A geotechnical engineering risk intelligent identification method and system based on geological system state inverse modeling

Abstract

The present application relates to a kind of geotechnical engineering risk intelligent identification method and system based on geological system state reverse modeling, method includes the following steps: collection and fusion multi-source heterogeneous monitoring data;Mechanism core of differentiable that describes seepage, stress and damage coupling process is built;Using the framework that physical information neural network is combined with variational assimilation, reverse solution and observation data physical self-consistent geological system state time-space evolution sequence;Risk mechanism tracing, engineering measure effect deduction and instability precursor identification are carried out based on the system state sequence.Through multi-source data fusion preprocessing, differentiable mechanism core is built, geological physical chemistry process is embedded in model, the physical interpretability and generalization ability of model are enhanced;Through physical information neural network and variational assimilation, reverse modeling is carried out, complete geological system state is inverted in high-dimensional hidden space, risk tracing and dynamic deduction are carried out based on complete system state, risk cause can be quantitatively analyzed and intervention measure effect can be simulated.

Description

Technical Field

[0001] This invention relates to the field of geological exploration, and in particular to a method and system for intelligent identification of geotechnical engineering risks based on geological system state inverse modeling. Background Technology

[0002] Geotechnical engineering geological risk identification is a core component of ensuring the safe construction and operation of major infrastructure projects. Its goal is to detect potential hazards such as landslides, collapses, and ground subsidence in advance and assess their evolution trends. Traditional methods rely on engineers' experience in interpreting discrete borehole, mapping, and monitoring data, which has inherent drawbacks such as low efficiency and high subjectivity, making it difficult to handle the massive, multi-source, and dynamic modern engineering monitoring data.

[0003] To enhance analytical capabilities, technologies in this field have gradually developed: initially, digital management and GIS applications of exploration data were achieved, but these were merely data archives lacking in-depth analytical capabilities. Subsequently, 3D geological visualization technology integrating BIM and GIS enabled the fusion and geometric analysis of multi-source spatial data, but the models were static and could not reflect the dynamic mechanical behavior and risk evolution of soil and rock masses. In recent years, intelligent methods, represented by machine learning and big data fusion, have become cutting-edge, such as predicting soil and rock parameters and constructing comprehensive risk indicators to match historical cases through knowledge graphs and graph neural networks. These methods have improved the level of automation, but fundamental limitations still exist.

[0004] Existing technologies face three main bottlenecks: First, there is a disconnect between data-driven approaches and physical mechanisms. Current methods primarily focus on feature correlation and statistical modeling at the data level. Even when physical formulas are introduced, they are often only superficially integrated with the data model, failing to embed physical conservation laws as core constraints throughout the reasoning process. This results in poor model interpretability and insufficient reliability of extrapolation predictions under sparse data or new operating conditions. Second, risk analysis remains at the level of superficial correlation, lacking state insight and mechanism tracing. Current early warning systems are mostly based on monitoring thresholds or case matching, which can indicate potential risks but cannot quantitatively analyze the contribution of different physical processes such as stress, seepage, and damage to the risk state, let alone quantitatively extrapolate the future effects of engineering measures based on physical laws. Third, there is a gap between observed data and the complete state of the system. Field sensors can only acquire limited physical quantities from sparse points, while geological risk is the result of complex physical field interactions within the soil and rock mass. Existing technologies lack an effective framework that can use sparse multi-source observation data as constraints to inversely reconstruct the complete geological system state that is full-field, continuous, and dynamically evolving.

[0005] In summary, current technologies face significant bottlenecks in three key areas: achieving deep coupling between physical mechanisms and observational data, reconstructing the complete dynamic state of the system, and conducting mechanism tracing and proactive deduction based on system states. These bottlenecks hinder the fundamental leap from risk identification to mechanism-based, proactive decision support. Therefore, an innovative solution is urgently needed to address these issues. Summary of the Invention

[0006] To address the problems of existing geotechnical engineering geological risk identification methods, such as the disconnect between physical mechanisms and observational data, the inability to reconstruct the dynamic state of a complete geological system, and the lack of mechanism-based tracing and extrapolation capabilities, this invention proposes an intelligent geotechnical engineering risk identification method and system based on inverse modeling of geological system states. This method enables the deciphering of the physically self-consistent complete geological system state from sparse observational data, and the quantitative analysis of risk causes and simulation of the effectiveness of prevention and control measures based on this system state.

[0007] This invention provides a method for intelligent identification of geotechnical engineering risks based on geological system state inverse modeling, comprising the following steps: S1. By deploying a multi-type sensor network in the target area and combining it with satellite remote sensing, UAV aerial survey and historical data, collect multi-source heterogeneous monitoring data, perform coordinate unification, time alignment and quality verification on the data, and generate a spatiotemporally aligned multi-source heterogeneous monitoring data set. S2. Based on the fundamental principles of rock and soil mechanics, seepage mechanics and damage mechanics, a set of geophysical and chemical process mechanism cores that are expressed, parameterized and differentiable in the form of partial differential equations or differential algebraic equations are constructed. The mechanism cores include at least models describing seepage-stress coupling, elastoplastic deformation and damage evolution. S3. Define the geological system state representing the complete stress field, seepage field, and damage field as a high-dimensional hidden state vector; use the differentiable mechanism kernel constructed in step S2 as the state equation, and the multi-source heterogeneous monitoring data obtained in step S1 as the observation constraint. Use a solution framework that combines physical information neural network and variational assimilation to solve the spatiotemporal evolution sequence of the geological system state that simultaneously satisfies physical laws and observation data, and obtain a data-calibrated digital geological system dynamic model. S4. Based on the spatiotemporal evolution sequence of the geological system, perform quantitative decomposition analysis of risk-dominant driving factors, forward numerical simulation prediction of the effects of engineering intervention measures, and comprehensive instability precursor pattern identification based on the system phase space. S5. Meta-learning-driven adaptive optimization of mechanism kernel architecture: Construct a knowledge base containing multiple candidate mechanism kernel models and their parameter priors; For new engineering scenarios, utilize initial multi-source heterogeneous monitoring data, and use meta-learning algorithms to search and determine the mechanism kernel combination method and parameter initialization strategy to accelerate the convergence process of the geological system state inverse modeling in step S3. S6. Multidimensional visualization and decision support of geological system states: The geological system states reconstructed in step S3 and deduced in step S4 are dynamically rendered and displayed through a 3D visualization engine. It supports multi-physics field overlay display, spatiotemporal backtracking, profile analysis, virtual measuring point data extraction and comparison of different scenarios. It also integrates and displays risk tracing results, intervention scheme simulation comparison reports and early warning information, and generates a comprehensive risk assessment report.

[0008] Correspondingly, a geotechnical engineering risk intelligent identification system based on geological system state inverse modeling includes: The multi-source data fusion preprocessing module is used to collect real-time monitoring time-series data of the soil and rock mass and the surface through a multi-type sensor network deployed in the target geotechnical engineering area and engineering structure; synchronously integrate wide-area surface deformation time-series data acquired by satellite remote sensing platform, high-precision topographic and geomorphological data acquired by UAV aerial survey platform, and historical geological survey and engineering activity data; perform coordinate system transformation, time series alignment and data quality verification on the multi-source heterogeneous data to generate a spatiotemporally aligned multi-source heterogeneous monitoring data set; the multi-type sensor network includes at least a stress sensor array for monitoring stress changes in soil and rock mass, a seepage sensor array for monitoring pore water pressure distribution, a microseismic monitoring array for capturing micro-fracture events in rock mass, and a displacement monitoring array for monitoring surface and structural displacement; The differentiable mechanism core construction module, based on the fundamental principles of rock and soil mechanics, seepage mechanics, and damage mechanics, constructs a set of parameterized and differentiable mathematical and physical models to characterize the key intrinsic processes of geological risk development. The differentiable mechanism core includes at least: a hydraulic coupling mechanism core describing the coupling of unsaturated seepage and stress field; a stress-strain mechanism core characterizing the elastoplastic deformation and structural surface effects of rock and soil; and a damage evolution mechanism core depicting the cumulative evolution of damage within the rock mass. Each mechanism core is expressed in the form of partial differential equations or differential-algebraic equations, and its model parameters are optimized and adjusted through a data-driven approach. The geological system state inverse modeling module defines the geological system state of the target area as a high-dimensional hidden state vector, which represents the stress field, strain field, pore water pressure field, and damage field of the soil and rock mass in the complete spatiotemporal domain. The constructed differentiable mechanism kernel is used as the state equation describing the evolution of the system state. An observation equation is established, which describes the obtained multi-source heterogeneous monitoring data as a mapping of the geological system state under a specific observation operator. A solution framework combining physical information neural networks and variational assimilation is adopted. With the multi-source heterogeneous monitoring data as constraints, the spatiotemporal evolution sequence of the geological system state that simultaneously satisfies the state equation and the observation equation, as well as the corresponding model parameter field, is solved in reverse by an optimization algorithm to obtain a digital geological system dynamic model calibrated by multi-source data. The risk intelligent analysis and deduction module is used for risk mechanism tracing and dynamic deduction analysis based on the geological system state: based on the reconstructed spatiotemporal evolution sequence of the geological system state, it performs quantitative decomposition analysis of risk-dominant driving factors, forward numerical simulation prediction of the effects of engineering intervention measures, and identification of comprehensive instability precursor patterns. The mechanism kernel architecture adaptive optimization module is used to construct a knowledge base containing multiple candidate mechanism kernel models and their parameter priors. For new engineering scenarios, it uses initial multi-source heterogeneous monitoring data to search and determine the mechanism kernel combination method and parameter initialization strategy through meta-learning algorithm to accelerate the convergence process of the geological system state inverse modeling. The multidimensional visualization and decision support module is used to dynamically render and display the reconstructed and deduced geological system state through a three-dimensional visualization engine. It supports multi-physics field overlay display, spatiotemporal backtracking, profile analysis, virtual measuring point data extraction, and different scenario comparison functions. It also integrates and displays risk tracing results, intervention scheme simulation comparison reports, and early warning information to generate a comprehensive risk assessment report.

[0009] This invention discloses a computer program instruction that, when executed by a processor, implements the method described in any of the above embodiments.

[0010] In addition, the present invention discloses a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method described in any of the above embodiments.

[0011] The present invention proposes an intelligent identification method and system for geotechnical engineering risks based on geological system state reverse modeling, which has the following technical advantages compared with the prior art: (1) This invention constructs a parameterized, differentiable core of geophysical and chemical process mechanisms and forcibly embeds it as a state equation into an inverse modeling framework based on physical information neural networks and variational assimilation, thereby achieving deep coupling between physical conservation laws and multi-source observation data. This solves the problems of poor physical interpretability and weak extrapolation ability of traditional data-driven methods, enabling the model to output system state predictions that conform to physical laws even in data-sparse regions.

[0012] (2) The geological system state inverse modeling method proposed in this invention can inversely derive the complete, continuous, and dynamically evolving stress field, seepage field, and damage field of the target area using sparse, multi-source, and noisy field observation data as constraints. This breaks through the limitation of existing technologies that can only obtain discrete point states or comprehensive indicators, and for the first time realizes the leap from local observation to full-field state reconstruction, providing an unprecedented panoramic view for a deeper understanding of the risk generation mechanism.

[0013] (3) Based on the reconstructed complete geological system state, this invention can quantitatively analyze the contribution of different physical processes (such as excavation unloading, pore water pressure increase, and damage accumulation) to a specific risk state, and realize the mechanism of risk tracing. At the same time, by substituting virtual engineering measures as new boundary conditions into the calibrated system dynamic model for forward extrapolation, the effects of different prevention and control schemes can be quantified, realizing the upgrade of the decision-making mode from passive alarm to active prediction and scheme optimization, and significantly improving the predictability and scientific nature of risk management. Attached Figure Description

[0014] Figure 1 The flowchart illustrates a method for intelligent identification of geotechnical engineering risks based on geological system state inverse modeling, as provided in this embodiment of the invention.

[0015] Figure 2 This is a framework diagram of a geotechnical engineering risk intelligent identification system based on geological system state reverse modeling, provided by an embodiment of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] like Figure 1 As shown, the present invention provides an intelligent identification method for geotechnical engineering risks based on geological system state inverse modeling, comprising the following steps: S1. Multi-source heterogeneous monitoring data acquisition and fusion preprocessing: Real-time monitoring time-series data of the soil and rock mass and the surface are collected through a multi-type sensor network deployed in the target geotechnical engineering area and engineering structure; wide-area surface deformation time-series data obtained by satellite remote sensing platform, high-precision topographic and geomorphological data obtained by UAV aerial survey platform, and historical geological survey and engineering activity data are integrated simultaneously; coordinate system transformation, time series alignment and data quality verification are performed on the multi-source heterogeneous data to generate a spatiotemporally aligned multi-source heterogeneous monitoring data set; the multi-type sensor network includes at least a stress sensor array for monitoring stress changes in soil and rock mass, a seepage sensor array for monitoring pore water pressure distribution, a microseismic monitoring array for capturing micro-fracture events in rock mass, and a displacement monitoring array for monitoring surface and structural displacement.

[0019] Step S1 aims to construct a unified and reliable spatiotemporal data foundation to address the fundamental problems of data silos and spatiotemporal inaccuracies caused by inconsistent data formats and varying spatiotemporal benchmarks in geotechnical engineering monitoring. This step systematically integrates three types of data sources: high-frequency real-time monitoring data reflecting the internal state are collected through arrays of stress, seepage pressure, microseismic, and displacement sensors deployed within the soil and structural bodies; wide-area surface deformation time-series data acquired by satellite remote sensing platforms and centimeter-resolution topographic data acquired by UAV aerial survey platforms; and historical geological survey and engineering activity data are incorporated to provide prior knowledge and geological background. To achieve deep fusion of multi-source heterogeneous data, this step executes a standardized preprocessing pipeline. First, a coordinate system transformation is performed, converting all spatial data to a preset engineering coordinate system to ensure spatial benchmark consistency. Second, time series alignment is performed, using Coordinated Universal Time (UTC) as the reference, interpolating and synchronizing data from different acquisition frequencies to generate a data sequence with a unified timestamp. Finally, data quality verification is performed, automatically identifying and processing outliers and missing values ​​to ensure data integrity and reliability. The final output is a spatiotemporally aligned, quality-controlled, multi-source heterogeneous monitoring dataset. This step, by constructing an integrated space-ground-air-ground three-dimensional sensing system, achieves complementary and cross-validation of multi-dimensional information at the point, surface, and volume levels. Its core improvement lies in transforming discrete, heterogeneous raw data into a structured data cube that can be directly used by subsequent physical models and intelligent algorithms. This not only automates and standardizes the tedious data processing work, but more importantly, it provides crucial high-quality input for subsequent physical information neural networks and variational assimilation modeling that rely on precise data input, ensuring the accuracy and stability of the entire analytical method from the outset.

[0020] S2. Construction of Differentiable Mechanisms for Geophysical and Chemical Processes: Based on the fundamental principles of rock and soil mechanics, seepage mechanics, and damage mechanics, a set of parameterized and differentiable mathematical and physical models are constructed to characterize the key intrinsic processes of geological risk development. The differentiable mechanism cores include at least: a hydraulic coupling mechanism core describing the coupling of unsaturated seepage and stress field; a stress-strain mechanism core characterizing the elastoplastic deformation and structural surface effects of rock and soil; and a damage evolution mechanism core describing the cumulative evolution of damage within the rock mass. Each mechanism core is expressed in the form of partial differential equations or differential-algebraic equations, and its model parameters are optimized and adjusted through a data-driven approach.

[0021] Step S2 involves constructing differentiable mechanism cores for geophysicochemical processes. This involves expressing the key intrinsic physicochemical processes controlling the development of geological risks as a set of structured, parameterized, and mathematically differentiable computational models. These models serve as the physical constraints for subsequent inverse modeling, constituting the differentiable mechanism cores. Specifically, based on the fundamental theories of rock and soil mechanics, seepage mechanics, and damage mechanics, three core mechanism cores are constructed. The first is the hydraulic coupling mechanism core, which, based on unsaturated seepage theory and solid mechanics equilibrium equations, constructs a set of governing equations describing the bidirectional coupling of pore water pressure and stress fields. The second is the stress-strain mechanism core, which, based on an elastoplastic mechanics framework, introduces equivalent continuous media or discrete interface element methods to describe the nonlinear deformation behavior of rock and soil masses and the contact friction effects of structural surfaces such as faults and joints. The third is the damage evolution mechanism core, which characterizes the material property degradation process caused by the accumulation of micro-fractures within the rock mass by defining damage internal variables and their evolution equations. Each mechanism core is precisely expressed in mathematical form using partial differential equations or differential-algebraic equations. The core of this step lies in the design of the differentiability of the mechanism kernels. The gradient of the output of the governing equations corresponding to each mechanism kernel with respect to the input variables and model parameters must be calculated accurately and efficiently. This property is achieved through automatic differentiation technology, allowing these equations representing physical conservation laws to be directly embedded as mandatory constraints into the physical information neural network framework optimized by the gradient descent algorithm. Furthermore, the model parameters contained in each mechanism kernel, such as permeability coefficients, mechanical property parameters, and damage evolution parameters, are all designed as adjustable variables. These parameters are not fixed in advance but are automatically inverted and optimized in subsequent modeling steps using multi-source observation data as the training target through a data-driven approach. Therefore, the differentiable mechanism kernel system constructed in this step transforms prior physical knowledge into a computational module that can be deeply integrated with observational data and drive adaptive model calibration, laying the sole physical-mathematical model foundation for high-fidelity inversion of the complete geological system state from sparse observations.

[0022] S3. Inverse Modeling of Geological System State Based on Physical Information and Data Assimilation: Define the geological system state of the target area as a high-dimensional hidden state vector, which represents the stress field, strain field, pore water pressure field, and damage field of the soil and rock mass in the complete spatiotemporal domain; use the differentiable mechanism kernel constructed in step S2 as the state equation describing the evolution of the system state; establish the observation equation, and describe the multi-source heterogeneous monitoring data obtained in step S1 as the mapping of the geological system state under a specific observation operator; adopt a solution framework combining physical information neural network and variational assimilation, and use the multi-source heterogeneous monitoring data as constraints, and use the optimization algorithm to solve the spatiotemporal evolution sequence of the geological system state that simultaneously satisfies the state equation and the observation equation, as well as the corresponding model parameter field, to obtain a digital geological system dynamic model calibrated by multi-source data.

[0023] The differentiability characteristic is a key prerequisite for the physical information neural network to embed the mechanism kernel equation as a constraint into the loss function and perform gradient optimization. The physical information neural network used in step S3 is an advanced modeling framework that deeply integrates known physical laws with data-driven learning. Its core principle lies in directly embedding the mathematical physical equations describing the intrinsic evolution of the geological system, i.e., the partial differential equations corresponding to the differentiable mechanism kernel constructed in step S2, as strong constraints into the training process of the neural network. The loss function of this network consists of two parts: one part is the data fitting term, which measures the difference between the network's predicted value and the actual observation data obtained in step S1; the other part is the physical law residual term, which calculates the degree to which the network's predicted output field satisfies the differentiable mechanism kernel control equation at the corresponding points by sampling a large number of configuration points in the solution domain. The key to the realization of this network lies in the differentiability characteristic of the mechanism kernel constructed in step S2. With the help of automatic differentiation technology, the gradient of the physical control equation with respect to the network input variables (spatiotemporal coordinates) and the adjustable parameters of the network can be accurately calculated. This allows the residuals of the physical equations to participate in and guide the gradient backpropagation and optimization update of the neural network weight parameters. During training, the optimization algorithm not only strives to narrow the gap between the network prediction and the measured data, but also forces the entire physical field distribution predicted by the network to tend to satisfy the governing equations, which serve as prior knowledge. Through this mechanism, the mapping relationship from input (spatial-temporal coordinates) to output (multi-physics state) learned by the physical information neural network is strictly constrained within the solution space determined by physical laws. This invention, through the physical information neural network, does not simply merge the physical model calculation results with the data-driven model prediction results through post-processing, but achieves the native unity of physical mechanisms and observational evidence in the underlying optimization objective of model training. This ensures that the final inverse solution obtained is a high-fidelity solution that best matches the multi-source observation data in the least squares sense, and whose predicted state still automatically follows the basic physical laws of rock and soil mechanics, seepage mechanics, etc., at any spatiotemporal location where no measuring points are set up.

[0024] S4. Risk mechanism tracing and dynamic inference analysis based on geological system state: Based on the spatiotemporal evolution sequence of the geological system state reconstructed in step S3, quantitative decomposition analysis of risk-dominant driving factors, forward numerical simulation prediction of the effects of engineering intervention measures, and identification of comprehensive instability precursor modes are carried out.

[0025] Step S4, based on the spatiotemporal evolution sequence of the geological system state calibrated by both physical laws and observational data, quantitatively calculates the contribution of each physical process to the risk state by analyzing the spatiotemporal coupling relationship and energy transfer path between the stress field, seepage field, and damage field in the sequence, thus achieving the mechanism tracing of risk causes; by parameterizing virtual engineering measures as new boundary conditions or internal parameters of the system dynamic model and performing forward numerical integration, the evolution process of the system state and changes in risk indicators after intervention are simulated and evaluated; by using nonlinear time series analysis methods to extract dynamic characteristic quantities characterizing the system instability critical point in the high-dimensional system state phase space, and constructing a machine learning recognizer, comprehensive intelligent identification of instability precursors based on the overall behavior of the system is achieved.

[0026] S5. Meta-learning-driven adaptive optimization of mechanism kernel architecture: Construct a knowledge base containing multiple candidate mechanism kernel models and their parameter priors; For new engineering scenarios, utilize initial multi-source heterogeneous monitoring data, and use meta-learning algorithms to search and determine the mechanism kernel combination method and parameter initialization strategy to accelerate the convergence process of the geological system state inverse modeling in step S3.

[0027] Step S5 constructs a knowledge base containing various candidate mechanism kernels (such as different elastoplastic models and damage evolution laws) and their corresponding parameter prior distributions. For new engineering scenarios, after acquiring a small amount of initial monitoring data, a model-independent meta-learning algorithm is used for adaptive optimization. The goal of this algorithm is to learn a set of meta-parameters that allow for rapid convergence with minimal gradient updates on new tasks. Specifically, a combination of candidate mechanism kernels is sampled from the knowledge base and used as the base learner. Using initial multi-source heterogeneous monitoring data, the loss of the base learner on the support set is calculated, and one or more gradient descent steps are performed through the inner loop to obtain task-specific model parameters. Then, the loss is calculated based on the query set, and the meta-parameters are updated through backpropagation in the outer loop. By iterating this process, the meta-learner learns how to quickly adjust the mechanism kernel combination and parameters based on the initial data, thus providing a high-quality initial solution for subsequent complete inverse modeling.

[0028] Step S5 involves meta-learning-driven adaptive optimization of the mechanism kernel architecture. Its principle lies in introducing machine learning to address the slow startup and difficult parameter tuning issues of physical information models when facing entirely new engineering scenarios. This step pre-builds a knowledge base containing multiple candidate mechanism kernel models and their prior parameter distributions. When applied to a new engineering scenario, the system utilizes the limited monitoring data acquired initially to drive the meta-learning algorithm to automatically search and determine the optimal mechanism kernel combination architecture and parameter initialization strategy within the knowledge base. Essentially, this process rapidly transfers model configuration experience learned from multiple historical scenarios to the new scenario, thus providing a highly adapted initial starting point for the core reverse modeling process in Step S3, significantly accelerating its optimization convergence. Traditional methods retrain general models in new scenarios, resulting in poor performance and slow convergence in the early stages when data is scarce. This step uses meta-learning to automate architecture search and initialization. In the early stages of a project with limited data, it automatically recommends the most suitable combination of physical models for the current geological and working conditions, reducing the model initialization and calibration cycle from months to days. It also reduces the heavy reliance on the experience of domain experts, enabling the system to autonomously adapt to different engineering types, from soft soil tunnels to hard rock slopes, thus improving the engineering practicality and universality of the method.

[0029] S6. Multidimensional visualization and decision support of geological system states: The geological system states reconstructed in step S3 and deduced in step S4 are dynamically rendered and displayed through a 3D visualization engine. It supports multi-physics field overlay display, spatiotemporal backtracking, profile analysis, virtual measuring point data extraction and comparison of different scenarios. It also integrates and displays risk tracing results, intervention scheme simulation comparison reports and early warning information, and generates a comprehensive risk assessment report.

[0030] Step S6, centered on a 3D visualization engine, dynamically renders and expresses the geological system state (including a spatiotemporal continuous data field of multiple physical fields such as stress, strain, seepage, and damage) obtained from the reverse modeling in Step S3 and the predicted geological system state in Step S4 through simulation, using scientific visual encoding schemes (such as color mapping, isosurfaces, vector fields, and transparency). Furthermore, as an information aggregation and decision-making terminal, it integrates and comprehensively presents the results of multi-source analysis, such as risk tracing, measure simulation, and early warning, and drives the automatic generation of structured reports.

[0031] Specifically, the construction of the hydraulic coupling mechanism core in step S2 includes: S211. Based on the theory of unsaturated seepage, a differentiable form of a three-dimensional equation with pore water pressure as the dependent variable is established, wherein the soil hydraulic property function is described by a parameterized model, and the gradient of the model is calculated by automatic differentiation technology. The seepage component of the hydraulic coupling mechanism core is based on unsaturated seepage theory, and its governing equation is a three-dimensional equation with pore water pressure head h as the dependent variable. The hybrid form of this equation is as follows: (1) Where: C(h)=∂θ / ∂h is the specific water density, representing the change in water content caused by a change in pressure head; θ is the volumetric water content of the soil; K(h) is the unsaturated hydraulic conductivity tensor, which is a function of pressure head h; z is the vertical coordinate (upward is defined as positive); ∇ is the spatial gradient operator, ∇· is the divergence operator; t is time.

[0032] This invention employs a parameterized model, wherein the mathematical expression for the effective saturation Θ is: (2) Relative hydraulic conductivity Kr(Θ) and unsaturated hydraulic conductivity K(h): (3)

[0033] Where θs and θr are the saturated water content and residual water content, respectively. α, n, and m are shape parameters. l is the pore connectivity parameter; Ks is the saturated hydraulic conductivity tensor.

[0034] The key shape parameters α, n, and saturated conductivity Ks are functions of the spatial position x. α=α(x), n=n(x), Ks=Ks(x).

[0035] This characterizes the spatial heterogeneity of the geological body's permeability and pore structure. In the inverse modeling framework, α(x), n(x), and Ks(x) will become the spatially distributed parameter field that needs to be inverted.

[0036] Simultaneously, a correction term coupled with the stress field and damage field is introduced to associate the saturated water content θs with the volumetric strain εv or the damage variable D: (4) θ{s0} represents the initial saturated water content, and βε and βD are coupling coefficients. Compression deformation (negative εv) reduces porosity, thereby lowering the saturated water content; damage development (increased D) generates new voids or penetrating cracks, thereby increasing the saturated water content. This achieves bidirectional coupling of the seepage field with the stress-strain field and the damage field, surpassing the limitations of traditional models that only describe static medium properties.

[0037] S212. The permeation volume force obtained from the pore water pressure field calculation is used as the source term and coupled to the equilibrium equation of the stress-strain mechanism core to construct a two-way coupled calculation framework for fluid pressure and solid skeleton deformation.

[0038] The relationship between pore water pressure p and pressure head is as follows: (5) Where ρw is the density of water and g is the acceleration due to gravity. The resulting osmotic volume force f is: (6) This volume force f, as an additional body force source term, is introduced into the momentum balance equation of the stress-strain mechanism core: (7) Where σ' is the effective stress tensor; ρ is the total density of the soil and rock mass.

[0039] S213. Mathematically describe the rainfall infiltration boundary, groundwater level fluctuation boundary, and engineering structure contact boundary, and encode them as boundary condition constraints embedded in the variational assimilation solution framework.

[0040] To achieve modeling that matches actual engineering conditions, complex boundaries must be accurately described mathematically and encoded as constraints within a variational assimilation solution framework. This mainly involves three types of boundaries.

[0041] Rainfall infiltration boundary: -K(h)∇(h+z)·n=qrain(t) (8) Where qrain(t) is the rainfall intensity (infiltration flux) that varies with time, and n is the boundary outward normal vector.

[0042] Groundwater level or known head boundary: h(x,t)=hspecified(x,t) (9) Here, hspecified is the pressure head value known at the specified location and time at the boundary.

[0043] Engineering structural contact boundary: The contact surface between tunnel lining, retaining wall and rock and soil can be treated as an impermeable boundary (-K(h)∇(h+z)·n=0) or a continuous flow boundary considering equivalent permeability.

[0044] In solving physical information neural networks, the aforementioned boundary conditions are implemented by sampling a large number of boundary configuration points on the corresponding boundaries of the computational domain. At these points, the difference between the network's predicted values ​​and the boundary condition-specified values ​​is added as an independent loss term to the total loss function. Automatic differentiation guides network training, thereby forcing the solution to satisfy the boundary conditions.

[0045] The hydraulic coupling mechanism core constructed in this step represents a leap from a homogeneous parameter model to a spatialized intelligent coupling model. Traditional methods treat soil hydraulic parameters as constants independent of spatial location and mechanical state, severely deviating from the heterogeneous and multi-field coupled nature of natural soil and rock masses. This invention defines key hydraulic parameters as functions of spatial coordinates and uses them as a distributed parameter field to be inverted, enabling the direct reconstruction of the true spatial variation of geological permeability and pore structure from multi-source heterogeneous monitoring data. Simultaneously, by innovatively introducing a coupling correction term for volumetric strain and damage variables into the saturated water content, an intrinsic mathematical correlation is established between seepage characteristics, stress state, and damage evolution, achieving deep bidirectional coupling of the seepage field with other physical fields at the constitutive relation level. This not only allows the model to more realistically depict complex geological conditions but also makes the inverted parameter field itself an important indicator for revealing hydrogeological structural anomalies and mechanical damage areas, thereby significantly improving the physical fidelity of the entire geological system's state-of-the-art inverse modeling and the depth of explanation for risk formation mechanisms.

[0046] Specifically, the construction of the stress-strain mechanism core in step S2 includes: S221. Based on the framework of elastoplastic mechanics, a differentiable expression for the equilibrium equations and geometric equations of soil and rock is established. The equilibrium equations of soil and rock state that the sum of the divergence of the stress tensor and the volume forces is zero. This equation enforces the static equilibrium condition at any point inside the object. The geometric equations, i.e., the strain tensor, are given by the symmetric part of the displacement gradient, and describe the geometric relationship between deformation and motion. To achieve the differentiability of the constitutive relation, a smoothed elastoplastic theory is adopted. By introducing a continuously differentiable smooth function to approximate the traditional yield surface and plastic potential surface, the derivative of stress with respect to strain, i.e., the constitutive tangential stiffness, exists continuously throughout the entire process, including the elastic state, yield point, and plastic flow. This eliminates the problem of derivative discontinuity at the yield point in the traditional model, providing the necessary conditions for gradient-based optimization algorithms.

[0047] S222. The equivalent continuous medium method or discrete interface element method is used to explicitly model the key structural surfaces of faults and joints in the rock mass. The normal and tangential mechanical behaviors of the structural surfaces are described by a differentiable nonlinear contact-friction model.

[0048] For discontinuities such as faults and joints in rock masses, a discrete interface element method is used for explicit modeling. The normal contact behavior is described by a continuously differentiable pressure embedding function, avoiding discrete jumps in the contact state. The tangential friction behavior is characterized by a smoothed Coulomb friction law, using a hyperbolic tangent function as the transition function. This ensures that the frictional force transitions continuously and smoothly from zero to maximum static friction near the zero-slip point, thus eliminating the singularity of the derivative at the slip initiation point. The friction coefficient can also be defined as a differentiable function of slip or slip velocity to simulate complex frictional degradation effects.

[0049] S223. The engineering excavation activity is simulated as a dynamic removal of material properties and release of equivalent loads in a specific area within the computational domain. The anchor bolt and anchor cable support structure is simulated as a one-dimensional rod element or prestressed cable element, and its interaction with the surrounding rock is described by a differentiable bond-slip constitutive model. The excavation and support processes are parameterized as functions of time and integrated into the mechanism core.

[0050] The excavation activity is simulated as a smooth decay of the material properties of the excavated area over time, with parameters such as elastic modulus and strength gradually decreasing from their original values ​​to zero via a smooth time function. Simultaneously, a correspondingly smooth-changing equivalent release load is applied at the excavation boundary to simulate the gradual unloading of in-situ stress. For support structures such as anchor bolts and cables, they are discretized into one-dimensional elements, and their material constitutive model employs a differentiable model. Their bond-slip relationship with the surrounding rock is also described by a continuously differentiable curve, which smoothly characterizes the entire process from linear bonding, peak strength, to the softening stage. The support elements are coupled to the surrounding rock displacement field through their nodes and contribute to the overall system equilibrium in a differentiable manner.

[0051] This stress-strain mechanism core, through the aforementioned comprehensive differentiable design, integrates the nonlinearity, contact discontinuity, and construction dynamics of soil and rock mass into a unified and smooth mathematical framework. This allows it to be seamlessly embedded into the training process of a physical information neural network, thereby supporting efficient, joint, and global optimization inversion of rock mass mechanical parameters, geostress field, structural surface properties, and construction effects.

[0052] Specifically, step S3 employs a solution framework combining physical information neural networks and variational assimilation, including: S31. Construct a physical information neural network with an encoder-decoder structure. The encoder is used to extract the temporal features of multi-source heterogeneous monitoring data, and the decoder takes spatiotemporal coordinates as input and outputs the corresponding multi-physics state prediction values.

[0053] Specifically, the encoder receives spatiotemporally aligned multi-source monitoring time-series data generated in step S1 as input. Internally, it consists of recurrent neural network layers or temporal convolutional layers, specifically designed to extract and fuse the implicit temporal evolution features and patterns within this data, outputting a feature vector representing the overall state of the current system. The decoder takes the coordinates and time of any point in the target spatiotemporal domain as input, including the three-dimensional spatial location and a specific moment. Essentially, the decoder is a functional approximator with spatiotemporal coordinates as independent variables and multiphysics states as dependent variables. Through learning, it can directly output the predicted value of the complete geological system state defined in step S2 corresponding to that coordinate point, including stress tensor, pore water pressure scalar, and damage variables.

[0054] S32. The residual of the partial differential equation corresponding to the differentiable mechanism kernel constructed in step S2 is incorporated into the loss function of the neural network as a mandatory physical constraint term, and configuration points are sampled in the entire spatiotemporal computation domain to force the network prediction to satisfy the physical constraint.

[0055] To achieve this goal, a large number of placement points are randomly or systematically sampled throughout the entire spatiotemporal computational domain to be solved. For each placement point, the multiphysics predictions output by the decoder are substituted into the governing partial differential equations corresponding to the differentiable mechanism kernels constructed in step S2. Specifically, the predicted pore water pressure field is substituted into the Richards equations, and the predicted stress-strain field is substituted into the equilibrium equations and geometric equations. The residuals of these equations at that point are calculated, i.e., the difference between the left and right sides of the equations. The loss function for training the neural network not only includes the difference between the decoder output at the sensor location and the measured data, but also incorporates the sum of squares and the mean of the residuals of the physical equations at all these placement points; simultaneously, the gradient of the loss function with respect to all weight parameters of the network is calculated using automatic differentiation techniques. The gradient descent optimization algorithm, while reducing data fitting errors, forces the network to adjust its internal parameters so that its output at all placement points satisfies the pre-set physical conservation laws as much as possible, thereby ensuring that the mapping relationship learned by the network has a solid physical basis.

[0056] S33. Using a four-dimensional variational assimilation method, the initial system state, boundary condition parameters, and spatially distributed material parameters are combined to form a set of variables to be optimized. By minimizing the difference between the network prediction value and the multi-source heterogeneous monitoring data within a long time window, while simultaneously satisfying physical constraints, the optimal geological system state evolution sequence and parameter field are obtained by joint inversion using the gradient descent algorithm.

[0057] This step places the aforementioned physical information neural network framework within a dynamic data assimilation optimization problem. The set of variables to be optimized includes not only the weight parameters of the neural network but also extends to the initial state of the entire system, boundary condition parameters, and parameter fields characterizing the spatial distribution of materials. The optimization objective is defined as minimizing the overall difference between the time-series data predicted by the neural network model and the actual multi-source monitoring time-series data over a relatively long time window, while strictly satisfying the aforementioned physical constraints. By solving this constrained optimization problem, gradient descent algorithms can be used to simultaneously and jointly derive the optimal spatiotemporal evolution sequence of the geological system state, the optimal spatial distribution of material parameters, and the corresponding initial and boundary conditions. Ultimately, the network decoder trained and converged through this process constitutes a digital dynamic model of the geological system calibrated with multi-source data and strictly adhering to physical laws, capable of reflecting the true mechanical, seepage, and damage evolution processes within the rock and soil mass with high fidelity.

[0058] Specifically, step S4 includes: S41. Quantitative Analysis of Risk Mechanism: Based on the spatiotemporal evolution sequence of the geological system reconstructed in step S3, analyze the spatiotemporal coupling relationship and energy transfer path between the stress field, seepage field and damage field, calculate the contribution of stress redistribution, pore water pressure change and damage accumulation to the state evolution of the currently identified risk anomaly areas (such as deformation acceleration zone and plastic zone), and output the dominant driving factor analysis report in the form of quantitative percentage.

[0059] This step, by analyzing a complete spatiotemporal sequence of the geological system calibrated with physical laws, clarifies the interactions and energy transfer relationships between the stress field, seepage field, and damage field, thereby calculating the quantitative contribution ratio of different physical processes to the state evolution of specific risk anomaly areas. This fundamentally changes the traditional approach's limitation of only being able to locate risks without explaining their causes, elevating the understanding of engineering risks from superficial phenomenological descriptions to intrinsic mechanism diagnosis, and providing direct quantitative scientific evidence for formulating precise and efficient engineering management strategies.

[0060] S42. Dynamic simulation of the effects of engineering intervention measures: Receive virtual engineering intervention measure parameters defined by the user for the dynamic model of the digital geological system, and transform the intervention measures into new boundary conditions or internal parameter modifications of the model; using the current system state obtained by assimilation in step S3 as the initial condition, run the model coupled with the differentiable mechanism kernel for forward numerical integration, simulate and visualize the evolution process of the key physical fields of the geological system state after the implementation of the intervention measures and the corresponding changes in risk indicators within a specified future time period, and quantitatively evaluate the effects of different intervention schemes.

[0061] Step S42 constructs a virtual testing and pre-evaluation platform for engineering measures based on a high-fidelity digital twin model. This function allows users to define virtual engineering parameters, directly converting different intervention schemes into boundary conditions or internal parameter modifications of a calibrated system dynamic model, and performing high-fidelity forward numerical simulations starting from the current real system state. This enables engineering decision-makers to intuitively observe and quantitatively evaluate the future dynamic impact of different design schemes on key physical fields and comprehensive risk indicators before construction implementation. It achieves a fundamental shift in engineering decision-making from passively responding to sudden risks to proactively designing and optimizing intervention schemes, greatly improving the scientific rigor, economy, and reliability of engineering measures.

[0062] S43. Intelligent identification of instability precursors based on system state phase space: Based on the historically assimilated geological system state time series, in the high-dimensional system state phase space composed of the stress field, seepage field, and damage field variables, nonlinear time series analysis is used to extract feature quantities characterizing the critical point of system dynamic instability; a machine learning model is used to learn the mapping relationship between the feature quantities and macroscopic instability events to construct a data-driven comprehensive instability precursor identifier; the real-time short-sequence system state obtained by online assimilation is input into the identifier, and a continuous instability proximity probability index is output.

[0063] This step represents a leap from relying on single monitoring indicator thresholds to intelligent judgment based on the overall dynamic behavior of complex systems. It involves analyzing the evolutionary trajectories of historical instability events in a high-dimensional system state-phase space, extracting universal critical dynamic characteristics of system instability, and constructing an identification model using machine learning. This method can keenly capture early instability precursor signals hidden in multi-physics coupled systems, preceding macroscopic damage, and thus outputs a continuous, physically meaningful instability proximity probability index. This effectively overcomes the dual shortcomings of high lag and high false alarm rate in traditional single-parameter alarm methods, providing earlier and more reliable advance warnings and reserving valuable time for engineering emergency response.

[0064] Specifically, the simulation of the effects of engineering intervention measures in step S42 includes: S421. Provides a graphical user interface to receive user definitions of the geometric parameters, material properties, and implementation sequence of drainage holes, anchoring components, or excavation support procedures.

[0065] Step S421 involves providing a graphical user interface to receive user-defined intervention parameters. The system, through an integrated 3D visualization engine, loads and presents the dynamic model of the digital geological system calibrated in step S3 and its associated engineering scenarios. Within this visualization environment, users can directly manipulate the system using a series of graphical tools. For drainage holes, users can directly select or draw on the model surface or within it to determine the borehole trajectory, and define specific geometric and material properties such as borehole diameter, length, filter pipe range, and opening ratio through the side parameter panel, while also setting the opening and closing times and design drainage flow rate. For anchoring components, users can select the construction location and direction in the model, and fill in attributes such as the length, diameter, prestress value, and grouting material parameters of the anchor bolt or cable, as well as specify the tensioning and locking sequence. For complex excavation and support procedures, users can define the start and end times and depth of excavation for different zones, as well as the timing and parameters of associated support structures such as shotcrete layers or steel arches, through the timeline editor. The system records and stores all user-defined parameters in a structured manner, forming a complete virtual engineering intervention plan that can be called upon by subsequent calculation modules.

[0066] S422. The user-defined intervention parameters are converted into corresponding boundary condition modifications, material property updates, or load condition changes in the dynamic model of the digital geological system.

[0067] The core task of step S422 is to automatically convert user-defined engineering semantic parameters into mathematical instructions that the underlying digital geological system dynamic model can recognize and execute. This is a crucial parameter mapping and model update process. For example, user-defined drainage holes are automatically converted by the system into a set of constant head boundary conditions or given flow boundary conditions for the seepage field control equations at the corresponding spatial locations. User-added anchors or cables are converted into a series of one-dimensional rod or cable elements embedded in the soil matrix, their material properties are updated to the mechanical parameters of steel, and integrated into the overall system through stiffness matrix superposition. Simultaneously, the applied prestress is converted into corresponding nodal loads. For user-programmed excavation procedures, the system processes them by performing dynamic passivation simulation of the material properties of the target area element group at a specified time step. That is, its elastic modulus, strength, and other parameters are smoothly degraded to zero according to a time function, and the equivalent nodal forces released by the removal of this part of the soil are calculated and applied simultaneously to accurately simulate the stress redistribution process.

[0068] S423. Using the currently calibrated system state as the initial condition, perform forward numerical integration with the modified model to simulate the evolution process of the geological system state within a specified future time period, and dynamically output the predicted results of displacement field, stress field, plastic zone and risk level indicators.

[0069] Step S423 is responsible for performing a high-fidelity forward numerical simulation starting from the current state. This step uses the complete geological system state, representing the current real state, obtained from the assimilation in step S3, as the initial condition for the simulation. The system calls a numerical solver that integrates all differentiable mechanism kernels to perform forward time integration on the coupled system modified in step S422. The solver, based on the forward mode of the finite element method, finite difference method, or physical information neural network, progressively calculates the dynamic evolution of the entire model under the influence of intervention measures within a specified prediction period in the future. The simulation calculation proceeds according to a set time step and dynamically outputs the spatiotemporal distribution results of key physical fields across the entire domain. These results include, but are not limited to: three-dimensional displacement field cloud maps and vector maps of the surface and deep subsurface, maximum and minimum principal stress and shear stress field distributions, the range of plastic zones indicating the yield or failure state of materials, and risk level zoning maps calculated and drawn in real time based on criteria such as displacement rate and safety factor.

[0070] S424. Compare and analyze the simulation results under different intervention programs, and quantitatively evaluate the impact of each program on risk indicators.

[0071] Step S424 aims to systematically compare and quantitatively evaluate the effects of different intervention schemes to support scientific decision-making. Users can define and submit multiple different virtual intervention schemes. The system will perform the simulation and deduction of step S423 sequentially or in parallel for each scheme and generate a dataset of prediction results containing a complete spatiotemporal sequence for each scheme. Subsequently, the built-in analysis engine will automatically extract and compare the key performance indicators of each scheme at the same prediction time point and the same key focus area. The quantitative indicators for comparison include: the maximum displacement value and percentage reduction of displacement rate at key monitoring points, the reduction ratio of the total volume of the plastic zone, the rate of change of the area of ​​high-risk areas, the distribution and maximum value of the internal forces of the support structure, and the safety factor evolution curve reflecting the overall stability over time. The system will finally generate a comprehensive analysis report containing comparison curves of multiple schemes, bar charts, effect difference cloud maps, and data tables, intuitively showing the quantitative advantages and disadvantages of different engineering intervention strategies in controlling deformation, inhibiting damage development, and reducing risk levels, thereby providing clear and direct data support for selecting the optimal risk management scheme.

[0072] Specifically, the comprehensive instability precursor pattern identification in step S43 includes: S431. Collect the geological system state time series data reconstructed before the occurrence of instability events in historical cases.

[0073] Step S431 is responsible for constructing a historical dataset for training. This involves collecting and organizing complete geological system state time-series data reconstructed using the reverse modeling framework described in step S3, prior to instability events such as landslides and cavern collapses in historical engineering cases. This data is not raw, sparse sensor readings, but rather spatiotemporally continuous sequences calibrated according to physical laws, containing complete information on stress fields, seepage fields, and damage fields. For each historical case, a sufficiently long system state evolution sequence reflecting the instability gestation process before instability is extracted as a positive sample. Simultaneously, a large number of system state sequences under stable conditions or undergoing only conventional deformation are collected as negative samples, together forming a time-labeled sample library.

[0074] S432. In the high-dimensional phase space of the geological system, calculate the index characterizing the dynamic characteristics of the system. The index includes the local Lyapunov exponent, the system restitution force, the order parameter extracted by principal component analysis, and the nonlinear coherence coefficient between multiple physics fields.

[0075] Step S432 aims to extract quantitative features characterizing dynamic stability from the high-dimensional system state phase space. The system state time series of each historical case is mapped to a high-dimensional phase space spanned by all field variables, with the trajectory representing the evolution path of the system state. Within this space, a series of key indicators sensitively reflecting the impending abrupt changes in the system's nonlinear dynamic behavior are calculated. These indicators include: local Lyapunov exponents, used to quantify the divergence rate of adjacent trajectories in the phase space, indicating an increased sensitivity of the system to initial conditions; system resilience parameters, assessed by analyzing the rate at which the system returns to equilibrium after perturbation or the morphological changes of attractor basins; order parameters extracted from principal component analysis, i.e., eigenvectors capturing the main evolutionary patterns of the system, whose abrupt changes often predict the formation of macroscopically ordered structures; and nonlinear coherence coefficients between multiphysics fields, used to measure the intensity and directionality changes of interactions between different fields such as stress, seepage, and damage.

[0076] S433. Using the aforementioned dynamic characteristic indicators as features and whether instability is imminent as a label, a machine learning classification model is trained, and the classification model constitutes a data-driven instability precursor identifier.

[0077] The core of step S433 is to construct a data-driven instability precursor identifier using machine learning. The various dynamic characteristic indicators calculated in the previous step are used as input feature vectors, with the determination of whether instability ultimately develops within a given time period serving as the binary classification label. Machine learning algorithms such as Support Vector Machines are employed for supervised learning training on a prepared historical dataset. The learning objective of this model is to discover common, generalizable abnormal patterns exhibited by the aforementioned high-dimensional dynamic features during the critical transition from a stable to an unstable state. The trained classification model then becomes an intelligent instability precursor identifier, capable of comprehensively evaluating multi-dimensional dynamic information rather than relying on a single threshold.

[0078] In the specific implementation of step S433, the dynamic indicator feature vectors extracted from historical cases and their corresponding stable or near-instability labels are first used to form a training dataset. Then, a support vector machine algorithm is used to train this dataset. The training process involves solving the aforementioned convex optimization problem using an optimization algorithm to determine the optimal hyperplane parameters or their equivalent representation under kernel function mapping. The trained model constitutes the core discriminant function of the instability precursor identifier. For a new short sequence of geological system states from a real-time assimilation system, its dynamic feature vector is first calculated according to step S432, and then input into the discriminant function. The function outputs a real value, the sign of which represents the classification result, and its absolute value reflects the distance of the sample from the classification decision boundary. By mapping this output value to a probability exponent, the continuous instability proximity probability assessment required in step S434 can be achieved. The advantage of this method is that it is based on the principle of minimizing structural risk, comprehensively considers all input feature dimensions, and uses kernel functions to capture complex nonlinear patterns, thereby achieving robust and generalizable intelligent identification of instability precursors.

[0079] S434. Input the real-time system state short sequence obtained by online assimilation into the instability precursor identifier, calculate and output the instability proximity probability index, and trigger an early warning when the index exceeds a preset threshold.

[0080] In actual operation, the system continuously performs online assimilation in step S3, constantly generating the latest real-time geological system state sequence within a short time window. This short sequence is then input in real-time into the instability precursor identifier trained in step S433. Based on the dynamic characteristics calculated from the real-time data, the identifier outputs a continuous value between zero and one, namely the instability proximity probability index. The higher the index, the greater the similarity between the current system state and historical pre-instability patterns. The system continuously monitors this index and sets different levels of warning thresholds. When the index exceeds the preset yellow warning threshold, the system issues a risk concern alert; when the index further rises and exceeds the red warning threshold, the system automatically triggers a high-level warning, prompting engineering personnel to immediately intervene for inspection and handling, thereby achieving advanced and intelligent early warning of potential instability disasters.

[0081] Correspondingly, a geotechnical engineering risk intelligent identification system based on geological system state inverse modeling includes: The multi-source data fusion preprocessing module is used to collect real-time monitoring time-series data of the soil and rock mass and the surface through a multi-type sensor network deployed in the target geotechnical engineering area and engineering structure; synchronously integrate wide-area surface deformation time-series data acquired by satellite remote sensing platform, high-precision topographic and geomorphological data acquired by UAV aerial survey platform, and historical geological survey and engineering activity data; perform coordinate system transformation, time series alignment and data quality verification on the multi-source heterogeneous data to generate a spatiotemporally aligned multi-source heterogeneous monitoring data set; the multi-type sensor network includes at least a stress sensor array for monitoring stress changes in soil and rock mass, a seepage pressure sensor array for monitoring pore water pressure distribution, a microseismic monitoring array for capturing micro-fracture events in rock mass, and a displacement monitoring array for monitoring surface and structural displacement.

[0082] The differentiable mechanism core construction module, based on the fundamental principles of rock and soil mechanics, seepage mechanics, and damage mechanics, constructs a set of parameterized and differentiable mathematical and physical models to characterize the key intrinsic processes of geological risk development. The differentiable mechanism core includes at least: a hydraulic coupling mechanism core describing the coupling of unsaturated seepage and stress field, a stress-strain mechanism core characterizing the elastoplastic deformation and structural surface effects of rock and soil, and a damage evolution mechanism core describing the cumulative evolution of damage within the rock mass. Each mechanism core is expressed in the form of partial differential equations or differential-algebraic equations, and its model parameters are optimized and adjusted through a data-driven approach.

[0083] The geological system state inverse modeling module defines the geological system state of the target area as a high-dimensional hidden state vector, which represents the stress field, strain field, pore water pressure field, and damage field of the soil and rock mass in the complete spatiotemporal domain. The constructed differentiable mechanism kernel is used as the state equation describing the evolution of the system state. An observation equation is established, which describes the obtained multi-source heterogeneous monitoring data as a mapping of the geological system state under a specific observation operator. A solution framework combining physical information neural networks and variational assimilation is adopted, with the multi-source heterogeneous monitoring data as constraints, and an optimization algorithm is used to solve the spatiotemporal evolution sequence of the geological system state that simultaneously satisfies the state equation and the observation equation, as well as the corresponding model parameter field, to obtain a digital geological system dynamic model calibrated by multi-source data.

[0084] The risk intelligent analysis and deduction module is used for risk mechanism tracing and dynamic deduction analysis based on the geological system state: based on the reconstructed spatiotemporal evolution sequence of the geological system state, it performs quantitative decomposition analysis of risk-dominant driving factors, forward numerical simulation prediction of the effects of engineering intervention measures, and identification of comprehensive instability precursor patterns.

[0085] The mechanism kernel architecture adaptive optimization module is used to construct a knowledge base containing multiple candidate mechanism kernel models and their parameter priors. For new engineering scenarios, it uses initial multi-source heterogeneous monitoring data to search and determine the mechanism kernel combination method and parameter initialization strategy through meta-learning algorithm to accelerate the convergence process of the geological system state inverse modeling.

[0086] The multidimensional visualization and decision support module is used to dynamically render and display the reconstructed and deduced geological system state through a three-dimensional visualization engine. It supports multi-physics field overlay display, spatiotemporal backtracking, profile analysis, virtual measuring point data extraction, and different scenario comparison functions. It also integrates and displays risk tracing results, intervention scheme simulation comparison reports, and early warning information to generate a comprehensive risk assessment report.

[0087] This invention discloses a computer program instruction that, when executed by a processor, implements the method described in any of the above embodiments.

[0088] In addition, the present invention discloses a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method described in any of the above embodiments.

[0089] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0090] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for intelligent identification of geotechnical engineering risks based on geological system state inverse modeling, characterized in that, Includes the following steps: S1. Multi-source heterogeneous monitoring data acquisition and fusion preprocessing: Real-time monitoring time-series data of the soil and rock mass and the surface are collected through a multi-type sensor network deployed in the target geotechnical engineering area and engineering structure; wide-area surface deformation time-series data acquired by satellite remote sensing platform, high-precision topographic and geomorphological data acquired by UAV aerial survey platform, and historical geological survey and engineering activity data are integrated simultaneously; coordinate system transformation, time series alignment and data quality verification are performed on the multi-source heterogeneous monitoring data to generate a spatiotemporally aligned multi-source heterogeneous monitoring data set; the multi-type sensor network includes at least a stress sensor array for monitoring stress changes in soil and rock mass, a seepage sensor array for monitoring pore water pressure distribution, a microseismic monitoring array for capturing micro-fracture events in rock mass, and a displacement monitoring array for monitoring surface and structural displacement; S2. Construction of Differentiable Mechanisms for Geophysical and Chemical Processes: Based on the fundamental principles of rock and soil mechanics, seepage mechanics, and damage mechanics, a set of parameterized and differentiable mathematical and physical models are constructed to characterize the key intrinsic processes of geological risk development. The differentiable mechanism kernel includes at least: a hydraulic coupling mechanism kernel describing the coupling of unsaturated seepage and stress field, a stress-strain mechanism kernel characterizing the elastoplastic deformation and structural surface effect of rock and soil, and a damage evolution mechanism kernel describing the cumulative evolution of damage inside the rock mass; each mechanism kernel is expressed in the form of partial differential equations or differential algebraic equations, and its model parameters are optimized and adjusted through data-driven methods. S3. Inverse modeling of the geological system state based on physical information and data assimilation: Define the geological system state of the target area as a high-dimensional hidden state vector, which represents the stress field, strain field, pore water pressure field, and damage field of the soil and rock mass in the complete spatiotemporal domain; use the differentiable mechanism kernel constructed in step S2 as the state equation describing the evolution of the geological system state; establish the observation equation, and describe the multi-source heterogeneous monitoring data obtained in step S1 as the mapping of the geological system state under the observation operator corresponding to the multi-source heterogeneous monitoring data; adopt a solution framework combining physical information neural network and variational assimilation, with the multi-source heterogeneous monitoring data as the constraint, and solve the spatiotemporal evolution sequence of the geological system state that simultaneously satisfies the state equation and the observation equation, as well as the corresponding model parameter field, through the optimization algorithm to obtain a digital geological system dynamic model calibrated by multi-source data; S4. Risk mechanism tracing and dynamic inference analysis based on geological system state: Based on the spatiotemporal evolution sequence of the geological system state reconstructed in step S3, quantitative decomposition analysis of risk-dominant driving factors, forward numerical simulation prediction of the effects of engineering intervention measures, and identification of comprehensive instability precursor modes are carried out. S5. Meta-learning-driven adaptive optimization of mechanism kernel architecture: Construct a knowledge base containing multiple candidate mechanism kernel models and their parameter priors; For new engineering scenarios, utilize initial multi-source heterogeneous monitoring data, and use meta-learning algorithms to search and determine the mechanism kernel combination method and parameter initialization strategy to accelerate the convergence process of the geological system state inverse modeling in step S3. S6. Multidimensional visualization and decision support of geological system states: The geological system states reconstructed in step S3 and deduced in step S4 are dynamically rendered and displayed through a 3D visualization engine. It supports multi-physics field overlay display, spatiotemporal backtracking, profile analysis, virtual measuring point data extraction and comparison of different scenarios. It also integrates and displays risk tracing results, intervention scheme simulation comparison reports and early warning information, and generates a comprehensive risk assessment report.

2. The intelligent identification method for geotechnical engineering risks based on geological system state inverse modeling according to claim 1, characterized in that, Step S2 involves constructing the hydraulic coupling mechanism kernel, specifically including: S21. Based on the theory of unsaturated seepage, a three-dimensional equation with pore water pressure as the dependent variable is established. The soil hydraulic property function in the three-dimensional equation is described by a parameterized model, and the gradient of the parameterized model is calculated by automatic differentiation technique. S22. The permeable volume force obtained from the pore water pressure field calculation is used as the source term and coupled to the equilibrium equation of the stress-strain mechanism core to construct a two-way coupled calculation framework for fluid pressure and solid skeleton deformation. S23. Mathematically describe the rainfall infiltration boundary, groundwater level fluctuation boundary, and engineering structure contact boundary, and encode them as boundary condition constraints embedded in the solution framework.

3. The intelligent identification method for geotechnical engineering risks based on geological system state inverse modeling according to claim 1, characterized in that, Step S2 involves constructing a stress-strain mechanism core, specifically including: S24. Based on the theoretical framework of elastoplastic mechanics, establish the differentiable expression of the equilibrium equations and geometric equations of soil and rock mass; S25. The equivalent continuous medium method or discrete interface element method is used to explicitly model the key structural surfaces of faults and joints in the rock mass. The normal and tangential mechanical behavior of the structural surfaces is described by a differentiable nonlinear contact-friction model. S26. The engineering excavation activity is simulated as the dynamic removal of regional material properties and the release of equivalent loads within the computational domain, as determined by the engineering excavation scheme. The anchor bolt and anchor cable support structure is simulated as a one-dimensional rod element or prestressed cable element, and its interaction with the surrounding rock is described by a differentiable bond-slip constitutive model. The excavation and support processes are parameterized as functions of time and integrated into the stress-strain mechanism kernel.

4. The intelligent identification method for geotechnical engineering risks based on geological system state reverse modeling according to claim 1, characterized in that, Step S3 employs a solution framework combining physical information neural networks and variational assimilation, specifically including: S31. Construct a physical information neural network with an encoder-decoder structure. The encoder is used to extract the temporal features of multi-source heterogeneous monitoring data, and the decoder takes spatiotemporal coordinates as input and outputs the corresponding multi-physics state prediction values. S32. The residual of the partial differential equation corresponding to the differentiable mechanism kernel constructed in step S2 is incorporated into the loss function of the neural network as a mandatory physical constraint, and configuration points are sampled in the entire spatiotemporal computation domain to force the network prediction to satisfy the physical constraint. S33. Using a four-dimensional variational assimilation method, the initial system state, boundary condition parameters, and spatially distributed material parameters are combined to form a set of variables to be optimized. By minimizing the difference between the network prediction value and the multi-source heterogeneous monitoring data within a long time window, while simultaneously satisfying physical constraints, the optimal geological system state evolution sequence and parameter field are obtained by joint inversion using the gradient descent algorithm.

5. The intelligent identification method for geotechnical engineering risks based on geological system state reverse modeling according to claim 1, characterized in that, Step S4 specifically includes: S41. Quantitative analysis of risk mechanism: Based on the spatiotemporal evolution sequence of the geological system reconstructed in step S3, analyze the spatiotemporal coupling relationship and energy transfer path between the stress field, seepage field and damage field, calculate the contribution of stress redistribution, pore water pressure change and damage accumulation to the state evolution of the currently identified risk anomaly area, and output the dominant driving factor analysis report in the form of quantitative percentage. S42. Dynamic simulation of the effects of engineering intervention measures: Receive virtual engineering intervention measure parameters defined by the user for the digital geological system dynamic model, and transform the intervention measures into new boundary conditions or internal parameter modifications of the geological system dynamic model; using the current system state obtained by assimilation in step S3 as the initial condition, run the geological system dynamic model for forward numerical integration, simulate and visualize the evolution of key physical fields of the geological system state and the corresponding changes in risk indicators after the implementation of the intervention measures in a future specified period, and quantitatively evaluate the effects of different intervention schemes. S43. Intelligent identification of instability precursors based on system state phase space: Based on the historical assimilation of the geological system state time series, in the high-dimensional system state phase space composed of stress field, seepage field, and damage field variables, nonlinear time series analysis is used to extract feature quantities characterizing the critical point of system dynamic instability; a machine learning model is used to learn the mapping relationship between the feature quantities and macroscopic instability events to construct a data-driven comprehensive instability precursor identifier; the real-time short-sequence system state obtained by online assimilation is input into the identifier, and a continuous instability proximity probability index is output.

6. The intelligent identification method for geotechnical engineering risks based on geological system state reverse modeling according to claim 5, characterized in that, Step S42, which simulates the effects of engineering intervention measures, specifically includes: S421. Provide a graphical user interface to receive user definitions of the geometric parameters, material properties, and implementation sequence of drainage holes, anchoring components, or excavation support procedures. S422. Transform the user-defined intervention parameters into corresponding boundary condition modifications, material property updates, or load condition changes in the dynamic model of the digital geological system. S423. Using the currently calibrated system state as the initial condition, perform forward numerical integration with the modified model to simulate the evolution process of the geological system state within a specified future time period, and dynamically output the predicted results of displacement field, stress field, plastic zone and risk level indicators. S424. Compare and analyze the simulation results under different intervention programs, and quantitatively evaluate the impact of each program on risk indicators.

7. The intelligent identification method for geotechnical engineering risks based on geological system state reverse modeling according to claim 5, characterized in that, Step S43, which involves the comprehensive identification of instability precursor patterns, specifically includes: S431. Collect the geological system state time series data reconstructed before the occurrence of instability events in historical cases; S432. In the high-dimensional phase space of the geological system, calculate the index characterizing the dynamic characteristics of the system. The index includes the local Lyapunov exponent, the system restitution force, the order parameter extracted by principal component analysis, and the nonlinear coherence coefficient between multiple physics fields. S433. Using the aforementioned dynamic characteristic index as a feature and whether it is close to instability as a label, train a machine learning classification model, wherein the classification model constitutes a data-driven instability precursor identifier. S434. Input the real-time system state short sequence obtained by online assimilation into the instability precursor identifier, calculate and output the instability proximity probability index, and trigger an early warning when the index exceeds a preset threshold.

8. A geotechnical engineering risk intelligent identification system based on geological system state reverse modeling, characterized in that, include: A multi-source heterogeneous monitoring data fusion preprocessing module is used to collect real-time monitoring time-series data of the soil and rock mass and the surface through a multi-type sensor network deployed in the target geotechnical engineering area and engineering structure; synchronously integrate wide-area surface deformation time-series data acquired by satellite remote sensing platform, high-precision topographic and geomorphological data acquired by UAV aerial survey platform, and historical geological survey and engineering activity data; perform coordinate system transformation, time series alignment and data quality verification on the multi-source heterogeneous monitoring data to generate a spatiotemporally aligned multi-source heterogeneous monitoring data set; the multi-type sensor network includes at least a stress sensor array for monitoring stress changes in soil and rock mass, a seepage pressure sensor array for monitoring pore water pressure distribution, a microseismic monitoring array for capturing micro-fracture events in rock mass, and a displacement monitoring array for monitoring surface and structural displacement; The differentiable mechanism core construction module, based on the fundamental principles of rock and soil mechanics, seepage mechanics and damage mechanics, constructs a set of parameterized and differentiable mathematical and physical models to characterize the key intrinsic processes of geological risk formation. The differentiable mechanism kernel includes at least: a hydraulic coupling mechanism kernel describing the coupling of unsaturated seepage and stress field, a stress-strain mechanism kernel characterizing the elastoplastic deformation and structural surface effect of rock and soil, and a damage evolution mechanism kernel describing the cumulative evolution of damage inside the rock mass; each mechanism kernel is expressed in the form of partial differential equations or differential algebraic equations, and its model parameters are optimized and adjusted through data-driven methods. The geological system state inverse modeling module defines the geological system state of the target area as a high-dimensional hidden state vector, which represents the stress field, strain field, pore water pressure field, and damage field of the soil and rock mass in the complete spatiotemporal domain. The constructed differentiable mechanism kernel is used as the state equation describing the evolution of the system state. An observation equation is established, which describes the obtained multi-source heterogeneous monitoring data as a mapping of the geological system state under the observation operator corresponding to the multi-source heterogeneous monitoring data. A solution framework combining physical information neural network and variational assimilation is adopted. With the multi-source heterogeneous monitoring data as constraints, the spatiotemporal evolution sequence of the geological system state that simultaneously satisfies the state equation and the observation equation, as well as the corresponding model parameter field, is solved in reverse by an optimization algorithm to obtain a digital geological system dynamic model calibrated by multi-source data. The risk intelligent analysis and deduction module is used for risk mechanism tracing and dynamic deduction analysis based on the geological system state: based on the reconstructed spatiotemporal evolution sequence of the geological system state, it performs quantitative decomposition analysis of risk-dominant driving factors, forward numerical simulation prediction of the effects of engineering intervention measures, and identification of comprehensive instability precursor patterns. The mechanism kernel architecture adaptive optimization module is used to construct a knowledge base containing multiple candidate mechanism kernel models and their parameter priors. For new engineering scenarios, it uses initial multi-source heterogeneous monitoring data to search and determine the mechanism kernel combination method and parameter initialization strategy through meta-learning algorithm to accelerate the convergence process of the geological system state inverse modeling. The multidimensional visualization and decision support module is used to dynamically render and display the reconstructed and deduced geological system state through a three-dimensional visualization engine. It supports multi-physics field overlay display, spatiotemporal backtracking, profile analysis, virtual measuring point data extraction, and different scenario comparison functions. It also integrates and displays risk tracing results, intervention scheme simulation comparison reports, and early warning information to generate a comprehensive risk assessment report.

9. A computer program instruction, characterized in that, When the instruction is executed by the processor, it implements the intelligent identification method for geotechnical engineering risks based on geological system state reverse modeling as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the intelligent identification method for geotechnical engineering risks based on geological system state reverse modeling as described in any one of claims 1-7.

Images