaptive contraction of the feasible region of soil mechanical parameters. The entire computational chain forms a complete closed-loop flow relationship. Upstream, it directly receives soil parameters and current load increments output by the GA-RBF mechanical inversion module; downstream, it uniformly outputs standardized tailings dam dynamic safety early warning results. Each computational step can be executed automatically and sequentially without manual intervention.

[0046] For example, during normal operation, the construction records of a Level 4 tailings dam are collected monthly to continuously update the load boundary. After each phase of X-band COSMO-SkyMed satellite CR-InSAR analysis of 14 time-series images, the complete set of inversion values ​​for soil weight, elastic modulus, Poisson's ratio, cohesion, and internal friction angle are simultaneously output. The time series of LOS deformation of 8 effective corner reflectors accumulated from November 2024 to July 2025 are all included in the unified calculation pool of the framework. The four processes of load update, safety factor calculation, model incremental learning, and parameter interval shrinkage run automatically and continuously. Operators do not need to separately retrieve load files, inversion results, and satellite deformation data to perform segmented calculations, avoiding problems such as data time series mismatch, confusion of multiple parameter versions, and omission of deformation samples that interfere with stability evaluation during the switching of multiple software. The entire closed-loop framework integrates CR-InSAR time-series processing and GA-RBF inversion calculation outputs. During the calculation process, the updated load and inverted soil parameters are synchronously imported into the FLAC numerical model to carry out strength reduction stability calculations. The risk is determined using the current "GB39496-2020" tailings dam safety specification classification threshold. After each new satellite deformation sample completes stability assessment, incremental learning is performed, only adding the new sample iterative network weights without restarting the complete genetic optimization process. Using the mechanical parameters obtained from the current inversion as a benchmark, the parameter search range for the next round is narrowed to within ±10% of the benchmark value. Relying on the physical prior constraints of 3D numerical simulation and the post-correction of satellite measured deformation data, a dual-drive verification system is formed to continuously optimize the inversion accuracy.

[0047] Taking soil mass weight as an example, if the current inversion baseline value is 19 kN / m³, the feasible region for the next period is automatically locked at 17 to 21 kN / m³. The elastic modulus, cohesion, and internal friction angle simultaneously execute the same interval shrinkage rule, narrowing the search range of the genetic algorithm population, reducing the computation time for 100 generations of complete evolutionary iterations, and improving the efficiency of soil parameter inversion for the next period. The incremental learning model uses the original three-layer fully connected Gaussian kernel RBF network structure. The input remains the single-period pile height combined with 8 sets of filtered LOS displacements, and the output consists of five types of soil mechanical indices. The network Gaussian kernel calculation formula is... Incremental iteration only updates the hidden layer to the output connection weight matrix of the network, and the weights are calculated using the regularized least squares formula. There is no need to re-perform the entire set of genetic operations, including population initialization, tournament selection, and adaptive mutation, on a 190-length real chromosome. Only local weight adjustments are made to the samples after the training set is expanded, balancing model update speed and inversion generalization performance. The entire technical chain of deformation monitoring, parameter inversion, stability calculation, and model self-optimization is fully connected.

[0048] A standardized access and analysis process for numerical models is configured to handle updated loads and inverted mechanical parameters, uniformly completing the preprocessing for dam stress field reconstruction and LOS displacement comparison and verification. After completing the inversion calculation of soil mechanical parameters to obtain a complete set of geotechnical indices and updating the current surcharge increment, a standardized access and analysis process is configured for two types of input materials: the updated dam surcharge load conditions and the current soil mechanical parameters output by the GA-RBF model. The entire preprocessing process includes two core processing components: reconstruction of the dam's global stress field and bidirectional comparison and verification between numerical simulation LOS displacement and CR-InSAR satellite measured deformation. Both types of materials must complete the preprocessing process before being imported into the FLAC numerical model for stability coefficient iteration calculation, thereby ensuring that the stress distribution within the model, the dam deformation simulation results, and the actual surcharge operation conditions at the mine site are consistent.

[0049] The stress field reconstruction reuses the layered stacking equilibrium calculation rules of the previous three-dimensional geological model, strictly following the stacking sequence of the mine in 12 phases, activating grid cells layer by layer, setting the single-layer stacking height to 2.0m, and using an unbalanced force ratio of less than 10 as the self-weight equilibrium iteration convergence threshold. -5 The two-way displacement verification consists of two steps. The first step converts the global three-dimensional displacement field output from the numerical model into line-of-sight data using a sparse block diagonal global topological projection matrix. The single-node transformation employs a Hilbert space linear inner product mapping relationship, with the corresponding formula: Global batch conversion matrix form: In the formula, Represents the three-dimensional displacement vector of a spatial node. G represents the radar line-of-sight unit direction vector at this point, and G is the sparse block diagonal global topological projection matrix. The diagonal blocks of the matrix are composed of the cosine vectors of the line-of-sight directions at each point. The second step is to compare and verify the difference between the batch output of the matrix simulation LOS time series data and the satellite measured angular reflection deformation sequence.

[0050] For example, the measured increase in the height of a Level 4 tailings dam is 3.5m. Simultaneously, five mechanical parameters—soil weight, elastic modulus, Poisson's ratio, cohesion, and internal friction angle—are input into the preprocessing flow. The system reactivates the mesh layer by layer according to the complete 12-phase layered tailings history, iteratively balancing and generating a global stress field matching the current tailings thickness. Then, it batch-calculates the simulated line-of-sight displacement of all monitoring points using a global topological projection matrix, comparing the deviation with the measured deformation of eight effective corner reflectors extracted from 14 phases of X-band COSMO-SkyMed satellite imagery. If the deformation deviation at a single point exceeds the reasonable range, the program automatically jumps to the GA-RBF inversion calculation module to verify the input tailings height and displacement characteristics, and to check the rationality of the network output parameters. After completing all steps of stress reconstruction and displacement bidirectional verification, the load boundary and soil parameter materials are allowed to be connected to the strength reduction stability calculation module to solve for the safety factor, avoiding stress field distortion and deformation simulation deviation from the actual working conditions caused by directly importing loads and parameters. The entire preprocessing procedure receives the output results of GA-RBF inversion and connects downstream with the dam stability safety factor calculation process, forming a complete technical chain of parameter inversion, data verification, and stability assessment.

[0051] The target tailings dam's graded safety threshold benchmark was retrieved, and updated loads and soil mechanical parameters were simultaneously imported into the numerical model. The strength reduction method was used to traverse the stability calculation process to solve for the dam's safety factor. The risk level was determined by comparing with the national standard threshold, and an early warning was output. The graded safety control benchmark stipulated in the "GB39496-2020" tailings dam safety regulations was retrieved, and the real-time surcharge boundary conditions after stress reconstruction and LOS displacement bidirectional verification preprocessing, along with a complete set of inverted soil mechanical parameters, were simultaneously imported into the FLAC three-dimensional numerical model. The numerical model uniformly adopted the Mohr-Coulomb constitutive model to characterize the soil shear failure characteristics throughout the entire domain. The strength reduction calculation method was used throughout the process to progressively reduce the soil shear strength index, completely traversing the entire process of potential slip formation, expansion, and evolution within the dam, and iteratively solving for the overall stability safety factor of the dam. The obtained safety factor was compared with the minimum limit value of the corresponding tailings grade specification, and different risk levels were divided according to the difference range between the coefficient and the threshold. Corresponding early warnings and complete original assessment records were output simultaneously. The strength reduction is performed by equivalent conversion of cohesion and internal friction angle. The reduction calculation formula is as follows: , In the formula: c, The original soil cohesion and internal friction angle obtained by inversion from the GA-RBF model; The intensity reduction factor used in a single iteration; , This represents the equivalent shear resistance parameter after reduction.

[0052] The program automatically zooms in gradually. The process continues until the model generates a continuous plastic slip surface. This critical reduction factor is the overall safety factor of the dam. For example, if the reservoir area to be assessed is a Class 4 tailings area, the national standard specifies that the minimum safety control threshold for normal operation of this class is 1.25. The load boundary corresponding to the current 3.5m surcharge increment and the complete set of inverted soil mechanical parameters are imported into the preprocessed numerical model. The program performs strength reduction iterations step by step and determines that the safety factor when the slip is continuous is 1.93. This value is higher than the standard threshold, and the system automatically determines that the dam is in a stable condition. If the subsequent surcharge is increased and the soil mechanical properties deteriorate, and the safety factor is found to be lower than 1.25 after repeated calculations, yellow and red grade warnings are output according to the deviation range. At the same time, the current surcharge ledger, GA-RBF inversion parameters, and CR-InSAR corner reflector time-series deformation original data are automatically retained as reference data. The retained data can be directly used for subsequent incremental learning of the proxy model and parameter interval shrinkage. The entire stability calculation process takes over the load and parameter preprocessing output results, and the downstream connection adds a SAR sample iterative optimization link, forming a complete closed-loop technical link of deformation monitoring - parameter inversion - stability assessment - model self-optimization.

[0053] Based on newly added SAR monitoring samples, GA-RBF incremental learning iteration is carried out. The current inversion parameters are used as a benchmark to shrink the feasible region of the next period's parameters to ±10%. Combined with the dual-driven assimilation framework of physical prior and data posterior, full-condition verification and optimization are completed, forming dynamic safety monitoring and early warning information for all types of tailings dams. After the single-period complete dam stability assessment process is completed, the system automatically retrieves the current period's CR-InSAR satellite time-series deformation monitoring samples and the complete set of soil mechanical parameters obtained from the corresponding GA-RBF model inversion, and starts the proxy model incremental iteration update process. Only the newly added pile height, LOS displacement, and soil parameter samples are incorporated into the original training dataset for local weight correction, without re-executing the entire set of calculations for population initialization and 100 generations of genetic optimization. At the same time, the soil weight, elastic modulus, Poisson's ratio, and cohesion obtained from the current period's inversion are used. The values ​​of force and internal friction angle are used as the search benchmarks for each parameter. The feasible domain of the parameters in the next round of Latin hypercube sampling is simultaneously narrowed to the range of 10% above and below the benchmark value. Throughout the process, the geotechnical mechanical boundary obtained by three-dimensional numerical simulation is used as the physical prior constraint, and the satellite measured time-series deformation data is used as the basis for a posteriori correction. A dual-drive assimilation and verification system is built to continuously correct the load-mechanical parameter matching relationship, iteratively optimize the inversion accuracy of the entire dynamic judgment system, and finally generate standardized dynamic safety early warning data for tailings dams that are adapted to different levels and different stacking conditions.

[0054] The entire incremental iterative operation uses the original three-layer fully connected Gaussian kernel RBF network structure. The network input is fixed as one set of dam heights combined with eight sets of filtered corner reflector LOS displacements. The output consists of five types of soil mechanical properties. The network weights are still solved using the regularized least squares formula. In the incremental phase, only the connection weights from the hidden layer to the output layer are updated. Instead of performing global optimization on the real chromosome of length 190 again, the network parameters are locally corrected after expanding the training sample set, which greatly reduces the computation time.

[0055] For example, the current inversion soil weight baseline value for a Class 4 tailings dam is 19 kN / m³, and the feasible region for the next sampling period is automatically locked at 17 kN / m³ to 21 kN / m³; the current elastic modulus inversion result is 36 MPa, and the parameter search range for the next round is adjusted to 32.4 MPa to 39.6 MPa, with cohesion, internal friction angle, and Poisson's ratio all following the same range contraction rule; the 8 corner reflector deformation samples corresponding to 14 sets of time-series images from the newly added X-band COSMO-SkyMed satellite are directly incorporated into the original 60 training sets, and only regularized least squares weighted samples are used. The new solution for network weights eliminates the need to repeatedly set a full set of genetic parameters, such as a 40-population base, a 0.15 base mutation probability, and 100 generations of iterations. The physical prior constraints adopt the boundary values ​​of 16~22kN / m³ and 20~80MPa for the entire geotechnical foundation from the previous study. Eight sets of time-series deformations measured by satellite are used as the basis for posterior correction. The parameter adaptation and correction are completed by comparing the deviation between the simulated LOS and the measured deformation. After multiple rounds of incremental sample iteration and interval adaptive shrinkage, the system can adapt to the dynamic assessment needs of various upstream, midline, and downstream tailings dams of levels 1 to 5, and output standardized graded safety early warning records.

[0056] like Figure 2 As shown, Figure 2 This is a flowchart illustrating the overall process of collaborative inversion of soil mechanical parameters and stability early warning for tailings dam dams based on CR-InSAR time-series monitoring and the GA-RBF algorithm. A "Start" node is set at the top of the diagram, and an "End" node is set at the bottom. The entire process is divided into five sequentially connected modules from top to bottom. Data is transferred between modules via downward arrows. The fourth module ends with a feedback arrow connecting to the second module, forming a closed-loop iterative logic. The internal branches of each module are as follows: The first module is for building an initial three-dimensional geological model. Two types of basic raw data are input on the left side of the module: borehole strata data and high-precision terrain data from UAVs. The two types of data are used to perform three-dimensional geological modeling calculations in sequence, and finally generate the initial three-dimensional geological model of the tailings dam. The model output data flows down into the second module.

[0057] The second module constructs the surrogate model using a GA-RBF (Gas-Induced-Free) approach. The process involves determining the range of soil physical and mechanical parameters and the surcharge increase interval; conducting batch simulations using a Latin hypercube (LHS) sampling input numerical model; outputting 3D displacement data and converting it to radar line-of-sight (LOS) displacement; performing normalized data preprocessing on all simulation samples; inputting the data into an RBF neural network using a genetic algorithm to perform global hyperparameter optimization; and generating a trained GA-RBF surrogate model. A feedback arrow is located on the right side of the module to receive closed-loop iterative update data from the fourth module. The module's output data flows downwards into the third module.

[0058] The third module is for data updating and collaborative inversion, internally divided into three parallel sub-branches. The first is the physical prior model-driven branch: it inputs the measured on-site pile height increment, activates the grid through a numerical model, and applies surface load stress to complete the load boundary update. The second is the data posterior data-driven branch: it inputs the original CSK satellite radar imagery, performs CR-InSAR standardized analytical processing, and extracts the time series data of line-of-sight displacement of effective monitoring points on the dam body. The third is the collaborative inversion branch: it couples the pile height increment and the LOS displacement of multiple measuring points as input materials and sends them to the GA-RBF proxy model. The model inversion calculates and outputs a complete set of soil physical and mechanical parameters. The text on the right side of the module indicates that the inverted soil parameters and load update data can participate in closed-loop iteration, and the overall output data of the module flows downward into the fourth module.

[0059] The fourth module is for stability analysis with coordinated dynamic updates of physical and mechanical parameters. It has a closed-loop feedback function. The workflow is as follows: the soil parameters output from the third module and the updated load are synchronously fed back to the three-dimensional numerical model; the simulated LOS displacement of the monitoring points in the numerical model is calculated and compared with the satellite measured displacement; the strength reduction algorithm is used to iteratively calculate the dam stability safety factor FOS; the calculated safety factor is compared with the standard safety threshold; this module has a feedback arrow to feed back the updated parameters and sample data to the second module for iterative optimization, and the main flow flows down to the fifth module.

[0060] The fifth module is for risk warning output, with a built-in diamond-shaped decision diagram. The decision condition of the diamond-shaped decision box is whether the safety factor FOS is greater than or equal to the standard threshold. There are two branch paths: if the decision result is yes, it flows to the left normal and stable branch, and the warehouse area is determined to be risk-free and no warning needs to be pushed; if the decision result is no, it flows to the right risk warning branch, and the system triggers and outputs the graded warning result; the two branch paths eventually merge and connect to the "end" node at the bottom of the diagram.

[0061] like Figure 3 As shown, Figure 3This is a schematic diagram of the topology of a three-layer RBF neural network. From left to right, it is divided into three main parts: the input layer, the hidden layer, and the output layer. The layers are fully connected, and all nodes are connected to each other by straight lines.

[0062] The input layer is located on the left side of the graph and contains... , Until There are n input nodes in total. The nodes are separated by ellipses to represent other input features, which correspond to network input feature variables such as dam height increment and LOS displacement of multiple measurement points. All input nodes are independently connected to each node in the hidden layer.

[0063] The hidden layer is located in the middle area of ​​the graph and is labeled as a hidden layer. , Until There are j radial basis kernel nodes. The ellipsis between the nodes represents the remaining Gaussian kernel center units. Each node corresponds to a set of independent Gaussian kernel operations. All hidden layer nodes simultaneously receive data transmitted from all input layer nodes and output the calculation results to all output layer nodes respectively.

[0064] The output layer is located on the right side of the diagram and is labeled as the output layer. , Until There are m output nodes in total. The ellipsis between the nodes represents the remaining output components, corresponding to five types of soil mechanical parameters: soil weight, elastic modulus, Poisson's ratio, cohesion, and internal friction angle. Each output node is connected to all hidden layer nodes, and the final output value is obtained by weighted summation of the basis functions. The overall topology fully presents the transmission logic of a three-layer fully connected network. After the input features are mapped by the Gaussian kernel of the hidden layer, the multi-dimensional mechanical parameter output results are obtained through linear weighted calculation.

[0065] like Figure 4 As shown, Figure 4 This is a flowchart of the complete iterative process of the genetic algorithm used to optimize the hyperparameters of the RBF network. The top of the diagram has a "Start" node, and the bottom has an "End" node. The process is arranged from top to bottom, with step boxes and logic judgment boxes arranged sequentially. When the iteration termination condition is not met, the process flows back to the fitness evaluation step, forming a cyclic iterative logic. The contents of each step are as follows: After the process begins, population initialization is performed first. The cluster centers output by the training set clustering algorithm are used as the initial reference for the central genes, and random perturbations are added to generate all individuals in the initial population. The width gene and the regularization coefficient gene are randomly initialized according to a logarithmic uniform distribution within their respective value ranges. Decoding is performed on each chromosome individual in the population to obtain all hyperparameters of the corresponding RBF network. Leave-one-out cross-validation is performed, and the negative root mean square error obtained from the calculation is set as the fitness value of that individual. After completing the fitness evaluation, the iteration termination condition judgment diamond diagram is entered.

[0066] The iteration termination logic is as follows: the determination condition is whether the pre-set maximum number of generations has been reached. There are two branch paths: one is that if the determination result is yes, the globally optimal chromosome is directly output and the process is connected to the "end" node; the other is that if the determination result is no, the process is executed sequentially in the order of elite retention, selection, crossover and mutation. After the mutation step is completed, it flows back to the fitness evaluation stage and starts a new round of iteration.

[0067] Sort all individuals in the current population according to their fitness values ​​from highest to lowest. Select a specified number of individuals with the highest fitness values ​​and directly retain them to the next generation population, without participating in subsequent selection, crossover, or mutation operations. Alternatively, use a tournament selection strategy to randomly select several individuals from the population, choosing the individuals with the highest fitness as parents for crossover operations. Then, use a simulated binary crossover operator to perform genetic recombination on the selected parent individuals according to a preset crossover probability, generating a new generation of offspring individuals.

[0068] For each gene locus in the offspring, perturbation is applied according to the preset mutation probability; the mutation space rules of different genes are distinguished, the central gene is subjected to Gaussian mutation in the original numerical space, and the width gene and the regularization coefficient gene are subjected to Gaussian mutation in the logarithmic space; if the iteration process reaches a stagnation state where the fitness does not improve for several consecutive generations, the program adaptively increases the mutation probability; after all mutations are completed, the process flows back to the fitness evaluation step to start a new cycle.

[0069] like Figure 5 As shown, Figure 5 This document presents a flowchart for the complete construction and testing process of the GA-RBF joint optimization proxy model. Starting with "Start," the process is divided into three main functional sections. Each step flows sequentially from top to bottom, with iterative loops using return arrows to form a closed loop for parameter optimization. The first section is the initial RBF neural network construction module, which performs two tasks simultaneously: on the left side of the module, the type of radial basis functions in the hidden layers and the number of hidden layer nodes are determined; on the right side, the basic network parameters are initialized, including the kernel center C and the Gaussian kernel width. The output layer weights W; the top frame of the module clearly defines the network input and output data types. The input layer includes dam load and time-series displacement data of each measuring point, and the output layer outputs soil mechanical parameters such as soil weight, elastic modulus, and Poisson's ratio.

[0070] The second section is the genetic algorithm population initialization and iterative optimization module. The population initialization rules are set with two types of parameters: the network kernel center gene generates initial individuals by superimposing random perturbations based on the cluster centers of the training set; the width and regularization coefficient genes are randomly assigned values ​​in a logarithmically uniform distribution within their corresponding value ranges. After initialization, the main loop begins, first calculating the fitness function F for all individuals in the population, and then entering the diamond-shaped decision box to determine whether the preset maximum number of generations has been reached.

[0071] If the iteration limit is not met, the three genetic operations of elite retention, selection crossover, and mutation are executed sequentially. After updating the population, the process returns to the fitness calculation step and continues to cycle. In the elite retention step, the best individual in the current population is retained. Then, genes are recombined through selection crossover. Finally, the mutation operation broadens the parameter search range and continuously optimizes the combination of network hyperparameters.

[0072] If the preset maximum number of generations is reached, the iteration loop exits, and the process proceeds to the model testing section. The third section is the GA-RBF model performance verification module, which sequentially completes four operations: generating a complete GA-RBF model based on the optimal hyperparameters obtained through iteration; inputting an independent test dataset into the model; outputting the predicted soil mechanics parameters from the model; calculating the root mean square error (RMSE) and correlation coefficient (R) to quantify the model's prediction accuracy. After all verification steps are completed, the process reaches the "End" node. The entire process first builds a basic neural network framework, then uses a genetic algorithm to globally iteratively optimize the network hyperparameters, and finally uses independent test samples to verify the model's performance, thus fully realizing the construction and accuracy evaluation of a GA-RBF-based surrogate model for soil mechanics parameters.

[0073] like Figure 6 As shown, Figure 6 This is an aerial photographic base map taken by a drone at the tailings dam site. It is a planar distribution map of the corner reflectors in the dam area, complete with coordinates, a north arrow, a scale bar, and a legend. The overall elements are described as follows: The base is an aerial photograph of the actual dam area. The four sides of the map frame are marked with geographical coordinates. The horizontal axis is east longitude, ranging from 113.164°E to 113.168°E; the vertical axis is north latitude, ranging from 25.841°N to 25.843°N, accurately defining the entire geographical area of ​​the tailings dam. The image shows upstream mountains and forests, the dam beach, downstream sedimentary dry beach, simple roads within the site, and small supporting buildings.

[0074] A north-facing compass is placed in the upper right corner of the map to indicate true north; a linear scale with a scale of 0 to 50 meters is placed in the lower right corner of the map for measuring actual distances on site; a legend is attached to the right side of the map, with a blue triangle symbol representing a corner reflector and indicating that the base is an optical map of a UAV.

[0075] The map shows 15 CR series corner reflection monitoring points marked with blue triangles, numbered CR1, CR2, CR3, CR4, CR5, CR6, CR7, CR8, CR9, CR10, CR11, CR12, CR13, CR14, and CR15. These points are evenly distributed along the stratified beach and dam slope access roads of the reservoir area, covering the entire area from upstream to downstream, and are used for satellite CR-InSAR deformation time-series monitoring and data acquisition.

[0076] like Figure 7 As shown, Figure 7 This is a schematic diagram of the elevation of a 3D numerical mesh model of a tailings dam, showing the layered construction mesh and simulation monitoring point layout of the dam area. The overall elements are described as follows: The main dam area of ​​the model adopts a red hexahedral structured mesh throughout, with a finer mesh size; the top of the dam is layered with multiple mesh layers of different colors from top to bottom, namely orange mesh layer, purple mesh layer, and cyan mesh layer, corresponding to the batch-by-batch layered construction levels of the tailings dam. The layered mesh distinguishes the filling body of different construction stages, intuitively reflecting the 12-phase layered construction modeling rules.

[0077] Multiple colored cube markers are discretely distributed on the red dam slope, representing monitoring points for internal displacement extraction in numerical simulation. These monitoring points are dispersed to various slip risk areas upstream, middle, and downstream of the dam. The spatial location of each point corresponds one-to-one with the actual location of 15 corner reflectors on site, which are used to extract three-dimensional node displacement data in batches. This data is then converted into radar LOS line-of-sight displacement and compared with satellite-measured deformation for two-way verification.

[0078] The model fully reproduces the tailings dam's narrow top and wide bottom outline. The grid is fine and uniform, and the layered colors are used to distinguish different filling stages. The slope measurement points provide data extraction points for subsequent LOS displacement conversion and comparison of simulation and measured data.

[0079] like Figure 8 As shown, Figure 8A line graph comparing the LOS-direction settlement values ​​of each effective monitoring point is used to demonstrate the matching relationship between the deformation measured by CR-InSAR satellite and the deformation simulated by FLAC3D. The elements of the graph are explained as follows: The vertical axis represents the LOS-direction settlement, in millimeters (mm), with a value range of -40 to 20; the horizontal axis represents the 8 groups of effective angular reflection monitoring points after screening, labeled CR2, CR4, CR5, CR8, CR9, CR10, CR12, and CR13 in sequence; the legend at the bottom distinguishes between two sets of data: the blue solid line dots represent the InSAR satellite measured data, and the red dashed triangles represent the FLAC3D numerical simulation calculated data.

[0080] The data curve distribution characteristics are as follows: From CR2 to CR9, the measured and simulated settlement values ​​are close to 0 mm, the settlement deformation is small, and the two curves are highly consistent with each other, with a slight difference. At CR12, two data points show a large abrupt change in settlement, which is the peak deformation position in this figure. The InSAR measured settlement is -25.00 mm, and the FLAC3D simulated settlement is -26.93 mm. The settlement at CR12 and CR13 gradually recovers, and the two curves rise synchronously, with the overall trend remaining consistent.

[0081] Next to each monitoring point, the corresponding measured and simulated settlement values ​​are simultaneously marked. Negative values ​​indicate that the dam body has settled and deformed along the radar line of sight. The overall trend of the two curves is highly consistent, with only a small numerical deviation. This can prove that the simulation results of the three-dimensional numerical model are in good agreement with the actual deformation measured by satellite on site. It can be used for two-way verification of the rationality of load and soil parameters.

[0082] In one embodiment, this application also provides a tailings dam load-mechanical parameter co-inversion system based on CR-InSAR time series and GA-RBF, comprising: The reservoir area 3D geological modeling module is used to collect multi-source exploration data of UAV LiDAR topography and borehole strata, build a 3D geological model of the reservoir area based on NURBS surface, and densify the mesh in the key deformation area of ​​the dam body; the initial stress field is generated by the Mohr-Coulomb constitutive model and the layered stacking equilibrium method, the joint feasible domain of load and mechanical parameters is delineated, and multiple physical constraints are introduced to eliminate invalid parameter combinations. The GA-RBF surrogate model training module is used to generate load-mechanical parameter samples by Latin hypercube sampling, conduct batch numerical simulations to output three-dimensional displacement fields, and convert radar LOS displacement using a global projection matrix. A Gaussian kernel RBF neural network is built, with stack height and LOS displacement as inputs. All network hyperparameters are uniformly encoded through a genetic algorithm. The fitness is achieved by using leave-one-out cross-validation error as the fitness and adaptive mutation to complete global optimization, thus completing the training of a high-precision GA-RBF surrogate model. The real-time inversion module for dam mechanical parameters is used to dynamically update the load boundary of the numerical model based on the measured dam height increment; process CR-InSAR satellite images and filter deformation noise, extract the time-series LOS displacement of the dam corner reflector; input the real-time dam height and measured deformation data into the GA-RBF model to invert the full set of soil mechanical parameters of the tailings dam in the current period. The tailings dam stability early warning and judgment module is used to transmit updated loads and inverted mechanical parameters back to the numerical model in a closed loop, calculate the dam safety factor using the strength reduction method, and trigger risk warnings by comparing with the national standard classification thresholds; it incrementally updates the proxy model based on newly added monitoring data, narrowing the parameter search range for the next period to ±10%; it completes the working condition verification and optimization based on the physical prior-data posterior dual-drive framework, and outputs dynamic monitoring and early warning information of the tailings dam.

[0083] The computer-readable storage medium provided in the above embodiments of this application and the tailings dam load-mechanical parameter co-inversion method based on CR-InSAR time series and GA-RBF provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0084] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating a method, system, electronic device, and readable storage medium for co-inversion of tailings dam load-mechanical parameters based on CR-InSAR time series and GA-RBF are basically similar to the aforementioned embodiment of a method for co-inversion of tailings dam load-mechanical parameters based on CR-InSAR time series and GA-RBF, so the description is relatively simple. Relevant parts can be referred to in the description of the aforementioned embodiment of a method for co-inversion of tailings dam load-mechanical parameters based on CR-InSAR time series and GA-RBF.

Claims

1. A method for the collaborative inversion of load-mechanical parameters of tailings dam body based on CR-InSAR time series and GA-RBF, characterized in that, include: We collected LiDAR topographic and borehole stratigraphic data from UAVs, constructed a three-dimensional geological model of the reservoir area based on NURBS surfaces, and refined the mesh in the key areas of the dam body. We used the Mohr-Coulomb constitutive model and the layered stacking equilibrium method to generate the initial stress field. We delineated the joint feasible region of load and mechanical parameters and added physical constraints to eliminate invalid solutions. Load-parameter samples were generated by Latin hypercube sampling, and three-dimensional displacements were obtained by batch numerical simulation. The displacements were then converted into radar LOS displacements by projection matrix. A Gaussian kernel RBF network was built, with stack height and LOS displacement as inputs. All hyperparameters of the network were uniformly encoded by a genetic algorithm. The fitness was determined by leave-one-out cross-validation error, and global optimization was completed by adaptive mutation. The GA-RBF surrogate model was then trained. The measured dam height increment is used to dynamically update the model's load boundary; CR-InSAR satellite imagery is processed to filter out noise and extract the time-series LOS displacement of the dam body's corner reflectors; the real-time dam height and measured displacement are input into the GA-RBF proxy model to invert the current soil mechanical parameters. The updated load and inversion parameters will be fed back to the numerical model, the safety factor will be calculated using the strength reduction method, and the warning will be classified according to the national standard threshold. The newly added monitoring data will be updated using incremental learning to update the surrogate model, narrowing the parameter search range for the next period to ±10%. The working condition verification and optimization will be completed based on the dual-drive framework of physical prior and data posterior, forming dynamic monitoring and early warning information for tailings dams.

2. The method as described in claim 1, characterized in that, Data were collected from UAV LiDAR topography and borehole stratigraphy. A three-dimensional geological model of the reservoir area was constructed based on NURBS surface, and the grid was densified in the key areas of the dam. The initial stress field was generated using the Mohr-Coulomb constitutive model and the layered stacking equilibrium method. Delineate the joint feasible region of load and mechanical parameters, and add physical constraints to eliminate invalid solutions, including: Combining the scenario of building a three-dimensional numerical model of tailings, and the technical characteristics of multi-source data fusion of terrain boreholes, stratigraphic surface reconstruction, mesh generation, stress generation, and parameter domain constraints, four core processing steps are introduced: multi-source data integration, three-dimensional numerical model building, initial stress solution, and parameter space constraints, to build an execution system for constructing a three-dimensional geological model of the reservoir area. Following the processing specifications of conforming the geological geometry to the site, mesh adaptation differential calculation, stress matching stacking process, and parameters meeting the rules of geotechnical physics, the three types of information, namely UAV LiDAR terrain data, borehole geological data, and basic geotechnical parameters, are jointly matched to complete the alignment of the three-dimensional numerical model modeling process. Based on the execution requirements of achieving model accuracy, realistic stress, and no physical contradictions in parameters, NURBS surface modeling rules, grid densification standards for key areas of the dam body, layered embankment stress calculation process, and geotechnical physical constraint judgment rules were set to complete the entire three-dimensional numerical modeling operation. The system inputs raw LiDAR terrain and borehole strata data from UAVs into the execution system, and outputs a three-dimensional numerical model of the tailings pond area that has been meshed, assigned values ​​by the Mohr-Coulomb constitutive model, generated an initial stress field by layered stacking equilibrium, and delineated the load-mechanism joint feasible domain and eliminated invalid solutions.

3. The method as described in claim 1, characterized in that, Load-parameter samples were generated using Latin hypercube sampling, and three-dimensional displacements were obtained through batch numerical simulations. These displacements were then converted to radar LOS displacements using a projection matrix. A Gaussian kernel RBF network was constructed, with stack height and LOS displacement as inputs. All hyperparameters of the network were uniformly encoded using a genetic algorithm. Leave-one-out cross-validation error was used as the fitness, and adaptive mutation was employed to complete global optimization. The GA-RBF surrogate model was trained, including: A load-mechanics augmented parameter space with multiple constraints including soil, time, and parameter correlation is defined. Samples are generated by Latin hypercube sampling. Three-dimensional displacements are output by batch calling numerical models through scripts. A sparse block diagonal global topological projection matrix is ​​constructed based on the linear inner product of Hilbert space to realize the unified conversion of global three-dimensional displacements into radar LOS displacements. A radial basis function neural network with independent Gaussian kernels in each dimension was constructed. The input was the dam height and multiple sets of LOS displacements, and the output was a complete set of soil mechanical parameters. The network center, kernel width in each dimension, and regularization coefficient were all integrated into a single real chromosome. The negative root mean square error was used as the fitness of the genetic algorithm by leave-one crossover verification. Differentiated Gaussian mutations are set to distinguish between the original space and the logarithmic space. When evolution stalls, the mutation probability is automatically increased. Combined with elite retention, tournament selection, and simulated binary crossover, the population is iteratively optimized. Finally, the network weights are solved by regularized least squares to complete the full training of the GA-RBF surrogate model.

4. The method as described in claim 1, characterized in that, The measured dam height increment dynamically updates the model's surcharge boundary; CR-InSAR processing of satellite imagery filters out noise and extracts the time-series LOS displacement of the dam's corner reflectors; real-time dam height and measured displacement are input into the GA-RBF surrogate model to invert current soil mechanical parameters, including: Based on the measured incremental time-series data of dam height, CR-InSAR satellite radar original images, and coherent deformation dataset of dam area corner reflectors, a real-time inversion calculation framework for soil mechanical parameters was established, and the complete calculation logic of dynamic update of load boundary, time-series deformation noise reduction and extraction, and GA-RBF proxy model inversion calculation was established. For satellite SAR raw image input, a standardized CR-InSAR analytical processing workflow is configured to uniformly complete dam slope geometric coupling noise filtering, time-series cumulative LOS displacement extraction, and effective monitoring point screening preprocessing. The real-time increase in dam height Δh is retrieved, and the LOS displacement sequence measured by the corner reflector is simultaneously matched and screened. The load variables and deformation monitoring data are coupled as input to the GA-RBF proxy model. The network inversion calculation process is traversed to complete the inversion calculation of the complete set of mechanical parameters of the tailings dam soil in the current period, including weight, elastic modulus, Poisson's ratio, cohesion, and internal friction angle.

5. The method as described in claim 4, characterized in that, The updated load and inversion parameters will be fed back to the numerical model, the safety factor will be calculated using the strength reduction method, and early warning will be issued according to the national standard threshold. New monitoring data will be updated using incremental learning to update the surrogate model, narrowing the parameter search range for the next period to ±10%. The working condition verification and optimization are completed based on a dual-driven framework of physical prior and data posterior, forming dynamic monitoring and early warning information for tailings ponds, including: Based on real-time updates of dam load boundaries, inversion of soil mechanical parameters, and time-series SAR monitoring datasets, a closed-loop simulation calculation framework for tailings dam stability evaluation is established, and a complete calculation logic is established for safety factor solution, graded early warning, incremental iteration of proxy model, and adaptive shrinkage of parameter interval. A standardized access and analysis process for numerical models is established to accommodate updated loads and inverted mechanical parameters, and preprocessing for dam stress field reconstruction and LOS displacement comparison verification is completed in a unified manner. The target tailings dam classification safety threshold benchmark is retrieved, and the updated load and soil mechanical parameters are synchronously imported into the numerical model. The strength reduction method is used to traverse the stability calculation process to solve the dam safety factor. The risk level is determined by comparing with the national standard threshold and an early warning is output. Based on newly added SAR monitoring samples, GA-RBF incremental learning iteration is carried out. The current inversion parameters are used as a benchmark to shrink the feasible region of the next period parameters to ±10%. Combined with the dual-drive assimilation framework of physical prior and data posterior, the full-condition verification and optimization are completed, forming dynamic safety monitoring and early warning information for all types of tailings ponds.

6. A tailings dam load-mechanical parameter co-inversion system based on CR-InSAR time series and GA-RBF, characterized in that, The system is used to execute executable instructions to perform the tailings dam load-mechanical parameter co-inversion method based on CR-InSAR time series and GA-RBF as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, include: First processor; And a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute, by executing the executable instructions, a method for co-inversion of load-mechanical parameters of tailings dam body based on CR-InSAR time series and GA-RBF as described in 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 the second processor, it implements the method for collaborative inversion of tailings dam load-mechanical parameters based on CR-InSAR time series and GA-RBF as described in any one of claims 1 to 5.

Images