A method for assimilating multi-physics parameters across a mine by integrating multi-source monitoring data
By employing a multi-source monitoring data collaborative constraint and physical enhancement proxy model approach, the problems of inconsistent mine monitoring data formats and multiple solutions in parameter inversion were solved. This approach enabled the dynamic assimilation and correction of multi-physics parameters across the entire mine, thereby improving the accuracy and efficiency of mine safety and stability analysis.
Patent Information
- Application Number
- CN202611236691.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-11-17
AI Technical Summary
Existing mine monitoring and numerical simulation technologies are insufficient to meet the high-precision and dynamic requirements of safety and stability analysis in deep mines. The data formats of multi-source monitoring are inconsistent and have poor compatibility. The parameter inversion results have serious multiple solutions and high computational costs, making them unsuitable for the safety management and control needs of complex mines.
By employing multi-source monitoring data collaborative constraints, physical augmentation proxy models, adaptive updates, and uncertainty quantification, a reasonable parameter value space is constructed through global sensitivity analysis and physical constraints. A multi-task spatiotemporal deep proxy model is established to perform robust probabilistic multi-objective optimization assimilation and time-varying parameter correction.
It enables dynamic assimilation and correction of multi-physics parameters across the entire mine, improves the application value of monitoring data, ensures the physical consistency and computational efficiency of parameter assimilation results, and meets the real-time requirements of mining engineering practice.
Smart Images

Figure CN123116794A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mine geological disaster monitoring and digital twin modeling technology, specifically involving a method for assimilating multi-physics field parameters of the entire mine by integrating multi-source monitoring data. Background Technology
[0002] As mining operations gradually shift from shallow to deep mining and from open-pit to underground mining, the geological environment of mining areas becomes increasingly complex. The mechanical behavior and seepage characteristics of rock masses exhibit complex features involving multi-scale and multi-field coupling. The emergence and development of geological hazards such as mine slope deformation, goaf collapse, and underground tunnel disturbance are closely related to the dynamic evolution of the stress field, seepage field, and geological structure in the mining area. To accurately grasp the overall stability of the mine and predict hazard development trends, the industry widely adopts multi-source monitoring technologies for comprehensive monitoring of mining areas. Mainstream monitoring methods include GNSS monitoring, InSAR remote sensing monitoring, deep rock mass inclination monitoring, and pore water pressure monitoring. These various monitoring methods can collect massive amounts of mine monitoring data from different spatial scales and temporal resolutions, providing fundamental data support for revealing the dynamic evolution mechanism of mine rock masses and analyzing deformation patterns in mining areas.
[0003] However, current mine monitoring and numerical simulation analysis technologies still have many shortcomings, making it difficult to meet the high-precision and dynamic requirements of safety and stability analysis in deep mines. Firstly, existing mine monitoring data is mostly discrete, surface-level, localized point-like monitoring information, only reflecting local deformation and seepage conditions at monitoring points. It cannot fully characterize the continuous spatial evolution characteristics of underground goaf areas, deep ore bodies, and the entire mining area, resulting in insufficient comprehensiveness and continuity of monitoring data. Simultaneously, significant differences exist in the spatial coordinate references, data sampling frequencies, and monitoring accuracy parameters of different monitoring equipment, leading to inconsistent formats and poor compatibility of multi-source monitoring data. This hinders effective integration and collaborative utilization, significantly reducing the application value of monitoring data. Secondly, the industry currently relies heavily on 3D geological modeling combined with numerical simulation to reconstruct the multi-physics field evolution process in mines. However, existing modeling and simulation technologies lack an efficient, closed-loop parameter feedback and correction mechanism between the 3D numerical model and field-measured monitoring data. The core parameters used in the model, such as rock mechanics parameters and seepage parameters, mostly rely on engineering experience and cannot be accurately calibrated using field-measured data. This leads to a systematic deviation between the numerical model calculation results and the actual deformation and seepage state at the mine site, highlighting the industry pain points of static models and empirical parameter assignment, resulting in low model simulation fidelity.
[0004] In addition, existing multiphysics analysis models for mines suffer from fragmented data processing. Surface monitoring data and deep underground monitoring data are typically processed and analyzed independently on different platforms, making it impossible to achieve coordinated constraints and joint parameter inversion within the same numerical model framework. The limited constraint dimensions of single-type monitoring data prevent accurate and unique calibration of multiple key parameters in the multiphysics coupled model, resulting in severe ambiguity and insufficient reliability in parameter inversion results. Furthermore, high-fidelity forward modeling of multiphysics numerical models for mines is time-consuming, and traditional iterative optimization-based parameter inversion methods require repeated calls to the numerical model for iterative calculations, resulting in a massive computational load and extremely high operational costs, failing to meet the real-time requirements of rapid parameter identification and dynamic model correction in mine engineering practice.
[0005] The aforementioned technical challenges severely limit the application of multi-source monitoring data in scenarios such as dynamic assessment of mine stability, precise early warning of geological disasters, and optimization of mining decisions. This makes it difficult for existing technical systems to meet the safety management needs of deep and complex mines. Therefore, there is an urgent need to provide a method for assimilating multi-physics parameters across the entire mine by integrating multi-source monitoring data. Summary of the Invention
[0006] To address the problems existing in the prior art, this invention provides a method for assimilating multi-physics parameters across a mine by integrating multi-source monitoring data. This method achieves dynamic assimilation and correction of mine rock mechanics and seepage parameters through collaborative constraints of multi-source monitoring data, physical enhancement proxy models, adaptive updates, and uncertainty quantification.
[0007] To achieve the above objectives, this invention provides a method for assimilating multi-physics parameters across a mine by integrating multi-source monitoring data, comprising the following steps: Step 1: Multi-source monitoring data acquisition and assimilation parameter screening; multi-source heterogeneous data covering surface deformation, deep displacement and pore water pressure are acquired by deploying various types of monitoring equipment and standardized preprocessing is performed; at the same time, a reasonable value space for key parameters is constructed by using global sensitivity analysis and physical constraints to achieve dimensionality reduction from high-dimensional candidate parameters to low-dimensional parameters to be assimilated. Step 2: Constructing a parameter-response sample set based on numerical simulation; guided by sensitivity distribution, using weighted optimal Latin hypercube sampling to obtain parameter samples; establishing a three-dimensional geological model including the mining process and performing fluid-structure interaction forward modeling to extract multiple types of physical responses containing spatial distribution characteristics; thereby constructing a spatiotemporally complete parameter-response sample set of the physical field. Step 3: Construction of Physically Enhanced Proxy Model and Uncertainty-Oriented Adaptive Update; Construct a multi-task spatiotemporal deep proxy model that integrates spatial coding and physical constraints; Simultaneously, introduce an active learning mechanism based on the proxy model to predict uncertainty, enabling the multi-task spatiotemporal deep proxy model to dynamically identify parameter drift and adaptively update during the mining process; Step 4: Robust probabilistic multi-objective optimization assimilation and time-varying parameter correction; Based on the multi-physics prediction and its confidence interval output by the surrogate model, a multi-objective optimization assimilation problem is constructed. Using a decomposition-based multi-objective evolutionary algorithm combined with uncertainty propagation, the Pareto optimal solution set and its probabilistic credibility are obtained. Robust time-varying parameter assimilation results are given through stochastic multi-criteria acceptability analysis.
[0008] As a preferred option, the process of multi-source monitoring data acquisition and assimilation parameter screening in step 1 is as follows: S11: Multi-source monitoring data acquisition and preprocessing; S11-1: Deploy multiple types of monitoring equipment in the target mining area to acquire multi-source monitoring data; S11-2: Synchronize the acquired multi-source monitoring data in time and unify the coordinates, and clean the data to obtain a spatiotemporally continuous monitoring dataset; S11-3: Organize the monitoring data according to time series and spatial location, and extract multi-dimensional feature parameters including slope deformation rate, pore pressure change rate and surface deformation gradient; S12: Determination of parameters to be assimilated and construction of the value space; S12-1: Based on geological survey data of the mining area, indoor rock mechanics test results, and in-situ test data, a preliminary set of candidate parameters affecting the multi-physics response of the mine is determined; the candidate parameters include the elastic modulus of the rock mass. Cohesion internal friction angle Permeability coefficient and creep viscosity coefficient ; S12-2: Using a variance-based global sensitivity analysis method, the main effects and total effects sensitivity indices of each candidate parameter on four types of monitoring responses—horizontal surface displacement, vertical surface displacement, deep displacement, and pore water pressure—are quantitatively calculated. Through ranking the total effect indices and analyzing interaction effects, several key parameters that significantly influence multi-physics responses are selected from the candidate parameters, forming a parameter vector to be assimilated. ,in This represents the total number of key parameters selected. S12-3: Based on the statistical distribution of indoor test data and the parameter variation range in historical monitoring data, determine the reasonable lower limit of each parameter to be assimilated. and upper limit A reasonable range of values for each parameter to be assimilated is constructed. ,in ; S12-4: Utilize the empirical relationships of rock mass parameters based on geological strength indices to determine reasonable value constraints between elastic modulus and cohesion parameters, avoid parameter combinations that violate the basic laws of rock mechanics within the value space, and finally construct a value space for parameters to be assimilated with physical consistency constraints.
[0009] As a preferred embodiment, in S11-1, the multi-source monitoring data includes: surface deformation data acquired by InSAR and GNSS, deep displacement data acquired by inclinometer, pore water pressure data acquired by piezometer, and mining disturbance information data.
[0010] As a preferred embodiment, in S11-2, the data cleaning includes noise filtering, outlier removal, and missing data completion; wherein missing data completion adopts a spatiotemporal kriging interpolation method that takes into account the spatiotemporal correlation of multiple physics fields, so as to maintain the spatial continuity and spatial statistical correlation between various physical quantities.
[0011] As a preferred embodiment, the process of constructing the parameter-response sample set based on numerical simulation in step 2 is as follows: S21: Parameter sample selection; Using the sensitivity indices of each parameter obtained in step 1, the sampling weights of the parameter space are determined: more dense sampling probabilities are assigned to dimensions with higher total effect indices; within the parameter value space considering physical consistency constraints, the sensitivity indices of each parameter are used as sampling weights, and optimal Latin hypercube sampling is adopted in the weighted space to make the samples obtain denser sampling resolution in the sensitive dimensions. S22: Establishment of a three-dimensional numerical model; Based on slope radar point cloud, geological exploration borehole columnar section, and mining production succession sequence data, a three-dimensional point cloud database containing major rock strata, fault structures, dynamic morphology of goaf areas, and support structures is established. A three-dimensional geological model reflecting the actual mining sequence is reconstructed using a three-dimensional geological modeling platform, and the model is meshed. At the same time, local mesh densification is implemented around the goaf areas and near monitoring points. S23: Multiphysics forward modeling; S23-1: Input each set of parameter samples selected in S21 into the three-dimensional numerical model, set boundary conditions including geostress, water pressure, mining unloading sequence and rainfall infiltration sequence, perform stress field-seepage field-displacement field coupled forward modeling calculation, and simulate the multi-physics field evolution of the entire mining process under various parameter combinations. S23-2: For each set of parameter samples, extract the surface key point displacement, deep key point displacement, and pore water pressure at the piezometer location. Simultaneously, extract physical response features reflecting spatial distribution variability. These physical response features include the horizontal gradient of the surface deformation field in the east-west and north-south directions, the tilt distribution curve of the deep displacement along the borehole direction, and the extreme values of the spatial gradient of pore water pressure, forming multiple types of numerical response vectors. The tilt distribution curve is the first derivative of the deep key point displacement along the borehole direction. S24: Construction of parameter-response sample set; S24-1: Construct a parameter-response sample database with each set of parameter samples as the input vector and its corresponding multi-type numerical responses as the output vector; S24-2: Divide the parameter-response sample database into a training set and a validation set according to the time series order or a set ratio, and ensure that the training set covers the main working conditions of the entire mining stage after the division.
[0012] As a preferred embodiment, in step 3, the process of constructing the physical enhancement agent model and the uncertainty-oriented adaptive update is as follows: S31: Training of the Physically Enhanced Agent Model; S31-1: Based on the training set data constructed in step 2, construct a multi-task spatiotemporal attention deep neural network proxy model based on the spatial coding module, parameter embedding module, spatiotemporal attention fusion module and multi-task decoding module; The spatial encoding module is used to encode the grid topology and spatial relationships of monitoring points of the 3D geological model using a graph convolutional network to generate spatial embedding features; the parameter embedding module is used to map the parameter vector to be assimilated into parameter latent variables through a fully connected layer; the spatiotemporal attention fusion module is used to use a multi-head attention mechanism to perform feature interaction between spatial embedding and parameter latent variables, as well as between different physical field prediction tasks, to achieve adaptive weighting of parameter influence; the multi-task decoding module is used to simultaneously output spatiotemporal sequence predictions of the surface horizontal displacement field, surface vertical displacement field, deep displacement along the borehole curve, and pore water pressure field; S31-2: Construct a total loss function based on data fitting loss and physical constraint loss; the data fitting loss adopts the mean square error between the predicted values of each physical quantity and the forward simulation values; the physical constraint loss is based on the discrete form of the solid equilibrium equation and the Darcy seepage equation, and substitutes the displacement field and pressure field output by the surrogate model into the control equation, calculates the equation residuals and uses them as penalty terms. During training, cross-validation and early stopping strategies are adopted. When the total loss function converges on the validation set and meets the preset criteria, the model weights are saved to obtain a surrogate model that can replace the time-consuming three-dimensional fluid-structure interaction forward modeling calculation. S32: Uncertainty-oriented adaptive update; S32-1: When new multi-source monitoring data enters the system, a trained surrogate model is used for prediction, and the Monte Carlo Dropout method is used to perform multiple random forward propagations of the predicted values to obtain the mean and variance of each prediction point, which serve as a comprehensive measure of the cognitive uncertainty of the prediction results and the random uncertainty that partially reflects data noise. S32-2: Calculate the residuals between the predicted values and the observed values at each monitoring location; simultaneously, detect concept drift based on the Kullback-Leibler divergence between the probability distributions of consecutive batches of predictions, and determine whether the surrogate model's predictive ability has significantly degraded; when the residuals exceed a preset threshold or the concept drift index rises significantly, the adaptive update process is automatically triggered. First, in regions of high uncertainty and significant concept drift, representative supplementary samples are generated using the forward modeling of the numerical model from step 2. Second, an exponential forgetting factor is introduced to adjust the time weights of the training samples, giving recent samples greater influence. Finally, the surrogate model is incrementally fine-tuned or lightly retrained to update the network weights, enabling the surrogate model to quickly adapt to the new parameter distribution.
[0013] As a preferred option, in step 4, the robust probabilistic multi-objective optimization assimilation and time-varying parameter correction process is as follows: S41: Construction of multi-objective optimization function; S41-1: Based on the monitoring results of InSAR, GNSS, inclinometer and piezometer, the measured time series data of surface horizontal displacement, surface vertical displacement, deep displacement and pore water pressure are obtained respectively. S41-2: With the objective of minimizing the difference between the surrogate model's predicted values and the measured values, a multi-objective vector consisting of four root mean square error sub-objective functions is constructed. ,in , , , These are the surface horizontal displacement error, surface vertical displacement error, deep displacement error, and pore water pressure error, defined as follows:
[0014]
[0015]
[0016]
[0017] In the formula, , , , These are the effective monitoring points for horizontal displacement, vertical displacement, deep displacement, and pore water pressure, respectively. The standard deviation of the monitored values for horizontal displacement at the Earth's surface; This represents the standard deviation of the monitored values for vertical displacement of the Earth's surface. The standard deviation of the deep displacement monitoring values; The standard deviation of the pore water pressure monitoring values; For parameter vectors Below, the surrogate model predicts the first One horizontal displacement of the ground surface; For the first Measured values of horizontal displacement of the ground surface at each measuring point; For parameter vectors Below, the surrogate model predicts the first Vertical displacement of the ground surface; For the first The measured values of the vertical displacement of the ground surface at each measuring point; For parameter vectors Below, the surrogate model predicts the first Displacement of a borehole at a depth; For the first Measured values of deep displacement at each measuring point in the borehole; For parameter vectors Below, the surrogate model predicts the first Pore water pressure; For the first Measured values of pore water pressure at each measuring point; S42: Robust probabilistic multi-objective assimilation solution; A multi-objective optimization search is performed using the decomposition-based multi-objective evolutionary algorithm MOEA / D driven by a proxy model. First, leveraging the uncertainty estimation capability of the surrogate model in S32, for each candidate parameter vector... Multiple random forward propagations are performed to obtain the mean and variance of four sub-objective functions, constructing a probabilistic Pareto front, where each objective vector is described in the form of a confidence interval. During the evolutionary process, the population evolution is guided by non-dominated sorting and probabilistic processing of Pareto dominance relations, while elite preservation and adaptive mutation strategies are adopted to ensure global convergence and population diversity. After the search is completed, a set of Pareto optimal solutions is obtained, with each solution accompanied by a confidence interval in the objective space. Furthermore, the Stochastic Multi-Criterion Acceptability Analysis (SMAA) is used to make robust decisions on these solutions: through Monte Carlo simulation, the central acceptability and global acceptability indices of each solution are calculated in the possible distribution space of the objective function. The parameter combination with high acceptability under multiple uncertainties is selected as the final robust assimilation parameter combination, based on the highest global acceptability.
[0018] Compared with the prior art, the present invention has the following technical advantages: 1. Deep collaboration of multi-source heterogeneous data enables global consistency constraint assimilation of multi-physics parameters across the entire mine. This invention integrates monitoring data from four different spatiotemporal resolutions and physical mechanisms: InSAR surface deformation, GNSS point displacement, deep borehole inclinometer data, and pore water pressure. A multi-objective assimilation system is constructed within a unified coordinate framework and time series. Compared to traditional inversion methods relying solely on single-physics data, the wide-area surface coverage provided by InSAR and the borehole internal distribution curves provided by deep inclinometers form complementary surface-line-point constraints. The four types of responses—horizontal surface displacement, vertical surface displacement, deep displacement, and pore water pressure—simultaneously participate in the assimilation, ensuring that the assimilated parameters match both the macroscopic deformation characteristics of the surface and the deformation and seepage response of deep rock masses. This effectively overcomes the problems of strong ambiguity and poor physical consistency of parameters in single-source data source inversion.
[0019] 2. Global sensitivity analysis and physical consistency constraints jointly drive parameter dimensionality reduction, enabling the rational construction of the parameter space to be assimilated. This invention employs Sobol's global sensitivity analysis method to quantitatively calculate the main effect and total effect sensitivity index of each candidate parameter to four types of monitoring responses. Key parameters are automatically screened through ranking and interaction effect analysis, avoiding the arbitrariness of traditional methods that rely on manual experience to select parameters to be assimilated. Simultaneously, empirical relationships of rock mass parameters based on the Geological Strength Index (GSI) are introduced to establish covariance constraints between parameters such as elastic modulus and cohesion, eliminating parameter combinations that violate the fundamental laws of rock mechanics within the value space. This ensures that subsequent sampling and assimilation always occur within a physically reasonable parameter range, guaranteeing the physical interpretability of the assimilation results from the outset.
[0020] 3. A sensitivity-oriented weighted sampling strategy achieves efficient coverage of the parameter space with limited forward modeling samples. This invention uses the total effect index obtained from Sobol's sensitivity analysis as the sampling weight to drive optimal Latin hypercube sampling, automatically focusing forward modeling resources on sensitive parameter dimensions that have a greater impact on the displacement and seepage fields. Compared to conventional uniform sampling, this strategy effectively improves the mapping accuracy of the surrogate model in the key parameter region with the same number of forward modeling iterations. Furthermore, in addition to extracting key point displacements and pore pressures during forward modeling, it also extracts multiple types of spatial distribution features, such as the horizontal gradient of the surface deformation field, the deep displacement tilt distribution curve, and the spatial gradient extrema of pore pressure. This expands the response information learned by the surrogate model from "points" to "fields," further enriching the information content of the samples.
[0021] 4. A data- and physics-driven multi-task spatiotemporal attention surrogate model that balances fitting accuracy and physical plausibility. The surrogate model constructed in this invention uses a graph convolutional network to encode the spatial topology of a 3D geological model and the spatial relationships of monitoring points. A multi-head attention mechanism enables adaptive feature interaction between parameter embedding and spatial embedding, as well as between different physical field prediction tasks. A multi-task decoder synchronously outputs spatiotemporal sequence predictions of the surface horizontal displacement field, surface vertical displacement field, deep displacement along the pore curve, and pore water pressure field. During training, not only is data fitting loss employed, but the discrete residuals of the solid equilibrium equation and the Darcy seepage equation are also used as physical constraint penalties. This ensures that the surrogate model adheres to physical laws even with limited training samples, effectively suppressing non-physical oscillations or artifacts that may arise from purely data-driven models, thus improving the reliability and generalization ability of extrapolation predictions. This surrogate model can replace time-consuming 3D fluid-structure interaction forward modeling calculations, significantly reducing the computational cost of assimilation iterations while maintaining computational accuracy.
[0022] 5. Uncertainty quantification and concept drift detection-driven proactive adaptive updates ensure the long-term prediction accuracy of the surrogate model throughout the entire mining cycle. This invention utilizes Monte Carlo Dropout to obtain the cognitive uncertainty and some accidental uncertainty in the surrogate model's predictions, and combines this with Kullback-Leibler divergence between the probability distributions of consecutive batch predictions to detect concept drift caused by parameter evolution. Compared to conventional update strategies that rely solely on residual threshold triggers, this mechanism can distinguish between accidental fluctuations caused by data noise and systematic drift caused by the actual evolution of parameters, avoiding unnecessary model retraining triggered by noise, while ensuring timely response when substantial changes occur in rock mass parameters. After triggering the update, an exponential forgetting factor is introduced to time-weight historical samples, enabling the surrogate model to maintain sensitivity to recent working conditions when tracking time-varying parameters, achieving a leap from static offline surrogate to dynamic online surrogate.
[0023] 6. Non-subjectively weighted probabilistic robust multi-objective assimilation quantifies and accommodates uncertainties in multi-source data and models. This invention abandons the traditional method of manually setting weights for multi-objective weighted summation. Instead, it normalizes the root mean square error to a dimensionless form using the standard deviation of each physical quantity's monitored values, eliminating dimensional differences between different physical quantities. Furthermore, it employs the decomposition-based multi-objective evolutionary algorithm MOEA / D to maintain the independence of each physical field objective, avoiding the introduction of biases in the assimilation results due to subjective weights and the omission of non-convex Pareto fronts. Simultaneously, through multiple forward propagations of Monte Carlo Dropout, the uncertainties of the surrogate model are propagated to the target space, constructing a probabilistic Pareto front described by confidence intervals. Stochastic multi-criteria acceptability analysis (SMAA) is used to calculate the acceptability index of each candidate parameter solution in the joint distribution space of uncertainty, selecting the robust optimal solution that maintains high acceptability under various uncertainties. This strategy imbues the final output parameters with credible quantitative information, providing a more prudent and reliable basis for engineering decisions.
[0024] 7. Online incremental assimilation and time-varying parameter tracking enable dynamic perception of rock mass parameters throughout the entire mining cycle. This invention integrates adaptive proxy updating and robust multi-objective assimilation into an online closed loop. As monitoring data arrives sequentially, it continuously executes a cycle of "monitoring, detection, updating, assimilation, and tracking," outputting the time-varying evolution trajectory of the parameter vector to be assimilated and calculating the time-varying rate and trend direction of each parameter. It generates anomaly warnings for parameters with excessive cumulative changes. Simultaneously, the assimilated parameters are substituted back into a three-dimensional numerical model for forward modeling verification. The assimilation accuracy is quantitatively evaluated using relative error, root mean square error, and coefficient of determination, forming a "assimilation-verification-feedback" closed loop. Compared to existing methods that only perform static inversion at a single time section, this mechanism can capture the gradual evolution and sudden jumps in rock mechanics and seepage parameters during mining, providing continuous time-series parameter input for dynamic assessment of mine slope stability and safety early warning.
[0025] 8. A complete closed-loop technology chain, with each step working synergistically to generate overall gain. This invention constructs a complete technology chain of "sensitivity-guided sampling → physical augmentation surrogate modeling → uncertainty-driven proactive updating → probabilistic robust assimilation → online time-varying tracking." The output of each step directly constitutes the input or driving condition for the next step: the results of sensitivity analysis determine the allocation of sampling weights; the sample set obtained from sampling supports the training of the physical augmentation surrogate model; the uncertainty quantification of the surrogate model output provides the trigger basis for adaptive updating and direction guidance for supplementary sampling; the confidence interval information in the probabilistic Pareto front provides distributed input for SMAA robust decision-making; and the bias of assimilation verification is fed back to the adaptive update module to form a closed loop. The steps are not simply sequentially spliced together, but rather involve multi-level causal nesting and information reuse. The overall accuracy, efficiency, and robustness gains exceed the simple sum of the independent effects of each step.
[0026] 9. Wide range of applications and strong engineering applicability. This method can be widely applied to long-term deformation analysis and parameter assimilation of various engineering bodies in mines, such as slopes, spoil heaps, tailings dams and banks, and surrounding rock of underground chambers in open-pit to underground mining areas, providing reliable technical support for mine stability evaluation and disaster prediction.
[0027] This method achieves the transformation of rock mechanics and seepage parameters in the multi-physics field of the entire mine from static empirical values to dynamic data-driven assimilation through the synergy of sensitivity-guided sampling, physical enhancement proxy modeling, uncertainty proactive updating, probabilistic robust assimilation and online time-varying tracking. Attached Figure Description
[0028] Figure 1 This is a flowchart of the assimilation method in this invention; Figure 2 This is a graph showing the fitting results of the training set for the proxy model of this invention; Figure 3 This is a graph showing the fitting results of the proxy model test set in this invention. Detailed Implementation
[0029] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. These embodiments are merely one of the preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention.
[0030] like Figure 1 As shown, this invention provides a method for assimilating multi-physics parameters across a mine by integrating multi-source monitoring data, characterized by comprising the following steps: Step 1: Multi-source monitoring data acquisition and assimilation parameter screening; Multi-source heterogeneous data covering surface deformation, deep displacement and pore water pressure are acquired by deploying various types of monitoring equipment and standardized preprocessing is performed; At the same time, global sensitivity analysis and physical constraints are used to construct a reasonable value space for key parameters, realizing dimensionality reduction from high-dimensional candidate parameters to low-dimensional parameters to be assimilated, providing guidance for subsequent numerical experimental design; As a preferred option, the process of multi-source monitoring data acquisition and assimilation parameter selection is as follows: S11: Multi-source monitoring data acquisition and preprocessing; S11-1: Deploy multiple types of monitoring equipment in the target mining area to acquire multi-source monitoring data; the multi-source monitoring data includes: surface deformation data acquired through InSAR and GNSS, deep displacement data acquired through inclinometers, pore water pressure data acquired through piezometers, and mining disturbance information data. S11-2: Time synchronization and coordinate unification are performed on the acquired multi-source monitoring data, and data cleaning is carried out to obtain a spatiotemporally continuous monitoring dataset; the data cleaning includes noise filtering, outlier removal, and missing data completion; the missing data completion adopts the spatiotemporal kriging interpolation method that takes into account the spatiotemporal correlation of multiple physics fields to maintain the spatial continuity and spatial statistical correlation between various physical quantities. S11-3: Organize the monitoring data according to time series and spatial location, and extract multi-dimensional feature parameters including slope deformation rate, pore pressure change rate and surface deformation gradient for subsequent parameter sensitivity analysis and model training. S12: Determination of parameters to be assimilated and construction of the value space; S12-1: Based on geological survey data of the mining area, indoor rock mechanics test results, and in-situ test data, a preliminary set of candidate parameters affecting the multi-physics response of the mine is determined; the candidate parameters include the elastic modulus of the rock mass. Cohesion internal friction angle Permeability coefficient and creep viscosity coefficient ; S12-2: Using a variance-based global sensitivity analysis method (preferably the Sobol global sensitivity analysis method), the main effects and total effects sensitivity indices of each candidate parameter on the four types of monitoring responses—horizontal surface displacement, vertical surface displacement, deep displacement, and pore water pressure—are quantitatively calculated. Through ranking the total effect indices and analyzing interaction effects, several key parameters that significantly influence the multi-physics response are selected from the candidate parameters, forming a parameter vector to be assimilated. ,in This represents the total number of key parameters selected. S12-3: Based on the statistical distribution of indoor test data and the parameter variation range in historical monitoring data, determine the reasonable lower limit of each parameter to be assimilated. and upper limit A reasonable range of values for each parameter to be assimilated is constructed. ,in ; S12-4: Utilizing the empirical relationships of rock mass parameters based on the Geological Strength Index (GSI), reasonable value constraints are determined among parameters such as elastic modulus and cohesion to avoid parameter combinations that violate the basic laws of rock mechanics within the value space, and finally construct a value space for parameters to be assimilated with physical consistency constraints. This technical solution integrates four types of heterogeneous monitoring data—InSAR, GNSS, inclinometer, and piezometer—to construct a collaborative observation system encompassing the surface, deep, and hydrological domains within a unified spatiotemporal framework, thus addressing the insufficient constraint capacity of a single data source. Sobol's global sensitivity analysis is employed to quantitatively identify key parameters that significantly influence multiphysics responses, avoiding the arbitrariness of manual selection. Furthermore, empirical relationships of rock mass parameters based on GSI are introduced to construct physical consistency constraints, eliminating parameter combinations that violate mechanical laws and ensuring the physical interpretability of the assimilation results from the outset.
[0031] In a preferred embodiment, the specific implementation method for multi-source monitoring data acquisition and assimilation parameter screening is as follows: The following monitoring equipment was deployed in the mining area: InSAR to acquire surface deformation data (covering an area of approximately 4 km). 2 The time span is from January 2025 to June 2026; surface displacement was obtained from 4 GNSS monitoring stations; deep displacement was obtained from 4 deep inclinometers (depth 60-120m); pore water pressure data was obtained from 6 piezometers; and mining disturbance information data, including mining progress, blasting records and rainfall sequences, were collected at the same time.
[0032] The acquired multi-source monitoring data underwent the following preprocessing: random noise in GNSS and InSAR data was filtered using the exponential moving average method; outliers in inclinometer and piezometer data were removed using the 3σ criterion; and spatiotemporal kriging interpolation, considering the spatiotemporal correlation of multiple physics fields, was used to complete the data for missing measurement points. After cleaning, all data were unified to the mining area's independent coordinate system and aligned according to time series to obtain a spatiotemporally continuous monitoring dataset. Furthermore, multi-dimensional characteristic parameters such as slope deformation rate, pore pressure change rate, and surface deformation gradient were extracted for subsequent parameter sensitivity analysis and model training.
[0033] Based on geological survey data of the mining area and indoor rock mechanics test results, five candidate parameters have been preliminarily determined: elastic modulus. Cohesion internal friction angle Permeability coefficient and creep viscosity coefficient The variance-based Sobol's global sensitivity analysis method was used to quantitatively calculate the main effects and total effects sensitivity indices of each candidate parameter on four types of monitoring responses: horizontal surface displacement, vertical surface displacement, deep displacement, and pore water pressure. The analysis results show that the total effect indices of all five parameters are significant, and there are strong interaction effects between elastic modulus and cohesion, and between permeability coefficient and creep viscosity coefficient. Therefore, all of these parameters are listed as parameters to be assimilated, forming a parameter vector to be assimilated. Based on the statistical distribution of indoor test data and historical monitoring data, reasonable lower and upper limits for each parameter were determined: elastic modulus. The cohesive strength is 5–30 GPa. The internal friction angle is 0.2–2.0 MPa. The angle is 20° to 45°, and the permeability coefficient is... 1×10 -7 ~1×10 -5 m / s, creep viscosity coefficient 1×10 14 ~1×10 16 Pa·s.
[0034] Furthermore, by utilizing the empirical relationships of rock mass parameters based on the Geological Strength Index (GSI), a covariance constraint between elastic modulus and cohesion is established to avoid parameter combinations that violate the basic laws of rock mechanics, such as high elastic modulus and low cohesion, within the value space. Finally, a value space for parameters to be assimilated with physical consistency constraints is constructed.
[0035] Step 2: Constructing a parameter-response sample set based on numerical simulation; guided by sensitivity distribution, using weighted optimal Latin hypercube sampling to obtain parameter samples; establishing a three-dimensional geological model including the mining process and performing fluid-structure interaction forward modeling to extract multiple types of physical responses containing spatial distribution characteristics; thereby constructing a spatiotemporally complete parameter-response sample set of the physical field to provide high-quality data for surrogate model training; As a preferred approach, the process of constructing the parameter-response sample set based on numerical simulation is as follows: S21: Parameter sample selection; Using the sensitivity indices of each parameter obtained in step 1, the sampling weights of the parameter space are determined: more dense sampling probabilities are assigned to dimensions with higher total effect indices; within the parameter value space considering physical consistency constraints, the sensitivity indices of each parameter are used as sampling weights, and optimal Latin hypercube sampling is adopted in the weighted space to make the samples obtain denser sampling resolution in the sensitive dimensions. S22: Establishment of a three-dimensional numerical model; Based on spatiotemporal data such as slope radar point clouds, geological exploration borehole columnar sections, and mining production succession sequences, a three-dimensional point cloud database containing major rock strata, fault structures, dynamic morphology of goaf areas, and support structures is established. A three-dimensional geological model reflecting the actual mining sequence is reconstructed using a three-dimensional geological modeling platform, and the model is meshed. At the same time, local mesh densification is implemented around the goaf areas and near monitoring points to ensure the accuracy and efficiency of forward modeling calculations. S23: Multiphysics forward modeling; S23-1: Input each set of parameter samples selected in S21 into the three-dimensional numerical model, set boundary conditions including geostress, water pressure, mining unloading sequence and rainfall infiltration sequence, perform stress field-seepage field-displacement field coupled forward modeling calculation, and simulate the multi-physics field evolution of the entire mining process under various parameter combinations. S23-2: For each set of parameter samples, extract the displacement of key surface points, the displacement of key deep points, and the pore water pressure at the piezometer location. Simultaneously, extract physical response features reflecting spatial distribution variability. These physical response features include the horizontal gradient of the surface deformation field in the east-west and north-south directions, the tilt distribution curve of the deep displacement along the borehole direction, and the extreme values of the spatial gradient of pore water pressure, forming multiple types of numerical response vectors. The tilt distribution curve is the first derivative of the deep key point displacement along the borehole direction. S24: Construction of parameter-response sample set; S24-1: Construct a parameter-response sample database with each set of parameter samples as the input vector and its corresponding multi-type numerical responses as the output vector; S24-2: Divide the parameter-response sample database into training and validation sets according to time series order or a set ratio, and ensure that the training set covers the main working conditions of the entire mining stage after the division, so as to provide data support for subsequent training of proxy models and evaluation of generalization ability; as a preferred option, the parameter-response sample database can be divided into training and validation sets according to a preset ratio (such as 8:2). In this technical solution, guided by the results of sensitivity analysis, the total effect index of each parameter is used as the sampling weight to drive optimal Latin hypercube sampling. This automatically focuses limited forward modeling resources on the sensitive parameter dimension, significantly improving the information density of the samples compared to conventional uniform sampling. In the forward modeling calculation, in addition to extracting key point displacements and pore pressures, multiple spatial distribution features are extracted, including the horizontal gradient of the surface deformation field, the deep displacement tilt distribution curve, and the spatial gradient extrema of pore pressure. This expands the numerical response from a discrete point scalar to a spatially variable field description. Simultaneously, the three-dimensional geological model incorporates the dynamic morphology of the goaf and the mining succession sequence, and implements local mesh refinement in key areas, ensuring the accuracy and efficiency of the full-cycle fluid-structure interaction forward modeling calculation.
[0036] In a preferred embodiment, the specific implementation method for constructing the parameter-response sample set based on numerical simulation is as follows: Using the Sobol' total effect sensitivity index obtained in step one, the sampling weights of the parameter space are determined: for elastic moduli with higher total effect indices... and cohesion The sampling probability is more densely distributed across dimensions. Within the parameter value space considering the GSI physical consistency constraint, the sensitivity index of each parameter is used as the sampling weight. Optimal Latin hypercube sampling is employed in the weighted space, selecting a total of 100 parameter samples to achieve a denser sampling resolution in sensitive dimensions such as elastic modulus and cohesion.
[0037] Based on spatiotemporal data such as slope radar point clouds, geological exploration borehole columnar sections, and mining production succession sequences, a three-dimensional point cloud database was established, including major rock strata, fault structures, dynamic morphology of goaf areas, and support structures. A three-dimensional geological model reflecting the actual mining sequence was reconstructed using a three-dimensional geological modeling platform. Mesh generation was performed on the FLAC3D platform, and local mesh densification was implemented around the goaf areas and near four GNSS monitoring points and four inclinometer boreholes to ensure the accuracy and efficiency of forward modeling calculations.
[0038] One hundred sets of parameter samples were sequentially input into a three-dimensional numerical model. Boundary conditions such as in-situ stress, water pressure, mining unloading sequence, and rainfall infiltration sequence were set. Coupled forward modeling of stress field, seepage field, and displacement field was performed to simulate the multi-physics field evolution during the entire mining process under various parameter combinations. For each set of parameter samples, the following numerical responses were extracted: horizontal and vertical surface displacements at four GNSS monitoring points; deep displacements at each depth measuring point (five measuring points per hole, totaling 20 measuring points) in four inclinometer boreholes; and pore water pressure at six piezometers. Simultaneously, the horizontal gradient of the surface deformation field in the east-west and north-south directions, the slope distribution curve of deep displacement along the borehole direction (i.e., the first derivative of the displacement of key deep points along the borehole direction), and the extreme values of the spatial gradient of pore water pressure were extracted to reflect the spatial distribution variability of physical response characteristics, forming multiple types of numerical response vectors.
[0039] A parameter-response sample database is constructed, using each set of parameter samples as the input vector and its corresponding multi-type numerical responses as the output vector. The input vector is a 5-dimensional parameter vector to be assimilated. The output vector is the core response vector. Including spatial distribution characteristic responses, a total of 100 samples were collected. The database was divided into a training set and a validation set in time series order (ratio approximately 8:2) to ensure that the training set covered the main working conditions throughout the entire mining phase.
[0040] Step 3: Construction of Physically Enhanced Proxy Model and Uncertainty-Driven Adaptive Update; Construct a multi-task spatiotemporal deep proxy model that integrates spatial coding and physical constraints to achieve rapid and accurate prediction of multi-physics responses; At the same time, introduce an active learning mechanism based on the uncertainty of the proxy model prediction, so that the multi-task spatiotemporal deep proxy model has the ability to dynamically identify parameter drift and adaptively update during the sampling process, maintaining long-term prediction accuracy. As a preferred approach, the process of constructing the physical augmentation agent model and the uncertainty-oriented adaptive update is as follows: S31: Training of the Physically Enhanced Agent Model; S31-1: Based on the training set data constructed in step 2, construct a multi-task spatiotemporal attention deep neural network proxy model based on the spatial coding module, parameter embedding module, spatiotemporal attention fusion module and multi-task decoding module; The spatial encoding module is used to encode the grid topology and spatial relationships of monitoring points of the 3D geological model using a graph convolutional network to generate spatial embedding features; the parameter embedding module is used to map the parameter vector to be assimilated into parameter latent variables through a fully connected layer; the spatiotemporal attention fusion module is used to use a multi-head attention mechanism to perform feature interaction between spatial embedding and parameter latent variables, as well as between different physical field prediction tasks, to achieve adaptive weighting of parameter influence; the multi-task decoding module is used to simultaneously output spatiotemporal sequence predictions of the surface horizontal displacement field, surface vertical displacement field, deep displacement along the borehole curve, and pore water pressure field; S31-2: Construct a total loss function based on data fitting loss and physical constraint loss; the data fitting loss adopts the mean square error between the predicted values of each physical quantity and the forward simulation values; the physical constraint loss is based on the discrete form of the solid equilibrium equation and the Darcy seepage equation, and substitutes the displacement field and pressure field output by the surrogate model into the control equation, calculates the equation residuals and uses them as penalty terms. During training, cross-validation and early stopping strategies are adopted. When the total loss function converges on the validation set and meets the preset criteria, the model weights are saved to obtain a surrogate model that can replace the time-consuming three-dimensional fluid-structure interaction forward modeling calculation. S32: Uncertainty-oriented adaptive update; S32-1: During the mining process, rock mass parameters will change over time with unloading damage and aging deformation. When new multi-source monitoring data enters the system, a trained surrogate model is used for prediction, and the Monte Carlo Dropout method is used to perform multiple random forward propagations of the predicted values to obtain the mean and variance of each prediction point, which serve as a comprehensive measure of the cognitive uncertainty of the prediction results and part of the random uncertainty reflecting data noise. S32-2: Calculate the residuals between the predicted values and the observed values at each monitoring location; simultaneously, detect concept drift based on the Kullback-Leibler divergence between the probability distributions of consecutive batches of predictions, and determine whether the surrogate model's predictive ability has significantly degraded (possibly due to parameter evolution exceeding the current effective domain); when the residuals exceed a preset threshold or the concept drift index rises significantly (the rise exceeds the set rise threshold), automatically trigger the adaptive update process: First, in regions of high uncertainty and significant concept drift, representative supplementary samples are generated using forward modeling from step 2. Secondly, an exponential forgetting factor is introduced to adjust the time weight of training samples, so that recent samples have a higher influence. Finally, incremental fine-tuning or lightweight retraining is performed on the surrogate model to update the network weights, enabling the surrogate model to quickly adapt to the new parameter distribution. Through this proactive update loop of "detection-supplementation-adjustment", the prediction accuracy of the surrogate model can be maintained continuously during long-term monitoring. In this technical solution, regarding the construction of the surrogate model, a graph convolutional network is used to encode the spatial topology of the 3D geological model. A multi-head attention mechanism is used to realize the feature interaction between parameter embedding and spatial embedding, as well as cross-physics prediction tasks. A multi-task decoder synchronously outputs the spatiotemporal sequence predictions of four types of physics fields, realizing a customized surrogate architecture design for multi-physics coupling problems. During training, the discrete residuals of the solid equilibrium equation and the Darcy seepage equation are introduced as physical constraint penalty terms, so that the model maintains physical rationality under limited sample conditions and effectively suppresses the extrapolation distortion that may be generated by the pure data-driven model. In terms of adaptive update, Monte Carlo Dropout is used to quantify prediction uncertainty. Combined with KL divergence between consecutive batch prediction probability distributions to detect concept drift, the update is triggered by the dual conditions of "residual exceeding the threshold + significant increase in drift". The exponential forgetting factor is used to adjust the sample time weight and incremental model retraining to form an active update closed loop of "detection-supplementation-adjustment", realizing the leap of the surrogate model from static offline to dynamic online, and from passive retraining to active adaptation.
[0041] In a preferred embodiment, the specific implementation of the physical augmentation agent model construction and uncertainty-oriented adaptive update is as follows: Based on the training set data constructed in step 2, a multi-task spatiotemporal attention deep neural network surrogate model is constructed. The spatial encoding module uses a graph convolutional network to encode the grid topology of the 3D geological model and the spatial relationship of monitoring points to generate spatial embedding features; the parameter embedding module maps the 5-dimensional parameter vector to be assimilated into parameter latent variables through a fully connected layer; the spatiotemporal attention fusion module uses a multi-head attention mechanism (with 4 heads) to perform feature interaction between spatial embedding and parameter latent variables, as well as between displacement field and seepage field prediction tasks; the multi-task decoding module simultaneously outputs spatiotemporal sequence predictions of the surface horizontal displacement field, surface vertical displacement field, deep displacement along the borehole curve, and pore water pressure field.
[0042] The total training loss function consists of two parts: data fitting loss and physical constraint loss. The data fitting loss uses the mean square error between the predicted and forward-simulated values of each physical quantity. The physical constraint loss is based on the discrete forms of the solid equilibrium equation and the Darcy flow equation. The displacement and pressure fields output by the surrogate model are substituted into the governing equations, and the residuals are calculated and used as a penalty term. During training, a 5-fold cross-validation and early stopping strategy (20 early stopping rounds) are used. The Adam optimizer is employed, with an initial learning rate of 0.001. After training, the surrogate model's coefficient of determination R on the validation set is... 2 With a resolution of 0.96 or higher, a single prediction takes less than 0.1 seconds, which can replace a single FLAC3D forward modeling calculation (approximately 4 hours).
[0043] When new multi-source monitoring data enters the system, a trained surrogate model is used for prediction, and the Monte Carlo Dropout method (with a Dropout ratio of 0.1 and 50 forward propagations) is employed to obtain the mean and variance of each prediction point. The residuals between the predicted values and the observed values at each monitoring location are calculated. Simultaneously, concept drift is detected based on the Kullback-Leibler divergence between the probability distributions of consecutive batches (with a monthly window). When the residual exceeds a preset threshold (15% of the measured value in this embodiment) or the concept drift index increases beyond a set threshold, an adaptive update process is automatically triggered: First, in high-uncertainty regions and regions with significant concept drift, representative supplementary samples are generated using the FLAC3D numerical model forward modeling; second, an exponential forgetting factor (set to 0.9 in this embodiment) is introduced to adjust the time weights of the training samples; finally, the surrogate model is incrementally fine-tuned, and the network weights are updated to allow the model to quickly adapt to the new parameter distribution.
[0044] Step 4: Robust probabilistic multi-objective optimization assimilation and time-varying parameter correction; Based on the multi-physics prediction and its confidence interval output by the surrogate model, a multi-objective optimization assimilation problem is constructed. The Pareto optimal solution set and its probabilistic credibility are obtained by using a decomposition-based multi-objective evolutionary algorithm combined with uncertainty propagation. Robust time-varying parameter assimilation results are given through stochastic multi-criteria acceptability analysis. As a preferred option, the process of robust probabilistic multi-objective parameter assimilation and time-varying correction is as follows: S41: Construction of multi-objective optimization function; S41-1: Based on the monitoring results of InSAR, GNSS, inclinometer and piezometer, the measured time series data of surface horizontal displacement, surface vertical displacement, deep displacement and pore water pressure are obtained respectively. S41-2: With the objective of minimizing the difference between the surrogate model's predicted values and the measured values, a multi-objective vector consisting of four root mean square error sub-objective functions is constructed. ,in , , , These are the surface horizontal displacement error, surface vertical displacement error, deep displacement error, and pore water pressure error, defined as follows:
[0045]
[0046]
[0047]
[0048] In the formula, , , , These are the effective monitoring points for horizontal displacement, vertical displacement, deep displacement, and pore water pressure, respectively. The standard deviation of the monitored values for horizontal displacement at the Earth's surface; This represents the standard deviation of the monitored values for vertical displacement of the Earth's surface. The standard deviation of the deep displacement monitoring values; The standard deviation of the pore water pressure monitoring values; , , and These are the standard deviations of the corresponding physical quantity monitoring values obtained based on historical monitoring data, used to normalize the root mean square error into a dimensionless form; For parameter vectors Below, the surrogate model predicts the first One horizontal displacement of the ground surface; For the first The measured values of the horizontal displacement of the ground surface at each measuring point were obtained through InSAR or GNSS. For parameter vectors Below, the surrogate model predicts the first Vertical displacement of the ground surface; For the first The measured values of the vertical displacement of the ground surface at each measuring point were obtained through InSAR or GNSS. For parameter vectors Below, the surrogate model predicts the first Displacement of a borehole at a depth; For the first The measured values of the deep displacement of each measuring point in the borehole were obtained using an inclinometer. For parameter vectors Below, the surrogate model predicts the first Pore water pressure; For the first The measured values of pore water pressure at each measuring point were obtained using a piezometer. This multi-objective vector does not undergo subjective weighted synthesis, but maintains the independent objectives of each physical field to fully reflect the trade-offs between multi-source data; S42: Robust probabilistic multi-objective assimilation solution; A multi-objective optimization search is performed using the decomposition-based multi-objective evolutionary algorithm MOEA / D driven by a proxy model. First, leveraging the uncertainty estimation capability of the surrogate model in S32, for each candidate parameter vector... Multiple random forward propagations are performed to obtain the mean and variance of four sub-objective functions, constructing a probabilistic Pareto front, where each objective vector is described in the form of a confidence interval. During the evolutionary process, the population evolution is guided by non-dominated sorting and probabilistic processing of Pareto dominance relations (such as based on expected hypervolume or probabilistic dominance criteria), while elite preservation and adaptive mutation strategies are adopted to ensure global convergence and population diversity. After the search is completed, a set of Pareto optimal solutions is obtained, with each solution accompanied by a confidence interval in the objective space. Furthermore, the Stochastic Multi-Criterion Acceptability Analysis (SMAA) is used to make robust decisions on these solutions: through Monte Carlo simulation, the central acceptability, global acceptability, and other indices of each solution are calculated in the possible distribution space of the objective function. The parameter combination with the highest global acceptability is selected as the final robust assimilation parameter combination, based on the criterion of the highest global acceptability.
[0049] This parameter not only fits various monitoring data, but also quantifies and accommodates the uncertainties of the data and models, providing more reliable parameter input for subsequent mine stability analysis and early warning.
[0050] S43: Verification of assimilation results; The final optimal parameter combination obtained by SMAA robust decision-making in S42 is... The input is fed into the three-dimensional numerical model established in step two. The forward modeling calculation is then performed again according to the multiphysics coupling forward modeling process described in S23. Simulated values of surface horizontal displacement, surface vertical displacement, deep displacement, and pore water pressure at each monitoring point are extracted and compared point-by-point with the preprocessed field measured data from S11. The root mean square error and coefficient of determination for each type of response are calculated to verify the degree of fit between the assimilation parameters and the multiphysics monitoring data. If the values of each sub-objective function are within the acceptable engineering range, the rationality and consistency of the assimilation results are confirmed. If the deviation exceeds the preset allowable limit, the process returns to S41 to adjust the weight distribution of each monitoring data point in the sub-objective function or to re-examine the physical constraint settings of the parameter value space in S12. After iterative improvement, the assimilation process is executed again until the assimilation parameters meet the engineering accuracy requirements in the multiphysics response, ensuring the reliability and engineering applicability of the final output parameters.
[0051] In this technical solution, the root mean square error is normalized to a dimensionless form using the standard deviation of each physical quantity's monitoring value, eliminating the dimensional difference between displacement and pore pressure. The multi-objective evolutionary algorithm MOEA / D, based on decomposition, maintains the independence of each physical field objective, avoiding the bias introduced by subjective weighting in traditional weighted summation methods and the omission of non-convex Pareto fronts. Simultaneously, the uncertainty of the surrogate model is transmitted to the target space through multiple random forward propagations, constructing a probabilistic Pareto front described by confidence intervals. The stochastic multi-criteria acceptability analysis (SMAA) is used to select the robust optimal parameter combination with the highest global acceptability in the joint uncertainty distribution space, achieving a paradigm shift from deterministic point estimation to probabilistic robust assimilation. Finally, the assimilated parameters are substituted back into the three-dimensional numerical model for forward modeling verification. The fitting accuracy is compared point-by-point using quantitative indicators, forming a "assimilation-verification-feedback" closed loop to ensure that the output parameters meet engineering reliability requirements in multi-physics responses.
[0052] In a preferred embodiment, the specific implementation of robust probabilistic multi-objective optimization assimilation and time-varying parameter correction is as follows: Based on the latest monitoring results from InSAR, GNSS, inclinometer, and piezometer, measured time-series data of surface horizontal displacement, surface vertical displacement, deep displacement, and pore water pressure were obtained. A multi-objective vector was constructed, consisting of four normalized root mean square error sub-objective functions: ; Standard deviation of each physical quantity monitoring value , , , The root mean square error is normalized to a dimensionless form (where... , , and These are the standard deviations obtained statistically from historical monitoring data. This multi-objective vector does not undergo subjective weighting and synthesis, thus maintaining the independent objectives of each physical field.
[0053] A decomposition-based multi-objective evolutionary algorithm, MOEA / D, is used to drive a surrogate model for multi-objective optimization search. The population size is set to 100, the maximum number of iterations is set to 200, the crossover probability is 0.9, and the mutation probability is 0.1. The Monte Carlo Dropout uncertainty estimation capability of the surrogate model is utilized to perform multi-objective optimization search on each candidate parameter vector. Fifty random forward propagations were performed to obtain the mean and variance of four sub-objective functions, constructing a probabilistic Pareto front, where each objective vector is described in the form of a confidence interval. During the evolutionary process, non-dominated ranking and a probabilistic dominance criterion based on expected hypervolume guided population evolution, while elite preservation and adaptive mutation strategies were employed. After the search, a set of Pareto optimal solutions was obtained, each solution accompanied by a confidence interval in the objective space. Furthermore, robust decision-making was performed using stochastic multi-criteria acceptability analysis (SMAA): through Monte Carlo simulations, the central and global acceptability of each solution were calculated in the possible distribution space of the objective function; the final robust optimal assimilation parameter was selected based on the highest global acceptability.
[0054] The optimal parameters obtained by robust decision-making using SMAA in this embodiment are: , .
[0055] The obtained optimal parameter combination was substituted back into the FLAC3D three-dimensional numerical model, and forward modeling was performed according to the actual mining progress and rainfall infiltration conditions at the corresponding time. The simulated values of four types of physical quantities at each monitoring point were extracted and compared with the field measured data point by point. The calculation results show that the average error of the horizontal displacement of the surface is 4.2%, the average error of the vertical displacement of the surface is 3.8%, the average error of the deep displacement is 5.1%, and the average error of the pore water pressure is 6.3%, all of which meet the engineering accuracy requirements (<10%), verifying the rationality and reliability of the assimilation results.
[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Those skilled in the art can make various modifications, substitutions, and variations to these embodiments without departing from the principles and spirit of the invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A full-mine multi-physical field parameter assimilation method for fusing multi-source monitoring data, characterized in that, Includes the following steps: Step 1: Multi-source monitoring data acquisition and assimilation parameter selection; Multi-source heterogeneous data covering surface deformation, deep displacement, and pore water pressure were acquired by deploying various types of monitoring equipment and standardized preprocessing was performed. At the same time, a reasonable value space for key parameters was constructed by using global sensitivity analysis and physical constraints to achieve dimensionality reduction from high-dimensional candidate parameters to low-dimensional parameters to be assimilated. Step 2: Constructing a parameter-response sample set based on numerical simulation; guided by sensitivity distribution, using weighted optimal Latin hypercube sampling to obtain parameter samples; establishing a three-dimensional geological model including the mining process and performing fluid-structure interaction forward modeling to extract multiple types of physical responses containing spatial distribution characteristics; thereby constructing a spatiotemporally complete parameter-response sample set of the physical field. Step 3: Construction of Physically Enhanced Proxy Model and Uncertainty-Oriented Adaptive Update; Construct a multi-task spatiotemporal deep proxy model that integrates spatial coding and physical constraints; Simultaneously, introduce an active learning mechanism based on the proxy model to predict uncertainty, enabling the multi-task spatiotemporal deep proxy model to dynamically identify parameter drift and adaptively update during the mining process; Step 4: Robust probabilistic multi-objective optimization assimilation and time-varying parameter correction; Based on the multi-physics prediction and its confidence interval output by the surrogate model, a multi-objective optimization assimilation problem is constructed. Using a decomposition-based multi-objective evolutionary algorithm combined with uncertainty propagation, the Pareto optimal solution set and its probabilistic credibility are obtained. Robust time-varying parameter assimilation results are given through stochastic multi-criteria acceptability analysis.
2. The full-mine multi-physical field parameter assimilation method of fusing multi-source monitoring data according to claim 1, characterized in that, In step 1, the process of multi-source monitoring data acquisition and assimilation parameter selection is as follows: S11: Multi-source monitoring data acquisition and preprocessing; S11-1: Deploy multiple types of monitoring equipment in the target mining area to acquire multi-source monitoring data; S11-2: Synchronize the acquired multi-source monitoring data in time and unify the coordinates, and clean the data to obtain a spatiotemporally continuous monitoring dataset; S11-3: Organize the monitoring data according to time series and spatial location, and extract multi-dimensional feature parameters including slope deformation rate, pore pressure change rate and surface deformation gradient; S12: Determination of parameters to be assimilated and construction of the value space; S12-1: According to the mine geological survey data, indoor rock mechanics test results and field in-situ test data, a candidate parameter set affecting the multi-physical field response of the mine is preliminarily determined; the candidate parameters include rock elastic modulus , cohesion , internal friction angle , permeability coefficient and creep viscosity coefficient ; S12-2: Using the variance-based global sensitivity analysis method, the main effect and total effect sensitivity index of each candidate parameter on the four types of monitoring responses of surface horizontal displacement, surface vertical displacement, deep displacement and pore water pressure are quantitatively calculated; through the sorting of total effect index and interaction effect analysis, a number of key parameters which have significant influence on multiple physical field responses are selected from the candidate parameters to form the assimilation parameter vector wherein is the total number of key parameters screened out S12-3: Based on the statistical distribution of indoor test data and the parameter variation range in historical monitoring data, determine the reasonable lower limit of each parameter to be assimilated. and upper limit A reasonable range of values for each parameter to be assimilated is constructed. ,in ; S12-4: Utilize the empirical relationships of rock mass parameters based on geological strength indices to determine reasonable value constraints between elastic modulus and cohesion parameters, avoid parameter combinations that violate the basic laws of rock mechanics within the value space, and finally construct a value space for parameters to be assimilated with physical consistency constraints.
3. The method for assimilating multi-physics parameters across a mine by fusing multi-source monitoring data according to claim 2, characterized in that, In S11-1, the multi-source monitoring data includes: surface deformation data obtained through InSAR and GNSS, deep displacement data obtained through inclinometers, pore water pressure data obtained through piezometers, and mining disturbance information data.
4. The method for assimilating multi-physics parameters across a mine by fusing multi-source monitoring data according to claim 2, characterized in that, In S11-2, the data cleaning includes noise filtering, outlier removal, and missing data completion; the missing data completion adopts the spatiotemporal kriging interpolation method that takes into account the spatiotemporal correlation of multiple physics fields, so as to maintain the spatial continuity and spatial statistical correlation between various physical quantities.
5. The method for assimilating multi-physics parameters across a mine by integrating multi-source monitoring data according to claim 1, characterized in that, In step 2, the process of constructing the parameter-response sample set based on numerical simulation is as follows: S21: Parameter sample selection; Using the sensitivity indices of each parameter obtained in step 1, the sampling weights of the parameter space are determined: more dense sampling probabilities are assigned to dimensions with higher total effect indices; within the parameter value space considering physical consistency constraints, the sensitivity indices of each parameter are used as sampling weights, and optimal Latin hypercube sampling is adopted in the weighted space to make the samples obtain denser sampling resolution in the sensitive dimensions. S22: Establishment of a three-dimensional numerical model; Based on slope radar point cloud, geological exploration borehole columnar section, and mining production succession sequence data, a three-dimensional point cloud database containing major rock strata, fault structures, dynamic morphology of goaf areas, and support structures is established. A three-dimensional geological model reflecting the actual mining sequence is reconstructed using a three-dimensional geological modeling platform, and the model is meshed. At the same time, local mesh densification is implemented around the goaf areas and near monitoring points. S23: Multiphysics forward modeling; S23-1: Input each set of parameter samples selected in S21 into the three-dimensional numerical model, set boundary conditions including geostress, water pressure, mining unloading sequence and rainfall infiltration sequence, perform stress field-seepage field-displacement field coupled forward modeling calculation, and simulate the multi-physics field evolution of the entire mining process under various parameter combinations. S23-2: For each set of parameter samples, extract the surface key point displacement, deep key point displacement, and pore water pressure at the piezometer location. Simultaneously, extract physical response features reflecting spatial distribution variability. These physical response features include the horizontal gradient of the surface deformation field in the east-west and north-south directions, the tilt distribution curve of the deep displacement along the borehole direction, and the extreme values of the spatial gradient of pore water pressure, forming multiple types of numerical response vectors. The tilt distribution curve is the first derivative of the deep key point displacement along the borehole direction. S24: Construction of parameter-response sample set; S24-1: Construct a parameter-response sample database with each set of parameter samples as the input vector and its corresponding multi-type numerical responses as the output vector; S24-2: Divide the parameter-response sample database into a training set and a validation set according to the time series order or a set ratio, and ensure that the training set covers the main working conditions of the entire mining stage after the division.
6. The method for assimilating multi-physics parameters across a mine by fusing multi-source monitoring data according to claim 1, characterized in that, In step 3, the process of constructing the physical augmentation agent model and the uncertainty-oriented adaptive update is as follows: S31: Training of the Physically Enhanced Agent Model; S31-1: Based on the training set data constructed in step 2, construct a multi-task spatiotemporal attention deep neural network proxy model based on the spatial coding module, parameter embedding module, spatiotemporal attention fusion module and multi-task decoding module; The spatial encoding module is used to encode the grid topology and spatial relationships of monitoring points of the three-dimensional geological model using a graph convolutional network, thereby generating spatial embedding features; The parameter embedding module is used to map the parameter vector to be assimilated into parameter latent variables through a fully connected layer; the spatiotemporal attention fusion module is used to use a multi-head attention mechanism to perform feature interaction between spatial embedding and parameter latent variables, as well as between different physical field prediction tasks, to achieve adaptive weighting of parameter influence; the multi-task decoding module is used to simultaneously output spatiotemporal sequence predictions of the surface horizontal displacement field, surface vertical displacement field, deep displacement along the pore curve, and pore water pressure field. S31-2: Construct a total loss function based on data fitting loss and physical constraint loss; the data fitting loss adopts the mean square error between the predicted values of each physical quantity and the forward simulation values; the physical constraint loss is based on the discrete form of the solid equilibrium equation and the Darcy seepage equation, and substitutes the displacement field and pressure field output by the surrogate model into the control equation, calculates the equation residuals and uses them as penalty terms. During training, cross-validation and early stopping strategies are adopted. When the total loss function converges on the validation set and meets the preset criteria, the model weights are saved to obtain a surrogate model that can replace the time-consuming three-dimensional fluid-structure interaction forward modeling calculation. S32: Uncertainty-oriented adaptive update; S32-1: When new multi-source monitoring data enters the system, a trained surrogate model is used for prediction, and the Monte Carlo Dropout method is used to perform multiple random forward propagations of the predicted values to obtain the mean and variance of each prediction point, which serve as a comprehensive measure of the cognitive uncertainty of the prediction results and the random uncertainty that partially reflects data noise. S32-2: Calculate the residuals between the predicted values and the observed values at each monitoring location; simultaneously, detect concept drift based on the Kullback-Leibler divergence between the probability distributions of consecutive batches of predictions, and determine whether the surrogate model's predictive ability has significantly degraded; when the residuals exceed a preset threshold or the concept drift index rises significantly, the adaptive update process is automatically triggered. First, in regions of high uncertainty and significant concept drift, representative supplementary samples are generated using the forward modeling of the numerical model from step 2. Second, an exponential forgetting factor is introduced to adjust the time weights of the training samples, giving recent samples greater influence. Finally, the surrogate model is incrementally fine-tuned or lightly retrained to update the network weights, enabling the surrogate model to quickly adapt to the new parameter distribution.
7. The method for assimilating multi-physics parameters across a mine by fusing multi-source monitoring data according to claim 1, characterized in that, In step 4, the robust probabilistic multi-objective optimization assimilation and time-varying parameter correction process is as follows: S41: Construction of multi-objective optimization function; S41-1: Based on the monitoring results of InSAR, GNSS, inclinometer and piezometer, the measured time series data of surface horizontal displacement, surface vertical displacement, deep displacement and pore water pressure are obtained respectively. S41-2: With the objective of minimizing the difference between the surrogate model's predicted values and the measured values, a multi-objective vector consisting of four root mean square error sub-objective functions is constructed. ,in , , , These are the surface horizontal displacement error, surface vertical displacement error, deep displacement error, and pore water pressure error, defined as follows: In the formula, , , , These are the effective monitoring points for horizontal displacement, vertical displacement, deep displacement, and pore water pressure, respectively. The standard deviation of the monitored values for horizontal displacement at the Earth's surface; This represents the standard deviation of the monitored values for vertical displacement of the Earth's surface. The standard deviation of the deep displacement monitoring values; The standard deviation of the pore water pressure monitoring values; For parameter vectors Below, the surrogate model predicts the first One horizontal displacement of the ground surface; For the first Measured values of horizontal displacement of the ground surface at each measuring point; For parameter vectors Below, the surrogate model predicts the first... Vertical displacement of the ground surface; For the first The measured values of the vertical displacement of the ground surface at each measuring point; For parameter vectors Below, the surrogate model predicts the first Displacement of a borehole at a depth; For the first Measured values of deep displacement at each measuring point in the borehole; For parameter vectors Below, the surrogate model predicts the first Pore water pressure; For the first Measured values of pore water pressure at each measuring point; S42: Robust probabilistic multi-objective assimilation solution; A multi-objective optimization search is performed using the decomposition-based multi-objective evolutionary algorithm MOEA / D driven by a proxy model. First, leveraging the uncertainty estimation capability of the surrogate model in S32, for each candidate parameter vector... Multiple random forward propagations are performed to obtain the mean and variance of four sub-objective functions, constructing a probabilistic Pareto front, where each objective vector is described in the form of a confidence interval. During the evolutionary process, the population evolution is guided by non-dominated sorting and probabilistic processing of Pareto dominance relations, while elite preservation and adaptive mutation strategies are adopted to ensure global convergence and population diversity. After the search is completed, a set of Pareto optimal solutions is obtained, with each solution accompanied by a confidence interval in the objective space. Furthermore, the Stochastic Multi-Criterion Acceptability Analysis (SMAA) is used to make robust decisions on these solutions: through Monte Carlo simulation, the central acceptability and global acceptability indices of each solution are calculated in the possible distribution space of the objective function. The parameter combination with high acceptability under multiple uncertainties is selected as the final robust assimilation parameter combination, based on the highest global acceptability.
