Geological disaster early warning method, system, device and medium based on digital twinning

Abstract

This application relates to a method, system, equipment, and medium for geological disaster early warning based on digital twins. The method includes: establishing an initial digital twin model based on multi-source monitoring data; extracting state vectors from the current digital twin model and generating a path set containing optimal and branching evolution paths through physical information deep reinforcement learning; updating the path set as a member of the set using a path trajectory similarity-weighted ensemble Kalman filter algorithm to obtain a posterior set of members and a calibrated digital twin model; extracting damage eigenstate variables from the calibrated digital twin model, calculating and classifying the instability probability, and generating graded and zoned early warning results through spatial-temporal dual constraints. This method improves the accuracy and timeliness of geological disaster early warning by integrating physical information deep reinforcement learning with path trajectory similarity-weighted ensemble Kalman filtering.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method, system, equipment and medium for geological disaster early warning based on digital twins. Background Technology

[0002] Geological disasters such as landslides and collapses are characterized by complex evolution processes, diverse triggering factors, and sudden occurrence, posing a serious threat to engineering construction, resident safety, and infrastructure operation and maintenance in mountainous areas. Developing high-precision, dynamic early warning technologies is a core requirement for improving geological disaster prevention and control capabilities. In recent years, digital twin technology, with its high-fidelity modeling and dynamic mapping capabilities for physical entities, has been widely applied in the field of geological disaster early warning. By constructing mechanical and geometric models of geological bodies, it has achieved digital reproduction of the disaster evolution process and trend prediction.

[0003] However, existing digital twin-based geological disaster early warning methods still suffer from several technical shortcomings. Firstly, traditional digital twin models often employ equal-weighted ensemble Kalman filtering algorithms for state updates, failing to consider the long-term matching differences between the evolution paths of ensemble members and measured data. The lack of targeted weight allocation leads to poor data assimilation, and the model is prone to state drift due to observational noise or evolutionary bias, failing to accurately reproduce the true mechanical state of the geological body. Secondly, existing geological disaster early warning methods lack differentiated state processing logic for initial and subsequent cyclical early warning scenarios. Specifically, the initial warning lacks historical assimilation data support, making it impossible to obtain a reliable initial state. Furthermore, subsequent warnings fail to fully utilize historical posterior states and evolutionary path information, resulting in data waste and reduced continuity and accuracy of state extrapolation. In addition, existing early warning grading relies heavily on threshold determinations of single-moment instability probabilities, neglecting to incorporate spatial connectivity constraints of geological units and temporal continuity verification of warning levels. This makes them susceptible to false alarms due to instantaneous anomalies or noise disturbances in local units, resulting in insufficient spatial-temporal consistency of warning results and hindering reliable tiered and zonal early warning decisions. Summary of the Invention

[0004] Therefore, it is necessary to provide methods, systems, equipment and media for geological disaster early warning based on digital twins to address the above-mentioned technical problems, aiming to improve the accuracy of digital twin model status updates and the continuity of evolutionary inference, and enhance the spatial-temporal consistency and hierarchical and zoning reliability of early warning results.

[0005] Firstly, this application provides a geological disaster early warning method based on digital twins, including:

[0006] S1: Based on the multi-source monitoring data of the geological disaster monitoring area, an initial digital twin model is established. The initial digital twin model includes discretized finite element mesh elements, the binding interface between each finite element mesh element and the corresponding monitoring data, and the initial mechanical parameter field, initial stress field and initial displacement field corresponding to each finite element mesh element.

[0007] S2: Extract the state vector, including the field parameters of each finite element mesh element, from the current digital twin model at the current moment; using the state vector as the starting state, obtain the evolution path set through the physical information deep reinforcement learning model; the evolution path set includes an optimal evolution path and multiple branch evolution paths, and each evolution path in the evolution path set contains a discrete sequence of state vectors from the current moment to the prediction time window and the evolution action corresponding to each time step; where, when the current moment is the first warning, the current digital twin model is the initial digital twin model;

[0008] S3: The evolution paths within the evolution path set are used as the current set members of the ensemble Kalman filter; based on the monitoring data received at the current moment, the predicted state of each set member at the current moment is dynamically updated using the ensemble Kalman filter algorithm with path trajectory similarity weighting, to obtain the posterior set of set members and the calibrated digital twin model. The posterior set of set members contains the posterior state vector of each set member at the current moment and its corresponding posterior weight.

[0009] S4: Extract the damage intrinsic state variables of each finite element mesh element from the calibrated digital twin model; calculate the instability probability of each finite element mesh element based on the posterior set membership; classify the instability probability to obtain the preliminary warning level of each finite element mesh element; apply spatial-temporal dual constraints to the preliminary warning level to generate graded and zoned warning results.

[0010] Secondly, this application also provides a geological disaster early warning system based on digital twins, including:

[0011] The twin modeling module is used to establish an initial digital twin model based on multi-source monitoring data of the geological disaster monitoring area. The initial digital twin model includes discretized finite element mesh elements, the binding interface between each finite element mesh element and the corresponding monitoring data, and the initial mechanical parameter field, initial stress field and initial displacement field corresponding to each finite element mesh element.

[0012] The evolutionary deduction module is used to extract the state vector, including the field parameters of each finite element mesh element, from the current digital twin model at the current moment. Starting with the state vector, an evolutionary path set is obtained through a physical information deep reinforcement learning model. The evolutionary path set includes an optimal evolutionary path and multiple branch evolutionary paths. Each evolutionary path in the set contains a discrete sequence of state vectors from the current moment to the prediction time window and the evolutionary actions corresponding to each time step. Among them, when the current moment is the first warning, the current digital twin model is the initial digital twin model.

[0013] The dynamic update module is used to take the evolution paths in the evolution path set as the current set members of the ensemble Kalman filter; based on the monitoring data received at the current time, the ensemble Kalman filter algorithm with path trajectory similarity weighting is used to dynamically update the predicted state of each set member at the current time, so as to obtain the posterior set of set members and the calibrated digital twin model. The posterior set of set members contains the posterior state vector of each set member at the current time and its corresponding posterior weight.

[0014] The early warning generation module is used to extract the damage intrinsic state variables of each finite element mesh element from the calibrated digital twin model; calculate the instability probability of each finite element mesh element based on the posterior set membership; classify the instability probability to obtain the preliminary early warning level of each finite element mesh element; and apply spatial-temporal dual constraints to the preliminary early warning level to generate graded and zoned early warning results.

[0015] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0016] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0017] The aforementioned geological disaster early warning methods, systems, equipment, and media based on digital twins firstly achieve accurate digital reproduction of the mechanical state and geometric morphology of geological bodies through digital twin modeling. Secondly, by generating optimal and branching evolution paths through deep reinforcement learning of physical information, it solves the problem of existing early warning methods having a single evolutionary deduction and being unable to depict the multimodal evolutionary trend of disasters, realizing multi-path parallel prediction of the state evolution trajectory from the current moment to the prediction window. Furthermore, through dynamic updating of ensemble Kalman filtering with path trajectory similarity weighting, it solves the problem of traditional equal-weight assimilation ignoring the long-term matching degree differences of members and easily leading to state drift, realizing adaptive fusion of observation data and evolution paths and accurate calibration of digital twin models. Finally, through damage eigenstate variable extraction and spatial-temporal dual-constraint early warning, it effectively solves the problem of existing hierarchical methods relying on single-moment thresholds and lacking spatiotemporal consistency verification, which easily leads to false alarms, realizing accurate quantification of instability probability and improving the reliability of early warning results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a geological disaster early warning method based on digital twins is provided as an exemplary embodiment of the present invention;

[0020] Figure 2 A schematic diagram of a geological disaster early warning system based on digital twins is provided as an exemplary embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] In one embodiment, such as Figure 1 As shown, a geological disaster early warning method based on digital twins is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0023] S1: Based on multi-source monitoring data of the geological disaster monitoring area, an initial digital twin model is established. The initial digital twin model includes discretized finite element mesh elements, the binding interface between each finite element mesh element and the corresponding monitoring data, and the initial mechanical parameter field, initial stress field and initial displacement field corresponding to each finite element mesh element.

[0024] Specifically, when modeling based on multi-source monitoring data such as displacement, stress, and hydrological environment collected within the geological disaster monitoring area, a finite element method (FEM) discretization can be used to divide the area into uniformly arranged finite element mesh units. This allows for the spatial refinement and representation of a large-scale geological monitoring area, accurately corresponding to the actual morphology and structural characteristics of geological bodies in different locations. By setting up binding interfaces between each finite element mesh unit and the monitoring data, a stable data interaction channel can be established between the digital model and the field monitoring equipment, ensuring that field-measured data can be accessed into the model in real time. Furthermore, the initial mechanical parameter field, initial stress field, and initial displacement field corresponding to each mesh unit can be configured within the model to reproduce the initial mechanical equilibrium state of the geological body in the monitoring area before disaster deformation. This lays the data and model foundation for subsequent disaster state evolution prediction from a physical perspective, realizing the mapping of natural geological entities to a digital virtual model.

[0025] S2: Extract the state vector of the current moment, including the field parameters of each finite element mesh element, from the current digital twin model; using the state vector as the starting state, obtain the evolution path set through the physical information deep reinforcement learning model; the evolution path set includes an optimal evolution path and multiple branch evolution paths, and each evolution path in the evolution path set contains a state vector sequence discrete by time step from the current moment to the prediction time window and the evolution action corresponding to each time step; where, when the current moment is the first warning, the current digital twin model is the initial digital twin model.

[0026] Specifically, various field parameters corresponding to each finite element mesh unit can be extracted and integrated from the currently running digital twin model to form a standardized state vector. This vector summarizes the core feature information such as geological deformation and mechanical changes within the model, providing input for subsequent intelligent inference calculations. The current digital twin model serves as the initial digital twin model during the first warning operation. This state vector is used as the starting condition for inference and integrated into a physical information deep reinforcement learning model. This model incorporates inherent physical laws such as geotechnical mechanics and geological deformation during the inference calculation process, avoiding the problem of evolutionary logic deviating from the actual geological scenario that easily occurs in purely data-driven inference methods. This generates a set of evolutionary paths, including the optimal evolutionary path and multiple branch evolutionary paths. Each evolutionary path simultaneously carries a temporal state vector sequence extending from the current moment to the prediction time window, along with the corresponding evolutionary actions at each time step, reflecting the state change patterns and evolutionary execution logic under different development trends of the geological body.

[0027] S3: The evolution paths within the evolution path set are used as the current set members of the ensemble Kalman filter; based on the monitoring data received at the current moment, the predicted state of each set member at the current moment is dynamically updated using the ensemble Kalman filter algorithm with path trajectory similarity weighting, to obtain the posterior set of set members and the calibrated digital twin model. The posterior set of set members contains the posterior state vector of each set member at the current moment and its corresponding posterior weight.

[0028] Specifically, various evolutionary paths that have already been generated can be directly used as set members in the ensemble Kalman filter operation process. This ensures that the member samples used in the filtering operation all closely match the actual evolution trend of geological hazards, distinguishing it from the coarse approach of traditional random construction of set members. Subsequently, based on the newly acquired monitoring data at the current moment, a path trajectory similarity weighted approach is used to perform ensemble Kalman filter state update operations. This allows for differentiated dynamic correction of the predicted state of each set member based on the degree of fit between the overall development trend of different evolutionary paths and the field measured data, solving the problem of one-sided state correction in the traditional equal weight update mode. After filtering and iterative updates, a posterior set of set members carrying the corresponding posterior state vector and matching posterior weights can be obtained. Simultaneously, based on the updated set member state data, the overall field parameters of the currently used digital twin model can be synchronously adjusted and calibrated, continuously reducing the state difference between the digital virtual model and the actual geological entity on site. This enables the digital twin model to dynamically iterate in real time following changes in the field geological conditions, always maintaining a high degree of realism.

[0029] S4: Extract the damage intrinsic state variables of each finite element mesh element from the calibrated digital twin model; calculate the instability probability of each finite element mesh element based on the posterior set membership; classify the instability probability to obtain the preliminary warning level of each finite element mesh element; apply spatial-temporal dual constraints to the preliminary warning level to generate graded and zoned warning results.

[0030] Specifically, damage eigenstate variables corresponding to each finite element grid unit can be extracted from the real-time calibrated digital twin model. These variables directly reflect the degree of damage to the internal structure and the deterioration of mechanical properties of geological bodies in different locations, serving as core characteristic indicators for assessing geological hazard risk. Subsequently, the instability probability corresponding to each grid unit is comprehensively calculated by combining multiple sets of extrapolated states and weight information within the posterior set membership. By integrating multiple hazard evolution trend results, the comprehensiveness and objectivity of the instability risk assessment results can be improved. The calculated instability probabilities are then classified according to preset risk assessment standards, enabling rapid preliminary definition of the hazard risk level for all grid units. Furthermore, spatial distribution constraints and temporal variation constraints can be introduced for dual correction processing. This eliminates abnormal risk assessment results caused by external factors such as instantaneous fluctuations in on-site monitoring data and environmental noise interference, avoiding misjudgments caused by risk assessments at a single moment. The final output is a graded and zoned early warning result that combines spatial distribution patterns and temporal stability characteristics. This result can clearly delineate the hazard risk level and impact range of different regions, providing accurate and effective decision-making basis for early prevention and control of geological hazards and zoned management.

[0031] The above method first constructs a standardized initial digital twin model based on regional multi-source monitoring data, completing the integrated digital modeling of the geological body's spatial and mechanical parameters. Second, it extracts model field parameters to construct state vectors and generates an evolution path set based on deep reinforcement learning of physical information. This set can distinguish different early warning scenarios and adapt to corresponding model bases, achieving rational deduction of disaster evolution trends. Furthermore, it constructs a filter set based on the evolution paths and uses path trajectory similarity weighting for state assimilation updates, calibrating the digital twin model's operating state in real time, effectively improving the model's accuracy in reproducing the actual state of the land body. Finally, it extracts damage variables and solves for instability probabilities based on the calibrated model, optimizing early warning levels by combining graded judgment and spatiotemporal constraints, outputting graded and zoned early warning results. This significantly reduces false early warnings caused by abnormal disturbances and significantly enhances the accuracy and practicality of dynamic geological disaster early warning.

[0032] In one embodiment, the geological disaster early warning method S2 based on digital twins provided in this application, starting with a state vector, obtains an evolution path set through a deep reinforcement learning model of physical information, which may include:

[0033] Multiple sets of state replicas are constructed based on the state vector. Gaussian noise is superimposed on each state replica to generate each initial set member. Each initial set member is assigned an equal initial weight.

[0034] Determine whether the current moment is the first warning. If the current moment is the first warning, set an initial exploration temperature for each initial set member. Otherwise, determine the exploration temperature of the corresponding initial set member based on the posterior weight of each set member obtained from the previous warning moment. The posterior weight of each set member is negatively correlated with the exploration temperature of the initial set member.

[0035] Based on the physical information deep reinforcement learning model, starting from the state corresponding to each initial set member, the optimal evolution path is generated by greedy deduction, and combined with the exploration temperature, multiple branch evolution paths are generated by Boltzmann exploration strategy.

[0036] The optimal evolution path is combined with multiple branch evolution paths to form an evolution path set. The state vector sequence of each evolution path in the evolution path set is recorded sequentially from the current time to the prediction time window. The evolution actions executed at each time step are saved. The evolution actions include seepage driving actions, mechanical loading actions, strength degradation actions, and deformation calculation actions.

[0037] Specifically, based on the finite element mesh discrete architecture built into the digital twin model corresponding to the current moment, a full-domain traversal collection of various physical field parameters corresponding to all finite element mesh units within the monitoring area can be performed. The extracted unit field parameters can include real-time mechanical parameters of the soil and rock mass, spatial stress parameters, displacement and deformation parameters, and pore water pressure and seepage parameters. These parameters can comprehensively reflect the current material properties, stress state, deformation characteristics, and hydrological environment characteristics of the geological body. Subsequently, the discretely distributed single-unit physical parameters can be sequentially concatenated according to a preset fixed-dimensional sorting rule to obtain a state vector. This state vector uniquely corresponds to the true comprehensive state of the geological body in the monitoring area at the current moment.

[0038] Furthermore, multiple sets of state replicas can be constructed in batches based on the state vector. Each set of state replicas inherits the global field characteristic information, such as the mechanical parameters, stress parameters, and deformation parameters of the finite element mesh elements contained in the state vector, to ensure that the initial base state of all replicas is completely aligned with the actual initial geological state of the monitoring area. On this basis, random noise conforming to a Gaussian distribution can be superimposed on each set of independent state replicas. This controlled random perturbation simulates the fluctuations of minute parameters within the geological body, slight errors in the monitoring system, and latent environmental interferences, generating multiple sets of independent initial set members with subtle state differences, thus enriching the diversity of the initial simulation samples. All initial set members are uniformly assigned equal initial weights, ensuring that each potential evolutionary sample has an equal probability of participating in the calculation during the initial stage of the early warning simulation.

[0039] Specifically, the exploration temperature parameters of each initial set member can be configured differently based on different scenarios at different warning time nodes. Exploration temperature is a core parameter for regulating the randomness and path diversity of the reinforcement learning model's exploration, directly determining the dispersion of branch evolution paths and the coverage of uncertain scenarios. If the current moment is the first warning, since no historical extrapolation data, observation matching data, or state update weight data have been accumulated, there is no effective historical prior information to support differentiated strategy regulation. Therefore, a fixed initial exploration temperature can be uniformly configured for all initial set members for the first round of multi-path extrapolation. If the current moment is not the first warning, the posterior weights of each set member statistically output after the iteration calculation of the previous complete warning cycle can be retrieved. Based on the negative correlation between the posterior weights and the exploration temperature, the exploration temperature of each initial set member can be adaptively adjusted one by one. A higher posterior weight value indicates a higher degree of alignment between the historical inference path of the corresponding set member and the monitoring data and actual geological evolution law, resulting in stronger path inference credibility. A lower exploration temperature is configured accordingly to constrain the random exploration range of the highly credible path and preserve its core evolution trend. Conversely, a lower posterior weight value indicates a larger historical deviation and lower reference value for the corresponding inference path. A higher exploration temperature can be configured accordingly to enhance the randomness of the path's exploration, enabling it to break away from the original deviation evolution logic and achieve dynamic iterative optimization of the inference strategy.

[0040] After configuring the initial set of members' states and exploration temperatures, a time-series evolution can be deduced based on a trained deep reinforcement learning model of physical information. This model can embed geotechnical equilibrium equations, seepage continuity equations, and Mohr-Coulomb strength constraints to ensure that the generated state iteration results strictly adhere to the physical evolution laws of the geological body. For example, the model can employ a dual-mode path generation mechanism. For generating the optimal evolution path, a greedy deduction approach is used throughout, selecting the evolutionary action with the highest value at each discrete time step for state iteration, continuously retaining the evolutionary direction that best fits the physical constraints and measured laws at each time step, iteratively generating a unique globally optimal evolution path. For generating branch evolution paths, the Boltzmann exploration strategy can be used to randomly sample evolutionary actions, combining the exploration temperature specific to each set member. The dispersion of the action sampling probability distribution is precisely controlled by the exploration temperature, generating multiple branch evolution paths with differentiated evolutionary trends to cover various uncertain scenarios in geological hazard evolution. Illustratively, the action sampling probability calculation formula can be:

[0041]

[0042] in, This represents the sampling probability of the model selecting a specified evolutionary action under the current geological conditions. Let V be the real-time model global state vector corresponding to any initial set member. The cumulative action value corresponding to a single evolutionary action in the current state. The evolutionary action of the target to be screened. This refers to all available evolutionary actions within the model's action space, excluding the target action. The exploration temperature configured for the corresponding initial set members.

[0043] Specifically, the iterative update actions during model derivation can be uniformly divided into four standardized evolution actions. Among them, the seepage-driven action can be used to dynamically update the global pore water pressure field and soil permeability parameter field of the model by combining real-time rainfall, groundwater level and other hydrological boundary conditions; the mechanical loading action can be used to iteratively update the global stress field distribution of each finite element mesh element based on the self-weight stress of the geological body and the external load boundary; the strength degradation action can be used to dynamically correct the shear strength parameters of the soil and rock mesh elements according to the strain softening characteristics and cumulative plastic deformation of the soil and rock mass, and simulate the long-term degradation and damage process of the geological body; the deformation calculation action can be used to solve the finite element mechanical equilibrium equations, iteratively update the global displacement deformation field, and accurately reproduce the deformation evolution process of the geological body.

[0044] By integrating the single optimal evolution path obtained from the deduction with the evolution paths of each branch, a complete set of evolution paths can be formed. At the same time, based on a preset unified discrete time step, the temporal state vector sequence of each evolution path in the set can be recorded from the current warning time to the end of the prediction time window, and the specific evolution actions executed in each discrete time step iteration can be retained. This enables the full-process traceability and reusability of all disaster evolution trajectories, state change processes and evolution execution logic, providing prior path sample data support for the subsequent state assimilation and update operation of ensemble Kalman filtering.

[0045] In one embodiment, when the current moment is the first warning, the geological disaster early warning method S3 based on digital twins provided in this application dynamically updates the predicted state of each set member at the current moment using a path trajectory similarity-weighted ensemble Kalman filter algorithm based on the monitoring data received at the current moment, to obtain the posterior set of set members and the calibrated digital twin model, which may include:

[0046] S21: Extract the state at the current time from the evolution path corresponding to each set member, and use the state at the current time as the predicted state of each set member at the current time; add a preset model error term to the predicted state to correct the state and obtain the corrected predicted state.

[0047] Specifically, in the geological state assimilation operation, traditional ensemble Kalman filtering typically uses random perturbation to generate batches of ensemble members. However, randomly generated members only possess numerical differences and lack corresponding geological evolution physical logic. Many members exhibit evolutionary patterns that deviate from the constraints of geotechnical mechanics and the actual evolutionary trend of disasters, easily leading to invalid filter update samples, insufficient state calibration accuracy, and distorted uncertainty coverage. Therefore, during the initial warning, the independent evolutionary paths within the derived evolutionary path set can be directly used as the ensemble members for the current operation of the ensemble Kalman filtering algorithm. Each ensemble member is an effective qualitative evolution trajectory derived following seepage laws, mechanical equilibrium conditions, and geotechnical strength deterioration mechanisms. Each ensemble member corresponds to a complete and reliable temporal evolution state of the geological body, ensuring that the ensemble member samples have sufficient diversity and uncertainty coverage, and ensuring that all samples conform to the actual evolution mechanism of geological disasters.

[0048] During state update calculations, the temporal state corresponding to the current warning time in each evolution path can be extracted according to the temporal alignment rules. This temporal state is determined as the original predicted state of the corresponding set member. This predicted state carries the global field information such as mechanical parameters, stress distribution, and deformation characteristics derived from the corresponding evolution path at the current time. It is the prior inference result before the model is corrected by incorporating measured data. To overcome the state distortion problems caused by inherent model biases, parameter truncation errors, and small unknown disturbances of geological bodies in numerical inference, a model error term conforming to a normal distribution can be superimposed on the original predicted state of each set member. This simulates the objective model uncertainties that exist in the evolution of geological hazards, and performs inclusive correction on the original predicted state, ultimately obtaining a corrected predicted state that balances the realism of the inference with the coverage of uncertainties.

[0049] S22: Obtain the monitoring data received at the current moment, map the corrected prediction state corresponding to each set member to the observation space through the preset observation operator, and solve the state residual between the predicted observation value corresponding to the corrected prediction state of each set member and the measured value corresponding to the monitoring data.

[0050] Specifically, the monitoring data currently accessed is real geological monitoring data collected in real time by on-site sensors, constituting the real constraints for model state correction. Since the global state space dimension of the digital twin model is much higher than the observation space dimension of the on-site monitoring, the grid cell field parameters within the model cannot be directly compared and calculated with the measured data. Therefore, a pre-constructed observation operator can be used to complete the spatial dimension matching and transformation. This observation operator has a fixed spatial mapping logic, which can accurately map the high-dimensional corrected predicted state of each ensemble member from the model state space to the low-dimensional observation space based on the spatial coordinates of the monitoring points, the grid cell binding relationship, and the interpolation rules, transforming it into predicted observation values ​​that perfectly match the dimension and physical quantity type of the measured data. Based on this, the difference between the predicted observation value and the corresponding measured value of each ensemble member can be calculated one by one to obtain the state residual, which can quantify the deviation of a single evolution path. The magnitude of the state residual directly reflects the degree of fit between the evolution path of the corresponding ensemble member and the actual geological evolution conditions.

[0051] S23: Based on the state residuals, obtain the initial residual weighted sum of squares for each set member, and calculate the path trajectory similarity weight for each set member based on the initial residual weighted sum of squares.

[0052] Specifically, for the initial warning scenario, without prior knowledge of historical iteration weight data, the reliability of each set member is entirely determined by the residual characteristics at the current moment. For example, the initial residual weighted sum of squares can be obtained by performing a weighted square operation on the state residuals of each set member. This indicator can comprehensively quantify the cumulative degree of deviation of the overall inference of a single evolution path. The larger the value of the initial residual weighted sum of squares, the greater the deviation between the inference result of the corresponding evolution path and the actual field conditions, and the lower the reliability of the path evolution. Based on the initial residual weighted sum of squares, the path trajectory similarity weight of each set member can be obtained through exponential normalization, realizing the differentiated quantitative distinction of the reliability of different evolution paths. This differs from the coarse calculation method of equal weights in traditional set Kalman filtering. The calculation formula is as follows:

[0053]

[0054] in, For the first The path trajectory similarity weights corresponding to each set member For the first The initial residual weighted sum of squares corresponding to each set member This is the normalized temperature adjustment coefficient, used to control the dispersion of the weight distribution. This represents the total number of all set members currently participating in the calculation. This formula achieves a negative correlation between deviation and weight through an exponential decay mechanism, allowing evolution paths with smaller residuals and closer to real-world conditions to receive higher weights, while deviation paths with larger residuals receive lower weights.

[0055] S24: Based on the path trajectory similarity weights corresponding to each set member, calculate the weighted set mean and weighted set covariance respectively, and combine the weighted set mean and weighted set covariance to obtain the Kalman gain matrix.

[0056] Specifically, after obtaining the differentiated path trajectory similarity weights, the weights can be used as confidence coefficients to perform a weighted summation operation on the corrected predicted states of all ensemble members. This yields a weighted ensemble mean that represents the highest priori state of the geological body at the current moment. This mean preferentially aligns with the characteristics of high-confidence evolution paths, effectively weakening the interference of biased paths on the overall model state. Based on the weighted ensemble mean, the perturbation deviation between the state of each ensemble member and the mean can be further solved, and weighted correction can be performed using the path trajectory similarity weights to obtain a weighted ensemble covariance that accurately reflects the uncertainty distribution characteristics of the model state. This weighted ensemble covariance accurately characterizes the error distribution law of the model's predicted state. Combined with the observation noise covariance, a simultaneous operation is performed to obtain the Kalman gain matrix. The Kalman gain matrix is ​​the core control matrix for state updates, adaptively balancing the confidence ratio between the model's projected state and the field measured data, determining the magnitude and direction of subsequent state corrections.

[0057] S25: Based on the Kalman gain matrix, match the state residuals corresponding to each set member, iteratively correct the predicted state of each set member at the current time, and generate the posterior state vector corresponding to each set member at the current time.

[0058] Specifically, by performing matrix matching operations on the obtained Kalman gain matrix and the state residuals corresponding to each ensemble member, an independent state correction amount can be obtained for each ensemble member. This correction amount is adaptively generated based on path reliability and measured deviation, enabling differentiated correction for ensemble members with different degrees of deviation. By iteratively updating the original predicted state of each ensemble member dimension by dimension using the state correction amount, the deviation components of the model's mechanical parameters, stress parameters, and deformation parameters are corrected one by one. This allows the inferred state of each ensemble member to converge towards the actual geological conditions, ultimately obtaining the posterior state vector of all ensemble members after data assimilation and correction. Each set of posterior state vectors corresponds to an optimized geological hazard evolution trajectory.

[0059] S26: Take the path trajectory similarity weights corresponding to each set member as the posterior weights of each set member at the current time, and combine the posterior state vectors and posterior weights of each set member at the current time to obtain the posterior set member set.

[0060] Specifically, the path trajectory similarity weights obtained in the initial warning scenario accurately represent the true credibility of each evolution path within the current warning period. These weights can be directly used as the posterior weights for each set member, achieving a one-to-one binding and matching between the weights and the optimized states. Integrating and encapsulating the posterior state vectors and corresponding posterior weights of each set member constructs a complete posterior set of members. This set contains multiple sets of differentiated optimal geological evolution states and quantifies the confidence probability of each state, providing uncertainty sample support for subsequent instability probability statistics and risk classification.

[0061] S27: Using the posterior state vectors of each set member within the posterior set as the calibration basis, perform global state synchronization calibration on the current digital twin model, update the mechanical parameter field, stress field, and deformation field within the current digital twin model, and obtain the calibrated digital twin model.

[0062] Specifically, after obtaining the posterior set of members, based on the confidence weights corresponding to each posterior state vector, a global weighted fusion calibration can be performed on the mechanical parameter field, stress field, and deformation field of the global finite element mesh elements. This eliminates the need for local single-point correction and achieves synchronous iterative updates of the model's global physical field. During this calibration process, optimization features from multiple high-confidence evolution paths can be incorporated, while interference from low-confidence deviation paths is avoided. This ensures that the global physical field parameters, stress distribution, and geological deformation state of the digital twin model closely match the current field measurement conditions, eliminating state drift issues caused by model extrapolation. Ultimately, dynamic model calibration is completed, resulting in a highly accurate calibrated digital twin model that accurately recreates the real-time state of the geological body.

[0063] In one embodiment, when the current moment is not the first warning, the posterior set membership can be obtained through the following steps:

[0064] S31: Based on the action sequence from the previous warning time to the current time in the evolution path corresponding to each set member, and taking the posterior state corresponding to each set member at the previous warning time as the starting point for deduction, state prediction is performed to obtain the predicted state corresponding to each set member at the current time; the predicted state is corrected by superimposing a preset model error term on the predicted state to obtain the corrected predicted state.

[0065] Specifically, the action sequence in S31 is a temporal control sequence generated by a physical information deep reinforcement learning model, containing geotechnical mechanics-driven logic. This sequence guides state deduction to follow the physical laws of seepage drive, mechanical loading, strength degradation, and deformation calculation, avoiding the distortion of physical logic caused by unconstrained pure data-driven deduction. In non-initial warning scenarios, the posterior state at the previous warning time has been integrated with historical monitoring data, making it more closely aligned with the actual mechanical state of the geological body compared to the initial state. Using this as a starting point for deduction ensures the continuity and rationality of the predicted state. To simulate the uncertainties brought about by inherent model biases, parameter truncation errors, and unknown micro-disturbances during the evolution of geological hazards, a preset model error term can be superimposed on the predicted state for state correction. This error term introduces process noise, simulating parameter uncertainties and environmental disturbances during geological evolution, preventing numerical divergence in the filtering process due to deterministic model assumptions, and improving the robustness of state prediction.

[0066] S32: Obtain the monitoring data received at the current moment, map the corrected prediction state corresponding to each set member to the observation space through the preset observation operator, and solve the state residual between the predicted observation value corresponding to the corrected prediction state of each set member and the measured value corresponding to the monitoring data.

[0067] Specifically, the role of the preset observation operator is to convert the state vectors in the digital twin model, such as mechanical parameter fields, stress fields, and deformation fields, into observation vectors consistent with the dimensions of the monitoring data, such as directly observable physical quantities like pore water pressure and surface displacement. This eliminates the dimensional and spatial differences between the model state and the measured data, providing a unified benchmark for residual calculation. Based on this, the difference between the predicted observation vector and the corresponding measured vector for each set member can be calculated to obtain the state residual. This residual provides a quantitative basis for subsequent weight allocation and state correction, reflecting the degree of deviation between the predicted state and the measured data.

[0068] S33: Retrieve the historical residual weighted sum of squares accumulated from the first warning to the previous warning time for each set member, add the state residual to the historical residual weighted sum of squares to obtain the latest residual weighted sum of squares for each set member, and calculate the path trajectory similarity weight for each set member based on the latest residual weighted sum of squares.

[0069] Specifically, in non-initial warning scenarios, the cumulative calculation results of all state residuals for each set member from the time of the first warning to the time of the previous warning have been retained, i.e., the weighted sum of squared historical residuals. This indicator quantifies the overall cumulative deviation of the corresponding evolution path since the warning was initiated. The state residuals obtained at the current moment are then added to the weighted sum of squared historical residuals according to a preset Mahalanobis distance squared calculation method to obtain the latest weighted sum of squared residuals for each set member from the first warning to the current moment. This indicator characterizes the overall degree of fit between the corresponding evolution path and the measured operating conditions throughout the entire warning period. Based on the latest weighted sum of squared residuals, the path trajectory similarity weights for each set member can be obtained through exponential normalization operations. The calculation formula is as follows:

[0070]

[0071] in, Indicates the first At the current moment, the members of the set are... Path trajectory similarity weights, For the current moment, The first warning during the early warning process, starting from the first warning. A historical warning moment, For the first Each set member in At each step, the corrected predicted state of each set member is mapped to the predicted observation vector in the observation space. for The measured observation vector corresponding to the monitoring data received at the current time. This represents the square of the Mahalanobis distance, used to quantify the deviation between predicted and actual observations. This distance is weighted by the observation noise covariance matrix, which can eliminate the influence of differences in noise levels across various dimensions of the monitoring data. To observe the noise covariance matrix, it is matched with a preset observation operator to characterize the differences in noise levels across various dimensions of the monitoring data; Temperature is a parameter used to control the degree of influence of path trajectory deviation on weight allocation, and can be calibrated according to the type of geological disaster and the monitoring scenario; This represents the total number of members in the set.

[0072] In the above formula, the numerator uses an exponential function to nonlinearly map the accumulated deviation term. The smaller the deviation of a set member, the larger its exponential term value, and the higher its corresponding weight. The denominator is the sum of the exponential terms of all set members, serving a normalization function. This ensures that the sum of the path trajectory similarity weights of all set members is 1, achieving standardized weight allocation. This weight calculation method based on accumulated deviation differs from the equal weight allocation mechanism of traditional ensemble Kalman filtering. It reflects the differences in the long-term matching degree between different evolutionary paths and measured data, with paths with smaller deviations receiving higher weights, thereby improving the accuracy of subsequent state updates.

[0073] S34: Based on the path trajectory similarity weights corresponding to each set member, calculate the weighted set mean and weighted set covariance respectively, and combine the weighted set mean and weighted set covariance to obtain the Kalman gain matrix.

[0074] Similarly, after calculating the path trajectory similarity weights for each set member, the corrected predicted states of each set member can be weighted and summed, similar to the calculation logic in the initial warning scenario. This amplifies the contribution of set members with high credibility to the statistical characteristics and reduces the influence of set members with larger deviations, resulting in a weighted set mean that represents the highest priori state of the geological body at the current moment. Based on the weighted set mean, the perturbation deviation between the corrected predicted state of each set member and the mean is further solved, and weighted correction is performed using the path trajectory similarity weights. This yields a weighted set covariance that accurately reflects the distribution characteristics of the model's state uncertainty. By simultaneously solving the weighted set covariance and the observation noise covariance matrix, the Kalman gain matrix can be obtained. This matrix is ​​the core control matrix for state updates, adaptively balancing the credibility ratio between the model's prior inferred state and the field measured data.

[0075] S35: Based on the Kalman gain matrix, match the state residuals corresponding to each set member, iteratively correct the predicted state of each set member at the current time, and generate the posterior state vector corresponding to each set member at the current time.

[0076] Specifically, similar to the calculation logic in the initial early warning scenario, the obtained Kalman gain matrix is ​​used to perform matrix matching operations with the state residuals corresponding to each set member. The state residuals are then introduced into the predicted state through the Kalman gain matrix, achieving the fusion of the predicted state and the measured data, resulting in independent state corrections for each set member. The predicted states of each set member are iteratively updated dimension by dimension using these state corrections, correcting the deviation components of the model's mechanical parameters, stress parameters, and deformation parameters one by one. This allows the inferred states of each set member to converge towards the actual geological conditions, ultimately yielding posterior state vectors for all set members after data assimilation and correction. Each set of posterior state vectors corresponds to an optimized geological hazard evolution trajectory. The corrected posterior state vectors retain the physical constraints of geological evolution while incorporating the measured data at the current moment, enabling a more accurate reflection of the true mechanical state of the geological body.

[0077] S36: Determine the path trajectory similarity weights corresponding to each set member as the posterior weights of each set member at the current time; combine the posterior state vectors and posterior weights corresponding to each set member to obtain the posterior set member set.

[0078] Specifically, the path trajectory similarity weights calculated at the current moment characterize the overall fit between each evolution path and the measured conditions throughout the entire warning period. These weights can be directly used as the posterior weights of each set member, achieving a one-to-one binding and matching between weights and optimized states, maintaining consistency in the credibility of set members, without requiring additional weight allocation calculations. Integrating the posterior state vectors of each set member with their corresponding posterior weights constructs a complete posterior set of members. This set of members provides fundamental data support for subsequent digital twin model calibration and the calculation of geological disaster instability probabilities.

[0079] In one embodiment, the geological disaster early warning method S4 based on digital twins provided in this application extracts the damage eigenstate variables of each finite element mesh element from the calibrated digital twin model; calculates the instability probability of each finite element mesh element based on the posterior set membership; and classifies the instability probabilities to obtain the preliminary early warning level of each finite element mesh element, which may include:

[0080] The geotechnical mechanics parameters and mechanical response parameters corresponding to each finite element mesh element are retrieved from the calibrated digital twin model. The geotechnical mechanics parameters include cohesion and internal friction angle, and the mechanical response parameters include equivalent plastic strain, mean stress and pore water pressure.

[0081] The geotechnical mechanics parameters and mechanical response parameters corresponding to each finite element mesh element are integrated to form the feature vector corresponding to each finite element mesh element. The feature vector is input into the sparse autoencoder to perform feature analysis calculation and output the damage eigenstate variable corresponding to each finite element mesh element. The value of the damage eigenstate variable is between 0 and 1.

[0082] Iterate through each set member in the posterior set member set, and pre-set an initial cumulative probability value for a single finite element mesh element. Then, determine whether each set member meets the preset conditions on a single finite element mesh element. The preset conditions include that the Mohr-Coulomb yield function is greater than the preset judgment value and the equivalent plastic strain rate exceeds the preset strain rate threshold.

[0083] If a single set member satisfies the judgment condition, the posterior weight corresponding to the set member is accumulated into the initial accumulated probability value of the corresponding finite element mesh element. After traversal, the final instability probability corresponding to each finite element mesh element is obtained.

[0084] The safety factor of each finite element mesh element in the calibrated digital twin model is retrieved, and low-risk, medium-risk, and high-risk thresholds are set for each finite element mesh element based on the safety factor.

[0085] The instability probability corresponding to each finite element mesh element is numerically compared with the low-risk threshold, medium-risk threshold, and high-risk threshold, and the preliminary warning level corresponding to each finite element mesh element is determined based on the comparison results.

[0086] Specifically, the real-time geotechnical mechanics parameters and mechanical response parameters can be retrieved one by one from the calibrated digital twin model for all finite element mesh elements within the monitoring area. The geotechnical mechanics parameters can include cohesion and internal friction angle, which are core control parameters for the shear strength of the soil and rock mass, directly reflecting the real-time deterioration of the inherent resistance to failure. The mechanical response parameters can include equivalent plastic strain, mean stress, and pore water pressure. Equivalent plastic strain characterizes the degree of cumulative irreversible deformation damage to the soil and rock mass; mean stress characterizes the current overall stress level of the soil and rock mass; and pore water pressure characterizes the weakening effect of groundwater on the effective strength of the soil and rock mass. These three types of response parameters reflect the real-time mechanical response characteristics of the geological body under external loads and hydrological environment. The parameters of the calibrated digital twin model have undergone multiple rounds of data assimilation and iterative correction, enabling accurate matching of the real-time evolution state of the geological body compared to static initial parameters, avoiding discrimination errors caused by the lag in static parameter updates.

[0087] After obtaining the parameters of a single finite element mesh, various parameters can be standardized and their dimensions unified to eliminate feature interference caused by differences in the dimensions of different physical quantities. These parameters are then integrated according to a fixed parameter sorting rule to construct a multidimensional feature vector for the corresponding mesh element. Its expression can be:

[0088]

[0089] in, For the first The eigenvectors corresponding to each finite element mesh element For the real-time cohesion of the grid cells, This represents the real-time internal friction angle of the mesh element. The equivalent plastic strain of the mesh element, The average stress of the mesh element. This represents the pore water pressure of the grid cell.

[0090] By inputting the feature vectors of each grid cell into a pre-trained sparse autoencoder model, the multi-dimensional coupled mechanical features can be reduced and purified through feature compression, sparse constraints, and feature parsing operations within the model. Redundant correlation features and numerical noise between parameters are filtered out, and the intrinsic damage patterns of soil and rock under multi-physics coupling are uncovered. Finally, a single scalar damage eigenstate variable is output at the bottleneck layer of the model. The value of this damage eigenstate variable ranges from 0 to 1. The closer the value is to 0, the less strength degradation has occurred in the corresponding grid cell soil and rock, and the better the structural integrity is. Conversely, the closer the value is to 1, the higher the degree of damage in the corresponding grid cell soil and rock, the more likely the shear strength has failed, and the structure is on the verge of instability and failure. This achieves a quantitative characterization of the latent damage state of soil and rock.

[0091] After extracting the intrinsic states of damage in the mesh elements, the instability probability of the mesh elements can be quantitatively solved based on the posterior set of members. This leverages the uncertainty characteristics of multiple evolution paths to improve the reliability of instability determination. For example, an initial cumulative probability value can be preset for each finite element mesh element, initially set to zero. Then, each member in the posterior set with an independent evolution trajectory and confidence weight is sequentially traversed, and the state of the element corresponding to each member is checked for each individual mesh element to ensure it meets the preset instability conditions. These preset conditions are the core criteria for determining dynamic instability and failure of the geological body, avoiding the pitfalls of traditional single-stress criteria. Specifically, they can include two constraints: the Mohr-Coulomb yield function exceeding a preset threshold and the equivalent plastic strain rate exceeding a preset strain rate threshold. The former determines whether the current stress state of the soil or rock mass exceeds its shear strength limit, and the latter determines whether the soil or rock mass undergoes continuous irreversible plastic deformation. When both conditions are met simultaneously, the corresponding mesh element under that evolution path is determined to be in a critical instability state.

[0092] Specifically, after a single set member determination is completed, the accumulated probability value of the mesh element can be updated based on the determination result. If the element state corresponding to the current set member simultaneously meets two preset conditions, the posterior weight corresponding to that set member is added to the initial accumulated probability value of the mesh element. If the preset conditions are not met, no accumulation is performed. After completing the traversal determination and weight accumulation of all set members, the final accumulated value is the instability probability of the corresponding finite element mesh element, and its calculation expression is as follows:

[0093]

[0094] in, For the first The instability probability of a finite element mesh element. The total number of set members within the posterior set membership set. For the first The posterior weights corresponding to each set member. For a binary indicator function, when the first... Mesh cells in the state of each set member The value is 1 when the dual instability preset condition is met, and 0 when it is not met. This calculation method differs from the traditional equal statistical probability algorithm. It can differentiate the contribution probability ratio based on the true credibility of each evolution path, which can significantly improve the accuracy of instability probability calculation.

[0095] After obtaining the high-precision instability probability of each grid cell, a differentiated threshold grading mechanism adapted to the spatial heterogeneity of the geological body can be used for early warning judgment. For example, the initial safety factor corresponding to each finite element grid cell in the calibrated digital twin model can be retrieved. The initial safety factor is an inherent stability performance index of the soil and rock mass calculated based on the inherent geological parameters of the cell, which can reflect the differences in the basic stability capacity of grid cells at different spatial locations. According to the magnitude of the safety factor of each grid cell, the corresponding low-risk, medium-risk, and high-risk thresholds are adaptively and differentiatedly configured. The lower the safety factor of the weak geological cell, the stricter the configured early warning threshold, so as to achieve sensitive early warning of high-risk weak areas. The higher the safety factor of the stable geological cell, the more lenient the configured early warning threshold can be, so as to effectively avoid invalid early warning interference in stable areas.

[0096] The real-time instability probability obtained by solving each finite element mesh element is compared with the corresponding low-risk, medium-risk, and high-risk thresholds for each element. Based on the threshold range in which the instability probability falls, the preliminary warning level corresponding to each finite element mesh element is defined. This yields a comprehensive preliminary warning result that includes spatial location, instability probability, and warning level, providing basic hierarchical data for subsequent spatiotemporal dual constraint correction.

[0097] In one embodiment, the geological disaster early warning method S4 based on digital twins provided in this application applies spatial-temporal dual constraints to the preliminary early warning level to generate graded and zoned early warning results, which may include:

[0098] Finite element mesh elements with an initial warning level of red are designated as red finite element mesh elements. Density clustering algorithm is used to extract the first connected mesh cluster with a first spatial area greater than the first preset area threshold. The first connected mesh cluster is designated as a candidate red warning area. The initial warning level of isolated red finite element mesh elements that are not included in any first connected mesh cluster is downgraded to orange.

[0099] Finite element mesh elements with an initial warning level of orange are used as orange finite element mesh elements. Density clustering algorithm is used to extract second connected mesh clusters with a second spatial area greater than a second preset area threshold. The initial warning level of isolated orange finite element mesh elements that are not included in any second connected mesh cluster is downgraded to yellow.

[0100] Retrieve historical warning level data from at least two consecutive warning moments prior to the current moment, perform time-series continuity verification on each candidate red warning area, check whether the warning level corresponding to each candidate red warning area remains red within three consecutive sampling periods, and obtain the verification result;

[0101] If the verification result shows that the candidate red warning area meets the time-series continuous verification requirements, then the final warning level of the corresponding finite element mesh element in the candidate red warning area is locked as red, and the final red warning area is obtained. Otherwise, the final warning level of the corresponding finite element mesh element in the candidate red warning area is downgraded to orange warning level, and the continuous detection status is marked.

[0102] The finite element mesh elements with an initial warning level of yellow are used as yellow finite element mesh elements. Combined with the final red warning area, orange finite element mesh elements, and yellow finite element mesh elements, the spatial distribution data is visualized. The centroid coordinates, coverage area, and average instability probability of the final red warning area are statistically analyzed, and the corresponding effective time window is determined to generate emergency red information. Based on the emergency red information and spatial distribution data, a graded and zoned warning result is generated. The average instability probability is obtained by calculating the mean of the instability probabilities of the finite element mesh elements corresponding to the final red warning area.

[0103] Specifically, after the initial warning level classification, the inherent physical characteristics of the spatially contiguous development and temporally progressive evolution of geological hazards can be combined to perform dual correction processing on the initial warning results, including spatial connectivity constraints and temporal persistence constraints. This process eliminates false warning signals caused by numerical calculation disturbances, instantaneous monitoring noise, and local numerical anomalies, and removes the problems of isolated misjudgments and short-term jump false warnings, ultimately outputting graded and zoned warning results. For example, spatial clustering correction can be performed on the initially identified red high-risk grid cells, defining all finite element grid cells with an initial warning level of red as red finite element grid cells. Furthermore, high-risk instability and damage of geological hazards all have the characteristics of contiguous development and overall slippage. A single isolated red grid cell usually does not have the geological conditions for the actual development of the hazard and is mostly a false risk signal caused by model calculation truncation errors and instantaneous fluctuations in monitoring data.

[0104] In a schematic manner, density clustering algorithms can be used to perform spatial clustering analysis on the entire red finite element mesh cells. Based on the spatial adjacency relationship and cell distribution density of the mesh cells, spatially continuous and densely distributed mesh cells are automatically aggregated to form connected mesh clusters, while discrete, single-point anomalous mesh cells are filtered out. After clustering, the actual spatial area covered by each connected mesh cluster can be calculated one by one, and connected structures with a spatial area greater than a first preset area threshold can be selected as the first connected mesh cluster. The spatial regions corresponding to all first connected mesh clusters are uniformly designated as candidate red warning areas. For red finite element mesh cells that are not included in any first connected mesh cluster and are isolated and discretely distributed, it can be determined that their risk signals do not have the physical support for geological disaster development, and their preliminary warning level is uniformly downgraded to the orange warning level, completing the spatial noise reduction correction of the red warning area.

[0105] Similarly, based on the above correction logic, spatial constraint correction can be performed on orange medium-risk grid cells. Illustratively, all finite element grid cells with an initial warning level of orange can be defined as orange finite element grid cells. The geological weakening risk corresponding to orange warnings also exhibits regional and contiguous distribution characteristics, but scattered and isolated orange grid cells cannot form effective disaster hazards; they are anomaly judgments caused by low-amplitude environmental disturbances and numerical deviations. Therefore, spatial clustering can be performed on all orange finite element grid cells using a density clustering algorithm to aggregate multiple sets of continuously distributed connected grid structures. The spatial area of ​​each set of connected grid structures is calculated, and connected structures with a spatial area greater than a second preset area threshold are selected as the second connected grid cluster. The orange warning level of the corresponding cells in the second connected grid cluster is retained. For isolated orange finite element grid cells not included in any second connected grid cluster, they can be determined to have no substantial medium-risk hazard, and their initial warning level is uniformly downgraded to the yellow warning level, completing the spatial accuracy optimization of the medium-risk warning area.

[0106] Furthermore, although the candidate red alert areas obtained after spatial denoising correction have eliminated spatially isolated false risks, there are still temporary high-level false alerts caused by sudden factors such as short-term monitoring disturbances and instantaneous micro-seismic shocks. The actual geological disaster instability evolution is a slow, cumulative, and continuously deteriorating temporal process, and there will be no sudden risk abrupt change that disappears immediately. Therefore, based on this geological evolution law, we can retrieve the historical warning level data of the entire region stored at least two consecutive warning sampling times before the current warning time, and combine it with the warning data at the current time to form a temporal warning dataset for three consecutive complete sampling periods. For each candidate red alert area, we can conduct a region-by-region and time-by-time temporal continuity verification to check whether the warning level of each candidate red alert area has been stably maintained at the red level throughout the three consecutive sampling periods. Based on the verification results, we can obtain the corresponding temporal verification results.

[0107] If the time-series verification results show that the candidate red warning area maintains a stable red warning level for three consecutive sampling periods, it indicates that the soil and rock damage and instability trend in this area is long-term and persistent. The risk evolution conforms to the gradual development law of geological disasters, and there is no instantaneous disturbance interference. Therefore, the final warning level of all finite element mesh units within the candidate red warning area can be locked as red, and all qualified areas can be integrated to form the final red warning area. If the candidate red warning area experiences a downgrade in any consecutive sampling period, the high-risk state of this area can be determined as a false warning caused by an instantaneous abnormal disturbance, lacking the conditions for sustained instability development. The final warning level of all finite element mesh units in this area should be uniformly downgraded to orange warning level, and the area should be marked as continuously monitored to track its risk evolution trend in subsequent warning periods, avoiding the problem of missing potential slowly changing disaster hazards.

[0108] Specifically, after completing the spatiotemporal dual-constraint correction, all finite element mesh elements with an initial warning level of yellow can be uniformly categorized and organized. Combined with the corrected final red warning area, the denoised orange finite element mesh elements, and the global yellow finite element mesh elements, complete spatial distribution data of geological hazard risk is generated. This data can be directly used for visualization rendering output, intuitively displaying the spatial location, boundary range, and distribution pattern of areas with different risk levels. For the finally identified high-risk red warning areas, risk parameters can be further calculated. The overall risk level of the area is obtained by calculating the average parameters of the regional mesh elements, using the following formula:

[0109]

[0110] in, This represents the average probability of instability corresponding to the final red alert area. This represents the total number of finite element mesh elements contained within the final red alert area. The first in the red alert area The probability of element instability corresponding to each finite element mesh element.

[0111] Specifically, by simultaneously calculating the spatial centroid coordinates and overall coverage area of ​​the final red alert area, the core location and impact scale of high-risk hazard areas are determined. Combining the current alert sampling time with the preset prediction duration, the effective time window corresponding to this alert result is delineated, clarifying the effective monitoring and control period for the alert risk. Integrating the average instability probability corresponding to the aforementioned final red alert area yields emergency red information. Finally, based on the spatial distribution data of the entire risk area and the refined emergency red information, by merging the spatial correction results and temporal verification results of risk areas at each level, geological disaster early warning results with spatial accuracy, temporal stability, and hierarchical zoning characteristics can be generated, effectively solving the problems of high false alarm rate, coarse granularity, and poor stability of traditional early warning methods.

[0112] Based on the same inventive concept, this application also provides a digital twin-based geological disaster early warning system for implementing the aforementioned digital twin-based geological disaster early warning method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the digital twin-based geological disaster early warning system provided below can be found in the above-described limitations of the digital twin-based geological disaster early warning method, and will not be repeated here.

[0113] In one exemplary embodiment, such as Figure 2 As shown, a geological disaster early warning system 40 based on digital twins is provided, including:

[0114] The twin modeling module 41 is used to establish an initial digital twin model based on multi-source monitoring data of the geological disaster monitoring area. The initial digital twin model includes discretized finite element mesh elements, the binding interface between each finite element mesh element and the corresponding monitoring data, and the initial mechanical parameter field, initial stress field and initial displacement field corresponding to each finite element mesh element.

[0115] The evolution inference module 42 is used to extract the state vector, including the field parameters of each finite element mesh element, from the current digital twin model at the current moment; using the state vector as the starting state, an evolution path set is obtained through a physical information deep reinforcement learning model; the evolution path set includes an optimal evolution path and multiple branch evolution paths, and each evolution path in the evolution path set contains a state vector sequence discretely time-step from the current moment to the prediction time window and the evolution action corresponding to each time step; wherein, when the current moment is the first warning, the current digital twin model is the initial digital twin model;

[0116] The dynamic update module 43 is used to take the evolution paths in the evolution path set as the current set members of the ensemble Kalman filter; based on the monitoring data received at the current time, the ensemble Kalman filter algorithm with path trajectory similarity weighting is used to dynamically update the predicted state of each set member at the current time, so as to obtain the posterior set of set members and the calibrated digital twin model. The posterior set of set members contains the posterior state vector of each set member at the current time and its corresponding posterior weight.

[0117] The early warning generation module 44 is used to extract the damage intrinsic state variables of each finite element mesh element from the calibrated digital twin model; calculate the instability probability of each finite element mesh element based on the posterior set membership; classify the instability probability to obtain the preliminary early warning level of each finite element mesh element; and apply spatial-temporal dual constraints to the preliminary early warning level to generate graded and partitioned early warning results.

[0118] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the geological disaster early warning method based on digital twins of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0119] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the digital twin-based geological disaster early warning method of this application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0120] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the patent scope of the embodiments of this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the embodiments of this application, and these all fall within the protection scope of the embodiments of this application.

Claims

1. A geological disaster early warning method based on digital twins, characterized in that, The method includes: S1: Based on the multi-source monitoring data of the geological disaster monitoring area, an initial digital twin model is established. The initial digital twin model includes discretized finite element mesh elements, the binding interface between each finite element mesh element and the corresponding monitoring data, and the initial mechanical parameter field, initial stress field and initial displacement field corresponding to each finite element mesh element. S2: Extract the state vector at the current moment, including the field parameters corresponding to each finite element mesh element, from the current digital twin model; using the state vector as the starting state, obtain an evolution path set through a physical information deep reinforcement learning model; the evolution path set includes an optimal evolution path and multiple branch evolution paths, and each evolution path in the evolution path set contains a state vector sequence discretely time-step from the current moment to the prediction time window and the evolution action corresponding to each time step; wherein, when the current moment is the first warning, the current digital twin model is the initial digital twin model; S3: The evolution paths within the set of evolution paths are used as the current set members of the ensemble Kalman filter; based on the monitoring data received at the current time, the predicted state of each set member at the current time is dynamically updated using the ensemble Kalman filter algorithm with path trajectory similarity weighting, to obtain the posterior set of set members and the calibrated digital twin model, wherein the posterior set of set members contains the posterior state vector of each set member at the current time and its corresponding posterior weight; S4: Extract the damage intrinsic state variables of each finite element mesh element from the calibrated digital twin model; calculate the instability probability of each finite element mesh element based on the posterior set membership; classify the instability probability to obtain the preliminary warning level of each finite element mesh element; apply spatial-temporal dual constraints to the preliminary warning level to generate a graded and partitioned warning result.

2. The method according to claim 1, characterized in that, In step S2, the evolution path set is obtained using the state vector as the starting state and through a physical information deep reinforcement learning model, including: Multiple sets of state replicas are constructed based on the state vector. Gaussian noise is superimposed on each of the state replicas to generate each initial set member. An equal initial weight is assigned to each of the initial set members. Determine whether the current moment is the first warning. If the current moment is the first warning, set an initial exploration temperature for each initial set member. Otherwise, determine the exploration temperature of the corresponding initial set member based on the posterior weight of each set member obtained from the previous warning moment. The posterior weight of each set member is negatively correlated with the exploration temperature of the initial set member. Based on the physical information deep reinforcement learning model, starting from the state corresponding to each of the initial set members, the optimal evolution path is generated by greedy deduction, and combined with the exploration temperature, multiple branch evolution paths are generated by Boltzmann exploration strategy. The optimal evolution path is combined with multiple branch evolution paths to form the evolution path set. The state vector sequence of each evolution path in the evolution path set extending from the current time to the prediction time window is recorded in sequence, and the evolution actions executed at each time step are saved. The evolution actions include seepage driving actions, mechanical loading actions, strength degradation actions, and deformation calculation actions.

3. The method according to claim 1, characterized in that, When the current time is the first warning, in step S3, based on the monitoring data received at the current time, a path trajectory similarity-weighted ensemble Kalman filter algorithm is used to dynamically update the predicted state of each ensemble member at the current time, resulting in a posterior ensemble member set and a calibrated digital twin model, including: S21: Extract the state at the current time from the evolution path corresponding to each of the set members, and use the state at the current time as the predicted state of each of the set members at the current time; add a preset model error term to the predicted state to correct the state and obtain the corrected predicted state; S22: Obtain the monitoring data received at the current time, map the corrected prediction state corresponding to each set member to the observation space through a preset observation operator, and solve the state residual between the predicted observation value corresponding to the corrected prediction state of each set member and the measured value corresponding to the monitoring data. S23: Based on the state residual, obtain the initial residual weighted sum of squares corresponding to each set member, and calculate the path trajectory similarity weight corresponding to each set member based on the initial residual weighted sum of squares; S24: Based on the path trajectory similarity weights corresponding to each set member, calculate the weighted set mean and weighted set covariance respectively, and combine the weighted set mean and weighted set covariance to obtain the Kalman gain matrix; S25: Based on the Kalman gain matrix, match the state residuals corresponding to each set member, iteratively correct the predicted state of each set member at the current time, and generate the posterior state vector corresponding to each set member at the current time; S26: Take the path trajectory similarity weights corresponding to each set member as the posterior weights of each set member at the current time, and combine the posterior state vectors of each set member at the current time with the posterior weights to obtain the posterior set member set; S27: Using the posterior state vectors of each member of the posterior set as the calibration basis, perform global state synchronization calibration on the current digital twin model, update the mechanical parameter field, stress field, and deformation field inside the current digital twin model, and obtain the calibrated digital twin model.

4. The method according to claim 1, characterized in that, When the current moment is not the first warning, the posterior set membership is obtained through the following steps: S31: Based on the action sequence from the previous warning time to the current time in the evolution path corresponding to each set member, and taking the posterior state corresponding to each set member at the previous warning time as the starting point for deduction, state prediction is performed to obtain the predicted state corresponding to each set member at the current time; the predicted state is corrected by superimposing a preset model error term on the predicted state to obtain the corrected predicted state. S32: Obtain the monitoring data received at the current time, map the corrected prediction state corresponding to each set member to the observation space through a preset observation operator, and solve the state residual between the predicted observation value corresponding to the corrected prediction state of each set member and the measured value corresponding to the monitoring data. S33: Retrieve the historical residual weighted sum of squares accumulated from the first warning to the previous warning time for each of the set members, add the state residual to the historical residual weighted sum of squares to obtain the latest residual weighted sum of squares for each of the set members, and calculate the path trajectory similarity weight for each of the set members based on the latest residual weighted sum of squares. S34: Based on the path trajectory similarity weights corresponding to each set member, calculate the weighted set mean and weighted set covariance respectively, and combine the weighted set mean and weighted set covariance to obtain the Kalman gain matrix; S35: Based on the Kalman gain matrix, match the state residuals corresponding to each set member, iteratively correct the predicted state of each set member at the current time, and generate the posterior state vector corresponding to each set member at the current time; S36: Determine the path trajectory similarity weights corresponding to each set member as the posterior weights corresponding to each set member at the current time; combine the posterior state vectors corresponding to each set member and the posterior weights to obtain the posterior set member set.

5. The method according to claim 1, characterized in that, In step S4, the damage eigenstate variables of each finite element mesh element are extracted from the calibrated digital twin model; and the instability probability of each finite element mesh element is calculated based on the posterior set membership. The instability probabilities are classified to obtain the preliminary warning level for each finite element mesh element, including: The geotechnical mechanics parameters and mechanical response parameters corresponding to each finite element mesh element are retrieved from the calibrated digital twin model. The geotechnical mechanics parameters include cohesion and internal friction angle, and the mechanical response parameters include equivalent plastic strain, mean stress and pore water pressure. The geotechnical mechanics parameters and mechanical response parameters corresponding to each finite element mesh element are integrated to form the feature vector corresponding to each finite element mesh element. The feature vector is input into a sparse autoencoder to perform feature analysis calculation and output the damage eigenstate variable corresponding to each finite element mesh element. The value of the damage eigenstate variable is between 0 and 1. Traverse each set member in the posterior set member set, and for each finite element mesh element, pre-set an initial cumulative probability value, and sequentially determine whether each set member satisfies a preset condition on the finite element mesh element; wherein, the preset condition includes the Mohr-Coulomb yield function being greater than a preset judgment value, and the equivalent plastic strain rate exceeding a preset strain rate threshold. If a single member of the set satisfies the determination condition, the posterior weight corresponding to the member of the set is accumulated into the initial accumulated probability value of the corresponding finite element mesh element. After traversal, the final instability probability corresponding to each finite element mesh element is obtained. Retrieve the safety factor of each finite element mesh element in the calibrated digital twin model, and set low-risk, medium-risk, and high-risk thresholds for each finite element mesh element based on the safety factor; The instability probability corresponding to each finite element mesh element is compared with the low-risk threshold, the medium-risk threshold, and the high-risk threshold, respectively. Based on the comparison results, the preliminary warning level corresponding to each finite element mesh element is determined.

6. The method according to claim 1, characterized in that, In step S4, the initial warning level is subject to both spatial and temporal constraints to generate tiered and zoned warning results, including: The finite element mesh cells with a preliminary warning level of red are used as red finite element mesh cells. A density clustering algorithm is used to extract the first connected mesh cluster with a first spatial area greater than a first preset area threshold. The first connected mesh cluster is designated as a candidate red warning area. The preliminary warning level of the isolated red finite element mesh cells that are not included in any of the first connected mesh clusters is downgraded to orange. The finite element mesh cells with an initial warning level of orange are used as orange finite element mesh cells. The density clustering algorithm is used to extract the second connected mesh clusters with a second spatial area greater than the second preset area threshold. The initial warning level of the isolated orange finite element mesh cells that are not included in any of the second connected mesh clusters is downgraded to yellow. Retrieve historical warning level data from at least two consecutive warning times prior to the current time, perform time-series continuity verification on each candidate red warning area, check whether the warning level corresponding to each candidate red warning area remains red within three consecutive sampling periods, and obtain the verification result; If the verification result shows that the candidate red warning area meets the time-series continuous verification requirements, then the final warning level of the finite element mesh unit corresponding to the candidate red warning area is locked as red, and the final red warning area is obtained; otherwise, the final warning level of the finite element mesh unit corresponding to the candidate red warning area is downgraded to orange warning level, and the continuous detection status is marked. The finite element mesh elements with a preliminary warning level of yellow are used as yellow finite element mesh elements. Combined with the final red warning area, the orange finite element mesh elements, and the yellow finite element mesh elements, the spatial distribution data is visualized and output. The centroid coordinates, coverage area, and average instability probability corresponding to the final red warning area are statistically analyzed, and the corresponding effective time window is determined to generate emergency red information. Based on the emergency red information and the spatial distribution data, the graded and zoned warning results are generated. The average instability probability is obtained by calculating the mean of the instability probabilities of the finite element mesh elements corresponding to the final red warning area.

7. The method according to claim 4, characterized in that, The formula for calculating the path trajectory similarity weight is as follows: in, Indicates the first At the current moment, the members of the set are... The path trajectory similarity weights; For the current time, The warning process begins from the first warning. A historical warning moment; For the first Each set member in The predicted observation vector is generated by mapping the corrected predicted state of each member of the set to the observation space at each moment. for The measured observation vector corresponding to the monitoring data received at the current time; The square of the Mahalanobis distance is used to quantify the deviation between the predicted and measured observations. To observe the noise covariance matrix, it is matched with the preset observation operator to characterize the noise level differences in each dimension of the monitoring data; For temperature parameters; This represents the total number of members in the set.

8. A geological disaster early warning system based on digital twins, characterized in that, The system includes: The twin modeling module is used to establish an initial digital twin model based on multi-source monitoring data of the geological disaster monitoring area. The initial digital twin model includes discretized finite element mesh elements, binding interfaces between each finite element mesh element and the corresponding monitoring data, and initial mechanical parameter fields, initial stress fields, and initial displacement fields corresponding to each finite element mesh element. The evolutionary deduction module is used to extract the state vector, including the field parameters corresponding to each finite element mesh element, from the current digital twin model at the current moment; using the state vector as the starting state, an evolutionary path set is obtained through a physical information deep reinforcement learning model; the evolutionary path set includes an optimal evolutionary path and multiple branch evolutionary paths, and each evolutionary path in the evolutionary path set contains a state vector sequence discretely time-step from the current moment to the prediction time window and the evolutionary action corresponding to each time step; wherein, when the current moment is the first warning, the current digital twin model is the initial digital twin model; The dynamic update module is used to take the evolution paths in the evolution path set as the current set members of the ensemble Kalman filter; based on the monitoring data received at the current time, the predicted state of each set member at the current time is dynamically updated using the ensemble Kalman filter algorithm with path trajectory similarity weighting, to obtain the posterior set of set members and the calibrated digital twin model. The posterior set of set members contains the posterior state vector of each set member at the current time and its corresponding posterior weight. The early warning generation module is used to extract the damage intrinsic state variables of each finite element mesh element from the calibrated digital twin model; calculate the instability probability of each finite element mesh element according to the posterior set membership; classify the instability probability to obtain the preliminary early warning level of each finite element mesh element; and apply spatial-temporal dual constraints to the preliminary early warning level to generate a graded and partitioned early warning result.

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

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images