An intelligent inversion system for three-dimensional deformation field of land subsidence in an area with super large water level drawdown

Abstract

The present application relates to the technical field of hydrogeological monitoring, in particular to a kind of super large water level drawdown area ground subsidence three-dimensional deformation field intelligent inversion system, comprising: multi-source data acquisition and forward modeling unit, construct three-dimensional geological structure model, and obtain theoretical subsidence field based on groundwater level drawdown data;Intelligent inversion and parameter optimization unit is used to build subsidence residual field, generate updated hydrogeological parameters;Risk assessment and grade division unit, calculate the model health index, and calculate the system safety distance in combination with the model health index, and then generate risk grade signal;Closed-loop control and strategy execution unit is used to respond to risk grade signal, executes control strategy to dynamically correct the operation of intelligent inversion and parameter optimization unit.

Description

Technical Field

[0001] This invention relates to the field of hydrogeological monitoring technology, and more specifically, to an intelligent inversion system for three-dimensional deformation field of ground subsidence in areas with ultra-large drawdown. Background Technology

[0002] In areas with extremely deep drawdowns, significant drops in groundwater levels often trigger large-scale land subsidence, posing a serious threat to the safety of urban infrastructure, buildings, and underground engineering projects. Current monitoring technologies primarily rely on surface deformation sensor networks, such as GPS, InSAR, and groundwater level monitoring equipment, to collect real-time data on surface displacement sequences and water level changes. However, in actual monitoring, the data is generally subject to significant noise interference, including atmospheric disturbances, equipment measurement errors, and random fluctuations such as environmental vibrations. Furthermore, some collected data have weak correlations with the physical mechanisms of subsidence, such as invalid signals generated by shallow soil disturbances or irrelevant environmental factors. Directly basing subsidence field prediction and risk assessment on such noisy data often leads to large discrepancies between theoretical models and actual monitoring results, failing to accurately reflect the deformation characteristics of three-dimensional geological structures. In addition, traditional inversion methods often employ simplified one-dimensional consolidation theories or empirical parameters, making it difficult to handle the nonlinear spatiotemporal variability of hydrogeological parameters. This results in inaccurate subsidence predictions, ambiguous risk level classifications, and potentially overlooking hazardous areas, delaying the implementation of prevention and control measures, and causing engineering safety accidents. Existing technologies have not effectively solved the problem of synergistic optimization of data noise filtering, high-precision inversion of three-dimensional deformation fields, and dynamic risk assessment, resulting in insufficient reliability of monitoring systems. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent inversion system for three-dimensional deformation field of ground subsidence in areas with ultra-large drawdown, which can effectively filter noise interference in monitoring data, realize high-precision inversion of three-dimensional deformation field, improve the accuracy and real-time performance of risk assessment, and avoid potential engineering safety accidents.

[0004] Specifically, this invention provides an intelligent inversion system for three-dimensional deformation field of ground subsidence in areas with ultra-large drawdown, including: a multi-source data acquisition and forward modeling unit, used to acquire surface deformation sequence and groundwater drawdown data, construct a three-dimensional geological structure model, and calculate the theoretical subsidence field based on the groundwater drawdown data; The intelligent inversion and parameter optimization unit is used to compare the theoretical settlement field with the actual monitored settlement field obtained from the surface deformation sequence to construct the settlement residual field, and to use a Bayesian neural network to invert the hydrogeological parameters of the three-dimensional geological structure model to generate updated hydrogeological parameters. The risk assessment and classification unit is used to calculate the model health index based on the settlement residual field and updated hydrogeological parameters, and to calculate the system safety distance in combination with the model health index, thereby generating a risk level signal. The closed-loop control and strategy execution unit is used to respond to risk level signals and execute control strategies to dynamically correct the operation of the intelligent inversion and parameter optimization unit.

[0005] Preferably, the calculation of the theoretical settlement field includes the following steps: based on the effective stress principle and the porous media seepage theory, by extending the Terzaghi one-dimensional consolidation theory in three-dimensional space, the deformation of the soil skeleton caused by the change in groundwater level is solved, which is characterized by the following formula:

[0006] in, This represents the theoretical total settlement. This represents the total thickness of the compressed soil layer; This is the layered compression factor, in units of... ; This represents the effective stress increment. For depth; For time.

[0007] Preferably, the intelligent inversion and parameter optimization unit takes minimizing the global norm of the settlement residual field as the optimization objective and performs nonlinear mapping and probabilistic inversion of hydrogeological parameters.

[0008] Preferably, the calculation of the model health index in the risk assessment and grading unit includes the following steps: S1: The degree of information conflict is obtained using the following formula:

[0009] in To the degree of conflict of new information, The number of grid points for the new observation data. This represents the predicted settlement value for the region corresponding to the new data using the current model. For the new observed settlement value, This represents the current maximum settlement across the entire area. S2: The normalized spatiotemporal mean settlement residual is obtained using the following formula:

[0010] in For the normalized spatiotemporal mean settlement residual, It is the total number of spatial grid points; S3: The norm of the relative rate of change of the key parameter is obtained using the following formula:

[0011] in The norm of the relative rate of change of the key parameter. and These are the hydrogeological parameter vectors for the current and previous time steps, respectively. The second norm of a vector; S4: Weighted summation of the normalized spatiotemporal mean settlement residual, the norm of the relative rate of change of key parameters, and the degree of innovation conflict is performed, and the reciprocal of the summation result is taken to generate the model health index.

[0012] This is the model's health index; These are their respective weighting coefficients.

[0013] Preferably, the formula for calculating the system's safe distance is:

[0014] in, For system safety distance; To allow maximum settlement; This represents the current maximum settlement. This is the health index of the model.

[0015] Preferably, the risk assessment and classification unit compares the system safety distance with a preset first threshold and a second threshold; When the system's safe distance exceeds the first threshold, a safety risk level signal is generated. When the system's safe distance is less than or equal to the first threshold and greater than the second threshold, a risk level signal of concern is generated. When the system's safe distance is less than or equal to the second threshold, a warning risk level signal is generated.

[0016] Preferably, when the risk level signal is safe, the closed-loop control and strategy execution unit executes the normal operation strategy to maximize the weights of the Bayesian neural network in order to achieve rapid parameter updates.

[0017] Preferably, when the risk level signal is of concern, the closed-loop control and strategy execution unit executes a prudent operation strategy, reduces the weight of the Bayesian neural network, increases the constraint weight of the forward physical model, and increases the data assimilation frequency.

[0018] Preferably, when the risk level signal is an early warning, the closed-loop control and strategy execution unit executes an emergency decision-making strategy, issues a risk warning, temporarily freezes the parameter update function of the intelligent inversion and parameter optimization unit, and performs a global forced recalibration of the three-dimensional geological structure model.

[0019] By adopting the above scheme, the present invention has the following advantages and beneficial effects: The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas provided by the present invention effectively solves the problems of data noise interference and high-precision inversion of three-dimensional deformation field by integrating multi-source data acquisition, Bayesian neural network intelligent inversion, dynamic evaluation of model health and closed-loop control mechanism. It realizes real-time optimization of risk assessment and early warning of engineering safety, effectively filters noise interference in monitoring data, realizes high-precision inversion of three-dimensional deformation field, improves the accuracy and real-time performance of risk assessment, and avoids potential engineering safety accidents. Attached Figure Description

[0020] Figure 1 This is a flowchart of the intelligent inversion system for three-dimensional deformation field of ground subsidence in ultra-large drawdown areas according to the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0026] Please see Figure 1 This invention provides an intelligent inversion system for three-dimensional deformation field of land subsidence in areas with ultra-large drawdown, comprising: The multi-source data acquisition and forward modeling unit is used to acquire surface deformation sequences and groundwater level drawdown data, construct a three-dimensional geological structure model, and calculate the theoretical subsidence field based on the groundwater level drawdown data. The intelligent inversion and parameter optimization unit is used to compare the theoretical settlement field with the actual monitored settlement field obtained from the surface deformation sequence to construct the settlement residual field, and to use a Bayesian neural network to invert the hydrogeological parameters of the three-dimensional geological structure model to generate updated hydrogeological parameters. The risk assessment and classification unit is used to calculate the model health index based on the settlement residual field and updated hydrogeological parameters, and to calculate the system safety distance in combination with the model health index, thereby generating a risk level signal. The closed-loop control and strategy execution unit is used to respond to risk level signals and execute control strategies to dynamically correct the operation of the intelligent inversion and parameter optimization unit. This invention provides an intelligent inversion system for three-dimensional deformation field of ground subsidence in areas with ultra-large drawdown. The system includes a multi-source data acquisition and forward modeling unit. The purpose of this unit is to integrate multi-dimensional observation data from the ground and underground, and to construct a physical model capable of initially predicting ground subsidence. In this embodiment, this unit continuously acquires surface deformation sequences covering the entire monitoring area using a spaceborne interferometric synthetic aperture radar and a GPS reference station deployed on the surface. Simultaneously, it acquires groundwater level drawdown data in real time using stratified markers and water level gauges installed in boreholes at different depths. All acquired data is transmitted to a central processor for gridding processing under a unified spatiotemporal reference. The surface deformation sequence refers to the line-of-sight deformation dataset periodically acquired using InSAR technology, which provides large-scale, high-density actual subsidence observations as the baseline true values ​​for subsequent model inversion. Spaceborne interferometric synthetic aperture radar (IASAR) data used to acquire surface deformation sequences can utilize C-band data from the ESA's Sentinel-1 satellite, with a revisit period of 6 days and a spatial resolution of up to 5 meters. Surface deformation sequences refer to line-of-sight deformation datasets obtained by processing periodic InSAR data using small baseline set technology to provide large-scale, high-density actual subsidence observations. Global Positioning System (GPS) reference stations can employ TrimbleNet R9 receivers, combined with RTK (Real-Time Kinematics) technology, achieving a vertical accuracy better than 3 millimeters, used for calibrating and constraining the absolute values ​​of InSAR data. This line-of-sight deformation can be solved into vertical and horizontal deformation components by combining multi-source observations, such as GPS or using rise and fall orbit data, to better characterize the surface deformation field; Groundwater level drawdown data refers to the magnitude of the decrease in groundwater level over time as monitored by a water level gauge. It serves as a key input to the physical model driving settlement. Based on prior information about the mechanical properties of soil and rock from the geological survey report, a three-dimensional geological structure model is constructed. This model represents the spatial distribution of the main aquifers and impermeable layers in the form of a three-dimensional mesh. The three-dimensional geological structure model is a discretized digital model that reflects the distribution of the underground medium, providing a spatial framework and initial parameters for physical calculations. This model is constructed by conducting three-dimensional geological surveys of the monitoring area to obtain information on soil layering, soil and rock compression modulus, permeability coefficient, and other parameters at different depths, and then discretizing them using tetrahedral or hexahedral meshes. For example, in a certain monitoring area, the geological structure can be divided into clay layer, silty clay layer, and sand layer, and the model will represent the spatial distribution and thickness of these soil layers in the form of a three-dimensional mesh. Based on this model, the unit uses groundwater drawdown data as input to calculate and generate a two-dimensional theoretical settlement field. The system further includes an intelligent inversion and parameter optimization unit. The purpose of this unit is to dynamically correct the uncertain parameters in the physical model using actual observation data, so as to improve the prediction accuracy of the model. In this embodiment, the unit compares the theoretical settlement field output by the aforementioned multi-source data acquisition and forward modeling unit with the actual monitored settlement field obtained by processing the surface deformation sequence within the same spatiotemporal grid, thereby constructing a settlement residual field. The settlement residual field refers to the difference distribution field between the theoretical calculation results and the actual observation results. Its function is to quantify the prediction error of the current model and serve as the objective function for subsequent parameter optimization. A pre-trained Bayesian neural network is used to invert the core hydrogeological parameters in the three-dimensional geological structure model, mainly the layered compression coefficients of each soil layer unit. The Bayesian neural network is a machine learning model that combines Bayesian probability theory and neural network structure. Its function is to establish a nonlinear mapping relationship between settlement residuals and hydrogeological parameters, and output the estimated values ​​of the parameters and their uncertainties in probabilistic form. The network can adopt a three-layer feedforward network structure, namely an input layer, a hidden layer, and an output layer. The number of neurons in the input layer matches the number of grid points in the settlement residual field. The hidden layer can contain 100 neurons, and the activation function can be the ReLU function. The number of neurons in the output layer matches the number of hydrogeological parameters to be inverted, such as the number of layered compression coefficients. The network uses Monte Carlo variational inference as the training algorithm, with the goal of minimizing the Kullback-Leibler (KL) divergence between the predicted and true values. The network aims to minimize the global settlement residual field. It adjusts the network weights through the backpropagation algorithm to generate a set of updated hydrogeological parameters, which are used to iteratively correct the three-dimensional geological structure model. The system also includes a risk assessment and classification unit. This unit aims to comprehensively and quantitatively assess the health status of the entire inversion system and the safety status of the external physical world, and output a clear risk level. In this embodiment, this unit calculates a comprehensive model health index based on the settlement residual field and the updated hydrogeological parameters output by the intelligent inversion and parameter optimization unit. The model health index is a dimensionless index used to quantify the prediction accuracy, parameter stability, and data consistency of the inversion system, and its role is to assess the reliability of the model itself. Then, combining the maximum allowable settlement in the region determined by the engineering design specifications with the currently monitored maximum settlement in the entire area, and integrating the aforementioned model health index, the system safety distance is calculated. The system safety distance is a dynamic risk index that comprehensively considers the physical safety margin and the model credibility, and its role is to directly link the abstract model state with the engineering safety in the real world. By comparing the system safety distance with two preset thresholds, a corresponding risk level signal is generated, which is divided into three levels: safe, attention, and warning. Preset threshold and This can be determined through regression analysis of historical data. For example, in a certain city, it can be... Set to 20, The threshold is set to 10; these values ​​are derived from statistical analysis of a large amount of historical settlement data and related engineering incident cases to ensure the statistical significance of the threshold setting and the effectiveness of risk management. The system includes a closed-loop control and strategy execution unit. The purpose of this unit is to adaptively adjust the operating strategies of the preceding units of the system based on the results of risk assessment, forming a complete feedback control closed loop. In this embodiment, this unit is designed to respond to different risk level signals generated by the risk assessment and classification unit. Upon receiving the signal, the unit will automatically select and execute the corresponding control strategy from the preset strategy library. This strategy directly affects the operating mode of the intelligent inversion and parameter optimization unit, such as adjusting the weight of its internal algorithm or the frequency of data processing, thereby achieving dynamic correction and proactive management of system performance. This invention constructs a complete system encompassing forward modeling, intelligent inversion, risk assessment, and closed-loop control, enabling precise simulation and proactive risk management of land subsidence in areas with extremely deep drawdowns. It not only continuously optimizes geological model parameters using actual observation data to improve prediction accuracy but also dynamically assesses the system's own credibility, combining this credibility with the safety margin of the physical world to generate a risk level that guides decision-making. Through adaptive closed-loop control, the system achieves a dynamic balance between computational efficiency, prediction accuracy, and long-term credibility, effectively mitigating the risk of silent collapse due to model failure and significantly enhancing the intelligence and foresight of land subsidence disaster prevention and control.

[0027] The calculation process of the theoretical settlement field is as follows: Based on the effective stress principle and the porous media seepage theory, the deformation of the soil skeleton caused by the change of groundwater level is solved by extending the integral of Terzaghi's one-dimensional consolidation theory in three-dimensional space. The driving input for the calculation is the effective stress increment calculated from the real-time monitored groundwater level drawdown data. The key physical parameter for the calculation is the stratified compressibility coefficient, and the initial value of the stratified compressibility coefficient is preset according to the geological survey report. In this embodiment, the calculation process of the theoretical settlement field is specified. The purpose of this specification is to clarify the core physical mechanism and mathematical model for predicting settlement. In this embodiment, the calculation process strictly follows the basic theories of rock and soil mechanics, namely, based on the effective stress principle and the seepage theory of porous media. The core mathematical implementation method is to extend the Terzaghi one-dimensional consolidation theory in three-dimensional space to solve for the deformation of the soil skeleton caused by changes in groundwater level. The calculation is represented by the following formula:

[0028] in, This represents the theoretical total settlement. The total thickness of the compressible soil layer is determined by a three-dimensional geological structure model. This is the layered compression factor, in units of... (or ); This represents the effective stress increment. For depth; For time; This formula is based on a simplified model of one-dimensional consolidation theory, mainly capturing the primary causal relationship between effective stress increment and compressive deformation. Although it simplifies complex physical processes such as soil anisotropy and interlayer coupling effects, this invention dynamically optimizes the layered compressibility coefficient through intelligent inversion units. This allows the parameter to adaptively absorb and characterize these un-explicitly modeled complex factors in a data-driven manner, thereby improving the overall physical fidelity of the model. Terzaghi's one-dimensional consolidation theory extended in three-dimensional space is usually achieved using numerical calculation methods such as the finite element method or the finite difference method. This method discretizes the three-dimensional geological structure model into grid cells and applies the basic equations of consolidation theory to each cell. By coupling the seepage and deformation equations, iteratively solving the effective stress change and soil deformation caused by the drawdown at the time step. Although Terzaghi's theory provides a solid physical foundation, it has limitations in dealing with the nonlinearity, rheological effects, and complex stress paths of soil. The innovation of this invention lies in this: through intelligent inversion and parameter optimization units, the system can use Bayesian neural networks to learn and correct the uncertainties and nonlinearities brought about by the simplified model in actual observation data, thereby dynamically adjusting parameters such as the layered compression coefficient, so that it can better characterize the complex mechanical behavior of soil in the real world. The calculation driver for this formula is the effective stress increment. It is calculated from real-time monitoring data of groundwater level drawdown; It depends on the depth and time The changing function directly reflects the additional load exerted on the soil skeleton by groundwater extraction activities; the key physical parameter in this calculation is the stratified compressibility coefficient. It characterizes the compressive deformation capacity of soil under a unit effective stress increment; Spatial distribution, with depth Changes are the core factor affecting the accuracy of settlement calculations; the initial value of the stratified compression coefficient is preset according to the geological survey report and is dynamically optimized in the subsequent inversion process; By employing a mathematical model based on well-defined physical mechanisms, the effective stress principle, and consolidation theory to calculate the theoretical settlement field, a solid theoretical foundation for the system's predictions is ensured. This not only makes the initial predictions of the model highly reasonable but also provides a physically meaningful optimization object for subsequent intelligent inversion, including the hierarchical compression coefficient. This avoids the black box problem of purely data-driven models, thereby improving the stability and interpretability of the entire system under complex geological conditions.

[0029] The intelligent inversion and parameter optimization unit takes minimizing the global norm of the settlement residual field as the optimization objective and performs nonlinear mapping and probabilistic inversion of hydrogeological parameters. The output of the Bayesian neural network includes the most probable values ​​and uncertainties of hydrogeological parameters. The most probable values ​​are used to update the three-dimensional geological structure model, and the uncertainties are used to evaluate the stability of the inversion process. In this embodiment, the internal operating mechanism of the intelligent inversion and parameter optimization unit is further defined. This definition aims to clarify the objectives and methods of parameter optimization to ensure the efficiency and scientific rigor of the inversion process. In this embodiment, the unit sets a clear optimization objective: minimizing the global norm of the settlement residual field. The global norm refers to a mathematical measure of the settlement residual values ​​of all spatial grid points within the entire monitoring area; for example, the L2 norm, or root mean square error, serves to condense the distributed error field into a single scalar value, facilitating optimization. Based on this objective, the unit utilizes a Bayesian neural network to perform nonlinear mapping and probabilistic inversion of hydrogeological parameters. The output of this Bayesian neural network is designed to include two parts: The most likely values ​​of hydrogeological parameters are the mean of their posterior probability distributions. The uncertainty of the parameters is the variance of the posterior probability distribution. The most likely value refers to the estimate that the model believes is closest to the true value of the parameter after combining the current observation data. It will be used to update the three-dimensional geological structure model to improve the accuracy of the next round of forward modeling. Uncertainty quantifies the model's confidence in the current parameter estimate. High uncertainty means that the data does not constrain the parameters enough. This information will be used to evaluate the stability of the inversion process and can be used as one of the inputs of the risk assessment unit. By setting the optimization objective as minimizing the global norm, a clear and quantifiable convergence direction is provided for the inversion process. Using a Bayesian neural network for probabilistic inversion not only yields the optimal estimates of the parameters but also quantifies their uncertainty. This inversion method, which understands both the "what" and the "why," greatly enhances the robustness of the system. It enables the system to distinguish whether the model is truly accurate or whether the fit is accidental due to insufficient data, thus providing richer and deeper information for subsequent risk assessment and control, and improving the scientific nature of parameter update decisions.

[0030] Example 2 The risk assessment and grading unit is further used for: Determine the normalized spatiotemporal average settlement residual to characterize the prediction accuracy; Quantify the norm of the relative rate of change of key parameters to characterize the stability of model parameters; Calculate the degree of innovation conflict to measure the model's compatibility with new observation data; The calculation method for this indicator is as follows:

[0031] in, : The degree of conflict in new information The number of grid points for the new observation data. This represents the predicted settlement value for the region corresponding to the new data using the current model. For the new observed settlement value, This represents the current maximum settlement across the entire area. We perform a weighted summation of the normalized spatiotemporal mean settlement residual, the norm of the relative rate of change of key parameters, and the degree of innovation conflict, and take the reciprocal of the summation result to generate the model health index. In this embodiment, the risk assessment and grading unit generates the model health index in detail. The purpose of this explanation is to construct a multi-dimensional assessment system that can comprehensively and dynamically quantify the reliability of the inversion system. To achieve this purpose, the generation of the model health index incorporates three key negative indicators. Determine the normalized spatiotemporal mean settlement residual to characterize the prediction accuracy; this index The mean square value of the settlement residual field over the entire spatial domain is calculated and normalized by dividing it by the currently monitored maximum settlement in the region. It directly reflects the degree of deviation between the model output and the real-world observations. The smaller the value, the more accurate the prediction. The norm of the relative rate of change of key parameters is quantified to characterize the stability of model parameters; this indicator By calculating the set of hydrogeological parameters between two consecutive update time steps The relative change of the vector L2 norm is obtained; a healthy model should exhibit smooth convergence of its internal parameters after receiving new data, rather than violent oscillations; the smaller this index value, the more stable the inversion process. The degree of conflict of new information is calculated to measure the model's compatibility with new observational data; this metric The calculation method is as follows: when a batch of new observation data is introduced but has not yet been used to update the model, the current model is used to make a prediction on the region corresponding to the new data, and the normalized residual between the predicted value and the new observation value is calculated; it measures the current model state's ability to accept new information that is about to come. The smaller the value, the stronger the model's generalization ability and the better its compatibility with new data. Normalized spatiotemporal mean settlement residual Key parameter relative rate of change norm and the degree of conflict between new information Perform a weighted summation and take the reciprocal of the summation result to generate the model health index. The calculation formula is as follows:

[0032] in, This is the model's health index; These are their respective weighting coefficients; To normalize the residuals, the calculation inputs are derived from the settlement residual field and the maximum settlement over the entire domain; The parameter change rate is calculated using the current and previous set of hydrogeological parameters. The conflict degree is calculated using the current model and new observation data as inputs; dimensionless weighting coefficients are used. These are preset parameters. The preset weight coefficients refer to values ​​predetermined based on expert experience or historical data regression analysis, reflecting the decision-maker's relative emphasis on prediction accuracy, model stability, and data compatibility, respectively. For example, a judgment matrix is ​​constructed using the analytic hierarchy process (AHP), and multiple domain experts are invited to conduct pairwise comparisons and scores of the relative importance of the three indicators. Finally, a normalized weight vector is calculated as the final weight. The value of ; through this design, The higher the value, the healthier the model;

[0033] in, It is the total number of spatial grid points;

[0034] in, and These are the hydrogeological parameter vectors for the current and previous time steps, respectively. The second norm of a vector; This implementation extends the assessment of model health from a single perspective of predictive accuracy to a multi-dimensional perspective that includes parameter stability and compatibility with new data. This comprehensive assessment mechanism is more thorough and insightful, and can effectively identify sub-healthy model states that, although their short-term predictive accuracy is acceptable, have experienced drastic fluctuations in internal parameters or are no longer adapted to new data. This enables early warning of the risk of silent collapse model failure, greatly improving the foresight and accuracy of system risk assessment.

[0035] Example 3 The calculation process for the system's safe distance is as follows: Obtain the maximum allowable settlement of the area as determined by the engineering design specifications and the maximum settlement of the entire area currently monitored, and calculate the physical settlement safety margin based on the two. The physical settlement safety margin is multiplied by the model health index to generate the system safety distance; The risk assessment and classification unit compares the system's safe distance with a preset first threshold and a second threshold; When the system's safe distance exceeds the first threshold, a safety risk level signal is generated. When the system's safe distance is less than or equal to the first threshold and greater than the second threshold, a risk level signal of concern is generated. When the system's safe distance is less than or equal to the second threshold, a warning risk level signal is generated; In this embodiment, the calculation process of the system safety distance and its application in risk level classification are explained in detail; this explanation aims to establish a bridge connecting the model state and the risks in the physical world, and to transform them into clear and operable risk levels; In this embodiment, the calculation of the system safety distance requires obtaining the maximum allowable settlement of the area as determined by the engineering design specifications. Compared with the current monitored maximum settlement in the entire area The maximum allowable settlement of a region is a fixed threshold set according to the safety requirements of buildings and infrastructure in the region, and is derived from local building safety regulations or engineering design documents. This involves processing observations obtained in real time from the surface deformation sequence; based on these two measurements, the physical settlement safety margin is calculated, which in this embodiment is represented as the physical settlement safety margin, and its value is... ; This physical settlement safety margin is compared with the model health index calculated above. Multiply to generate the system's safe distance. The calculation formula is:

[0036] in, For system safety distance; To allow maximum settlement; This represents the current maximum settlement. This serves as a model health index; this design implements a dynamic penalty mechanism: even if the physical settlement is far from the safety threshold, i.e., the physical margin is large, if the model itself is unreliable, The value is very low, and the calculated system safety distance is... It will also be significantly lowered; In calculation Then, the risk assessment and classification unit compares the system's safe distance with a preset first threshold. Second threshold The comparison is performed; the preset threshold refers to the boundary value predefined according to the risk management strategy to classify different risk levels; the basis for setting it is statistical analysis of a large amount of historical settlement data and related engineering incident cases, for example, historical cases of engineering events of concern level can be used. The 80th percentile of the value is set as An engineering event at the warning level will occur. The 80th percentile of the value is set as This is to ensure the statistical significance of the threshold setting and the effectiveness of risk management; Preset threshold and Having a unit of length, for example, will Set to 20 mm, The value was set at 10 millimeters. These values ​​were derived from statistical analysis of a large amount of historical settlement data and related engineering incident cases. While this model quantifies risk as the product of physical safety margin and model health, simplifying the complexities of multiple factors in reality, this design captures the core relationship between physical and perceived risk, providing a crucial comprehensive indicator for decision-making. Future versions of the system could consider incorporating more risk variables, such as historical groundwater trends and regional infrastructure sensitivity indices, into the calculation of system safety distances to further enhance the model's practicality and robustness. The comparison logic is as follows: When the system safety distance Greater than the first threshold At that time, a security risk level signal is generated; When the system safety distance Less than or equal to the first threshold And greater than the second threshold At that time, a risk level signal is generated; When the system safety distance Less than or equal to the second threshold At that time, a warning risk level signal is generated; To improve the robustness of the system, when the maximum settlement over the entire area is monitored... When the value approaches zero or falls below a certain small threshold, the system will default to a safe state and skip the calculation of the system's safe distance, directly generating a safety risk level signal. In addition, when a data source, such as InSAR or GPS, experiences long-term data loss, the closed-loop control and strategy execution unit will automatically reduce the weight of that data source in the intelligent inversion unit or temporarily freeze its parameter updates to avoid system risks introduced by unreliable data, thereby ensuring the stable operation of the model under various complex and extreme conditions. By defining the system's safety distance by multiplying the physical safety margin by the model health index, this implementation method creatively quantifies the core risk control concept that safety judgments based on unreliable models are inherently unsafe. This allows risk assessment to move beyond focusing solely on the absolute value of physical quantities and simultaneously examine the reliability of the tools upon which those physical quantities are derived, i.e., the model. Combined with a clear, dual-threshold-based grading logic, the system can generate more comprehensive and reliable risk signals that reflect both physical reality and model confidence, providing high-quality decision input for downstream closed-loop control.

[0037] Example 4 When the risk level signal is safe, the closed-loop control and strategy execution unit executes the normal operation strategy to maximize the weights of the Bayesian neural network in order to achieve rapid parameter updates. When the risk level signal is of concern, the closed-loop control and strategy execution unit implements a prudent operation strategy, reduces the weight of the Bayesian neural network, increases the constraint weight of the forward physical model, and increases the data assimilation frequency. When the risk level signal is an early warning, the closed-loop control and strategy execution unit executes the emergency decision-making strategy, issues a risk warning, temporarily freezes the parameter update function of the intelligent inversion and parameter optimization unit, and performs a global forced recalibration of the three-dimensional geological structure model. Updated hydrogeological parameters, such as stratified compressibility coefficient This will be used to replace the old parameter values ​​of the corresponding grid cells in the 3D geological structure model, thereby completing the iterative correction of the model. For the meshed 3D geological model, the attribute parameters of each grid cell will be updated according to the inversion results to ensure that the next forward modeling calculation uses the latest corrected parameter set; In this embodiment, the functions of the closed-loop control and strategy execution unit in response to different risk level signals are defined in detail. The purpose of this definition is to build an adaptive control mechanism with hierarchical response and intelligent trade-offs to ensure that the system can operate in the optimal mode under different risk conditions. When the risk level signal is safe, it indicates that the current physical settlement margin is sufficient and the model is healthy. At this time, the closed-loop control and strategy execution unit executes the normal operation strategy. This strategy aims to maximize data assimilation efficiency and model update speed. The measure is to maximize the weight of the Bayesian neural network in the intelligent inversion and parameter optimization unit. This means that the system trusts the intelligent inversion results driven by new data more to achieve rapid parameter updates and ensure that the model can quickly keep up with the latest changes in the environment. In the intelligent inversion and parameter optimization unit, the weights of the Bayesian neural network can be adjusted by changing the regularization coefficient in its loss function. To control, among which The larger the value, the heavier the penalty to the network, and the lower its weight; when the risk level is safe, It can be set to 0.1 for faster updates; when the risk level is "concerned," It can be increased to 0.5 to reduce network weight; global forced recalibration refers to using regional historical geological data and engineering experience, through numerical simulation methods such as finite element or finite difference, to reinitialize and adjust all parameters in the three-dimensional geological structure model in order to reconstruct the physical basis of the model; When the risk level signal is "Concern," it means that the system's safe distance has narrowed, possibly due to an increase in physical settlement or a decrease in model health. Therefore, the closed-loop control and strategy execution unit implements a prudent operation strategy. The core of this strategy is to improve the model's stability and reliability. Measures include: in the intelligent inversion and parameter optimization unit, actively reducing the weight of the Bayesian neural network while correspondingly increasing the constraint weight of the forward physical model, i.e., placing greater emphasis on prior knowledge with clear physical mechanisms to suppress potential parameter overfitting; in addition, the system automatically increases the data assimilation frequency, for example, by 50%-100%, to obtain more observational information to constrain the inversion process. When the risk level signal is a warning, it indicates that the system's safety distance has reached the safety threshold, and the system faces a high risk. At this time, the closed-loop control and strategy execution unit executes an emergency decision-making strategy. This strategy prioritizes risk avoidance and ensuring decision robustness. Its operations include: the system automatically issues a risk warning to decision-making managers, clearly indicating that the reliability of the current model prediction results has significantly decreased; in the intelligent inversion and parameter optimization unit, its parameter update function is temporarily frozen, so that the system relies entirely on the forward model with clear physical meaning for inference, avoiding the introduction of greater risks by unreliable parameter updates; the system will automatically mobilize a large amount of computing resources to perform a global forced recalibration of the three-dimensional geological structure model in order to fundamentally rebuild the reliability of the model. Through this hierarchical response closed-loop control strategy, the system possesses an immune regulatory capacity similar to that of an organism. It can intelligently and dynamically balance aggressive updates and conservative stability based on its own health status, model health, external environmental threats, and physical subsidence. This adaptive mechanism ensures that the system can operate efficiently when it is safe, make prudent adjustments when risks emerge, and take the most robust emergency measures decisively when facing serious risks. This achieves proactive, closed-loop, and refined management of the entire life cycle of ground subsidence risk, significantly improving the system's long-term stability and decision support capabilities.

[0038] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A smart inversion system for three-dimensional deformation field of land subsidence in areas with ultra-large drawdown, characterized in that, include: The multi-source data acquisition and forward modeling unit is used to collect surface deformation sequences and groundwater level drawdown data, construct a three-dimensional geological structure model, and calculate the theoretical subsidence field based on the groundwater level drawdown data. The intelligent inversion and parameter optimization unit is used to compare the theoretical settlement field with the actual monitored settlement field obtained from the surface deformation sequence to construct the settlement residual field, and to use a Bayesian neural network to invert the hydrogeological parameters of the three-dimensional geological structure model to generate updated hydrogeological parameters. The risk assessment and classification unit is used to calculate the model health index based on the settlement residual field and updated hydrogeological parameters, and to calculate the system safety distance in combination with the model health index, thereby generating a risk level signal. The closed-loop control and strategy execution unit is used to respond to risk level signals and execute control strategies to dynamically correct the operation of the intelligent inversion and parameter optimization unit.

2. The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas according to claim 1, characterized in that, The calculation of the theoretical settlement field includes the following steps: Based on the effective stress principle and the porous media seepage theory, the deformation of the soil skeleton caused by changes in groundwater level is solved by extending the Terzaghi one-dimensional consolidation theory in three-dimensional space, and is characterized by the following formula: in, This represents the theoretical total settlement. This represents the total thickness of the compressed soil layer; This is the layered compression factor, in units of... ; This represents the effective stress increment. For depth; For time.

3. The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas according to claim 1, characterized in that, The intelligent inversion and parameter optimization unit takes minimizing the global norm of the settlement residual field as the optimization objective and performs nonlinear mapping and probabilistic inversion of hydrogeological parameters.

4. The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas according to claim 1, characterized in that, The risk assessment and grading unit, the calculation of the model health index includes the following steps: S1: The degree of information conflict is obtained using the following formula: in To the degree of conflict of new information, The number of grid points for the new observation data. This represents the predicted settlement value for the region corresponding to the new data using the current model. For the new observed settlement value, This represents the current maximum settlement across the entire area. S2: The normalized spatiotemporal mean settlement residual is obtained using the following formula: in For the normalized spatiotemporal mean settlement residual, It is the total number of spatial grid points; S3: The norm of the relative rate of change of the key parameter is obtained using the following formula: in The norm of the relative rate of change of the key parameter. and These are the hydrogeological parameter vectors for the current and previous time steps, respectively. The second norm of a vector; S4: Weighted summation of the normalized spatiotemporal mean settlement residual, the norm of the relative rate of change of key parameters, and the degree of innovation conflict is performed, and the reciprocal of the summation result is taken to generate the model health index. This is the model's health index; These are their respective weighting coefficients.

5. The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas according to claim 4, characterized in that, The formula for calculating the system's safe distance is: in, For system safety distance; To allow maximum settlement; This represents the current maximum settlement. This is the health index of the model.

6. The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas according to claim 5, characterized in that, The risk assessment and classification unit compares the system safety distance with a preset first threshold and a second threshold. When the system's safe distance exceeds the first threshold, a safety risk level signal is generated. When the system's safe distance is less than or equal to the first threshold and greater than the second threshold, a risk level signal of concern is generated. When the system's safe distance is less than or equal to the second threshold, a warning risk level signal is generated.

7. The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas according to claim 6, characterized in that, When the risk level signal is safe, the closed-loop control and strategy execution unit executes the normal operation strategy to maximize the weights of the Bayesian neural network in order to achieve rapid parameter updates.

8. The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas according to claim 6, characterized in that, When the risk level signal is of concern, the closed-loop control and strategy execution unit implements a prudent operation strategy, reduces the weight of the Bayesian neural network, increases the constraint weight of the forward physical model, and increases the data assimilation frequency.

9. The intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large drawdown areas according to claim 6, characterized in that, When the risk level signal is an early warning, the closed-loop control and strategy execution unit executes an emergency decision-making strategy, issues a risk warning, temporarily freezes the parameter update function of the intelligent inversion and parameter optimization unit, and performs a global forced recalibration of the three-dimensional geological structure model.

Images