A fluid boundary prediction method for instability breach of a waste dump

Abstract

This invention discloses a fluid boundary prediction method for the instability and collapse of waste disposal sites, comprising: perception and dynamic inversion: dynamic inversion obtains rheological parameters reflecting the current material state; core intelligent simulation: current terrain data and rheological parameters are input into a pre-trained physical information neural operator, which performs forward calculations at the second level to predict the fluid motion field and terrain change field in future time periods; assimilation and rolling deduction: the prediction results are assimilated with the new observation data acquired at the next moment to correct the system state and parameters, and the corrected state is used as the new initial condition to realize the dynamic closed-loop prediction of the disaster evolution process; by constructing a rolling closed loop of "perception-simulation-assimilation", the system can continuously integrate the latest observation data, automatically correct prediction deviations and update model parameters, so that the prediction can be dynamically optimized along with the disaster process, significantly improving the adaptability and forecast reliability of complex evolution processes.

Description

Technical Field

[0001] This invention relates to the field of disaster monitoring technology, specifically to a fluid boundary prediction method for the instability and collapse of spoil heaps. Background Technology

[0002] Spoil disposal sites are large-scale loose deposits formed by major engineering activities such as mining, tunnel construction, and water conservancy and hydropower projects. Their structural stability is poor, making them highly susceptible to sudden overall instability or gradual collapse under external triggers such as heavy rainfall and earthquakes. This results in high-speed, high-density debris flows or mudslides, causing devastating damage to downstream residential areas, infrastructure, and the ecological environment. Therefore, accurate and rapid prediction of the fluid movement path, impact range, and topographic erosion evolution during the collapse process is a core prerequisite for developing scientific risk mitigation plans and reducing disaster losses.

[0003] Currently, mainstream prediction and evaluation technologies mainly follow the following three paths, but all of them have significant bottlenecks: 1. Numerical simulation methods based on high-fidelity physical models: These methods rely on mechanistic models such as computational fluid dynamics (CFD) or discrete element-fluid coupling (e.g., using commercial software or open-source code like FLO-2D or RAMMS). Their core drawback lies in the high uncertainty of input parameters: the model depends on key rheological parameters (such as yield stress) in constitutive relations (e.g., Bingham and Voellmy models). Viscosity These parameters, typically calibrated through limited field sampling and laboratory testing, are insufficient to accurately reflect the dynamic rheological characteristics of materials at the collapse site due to sorting, mixing, and changes in moisture content, leading to distortion of the simulation source.

[0004] The computational efficiency is extremely low, and dynamic coupling cannot be achieved: To realize the interaction between fluid and terrain, existing methods (such as the published patent CN120235082B) mostly adopt an external iterative loop of "simulation-judgment-update terrain-resimulation". Each terrain update requires restarting the time-consuming CFD solution, and a single complete simulation often takes several hours or even days, which is completely unable to meet the realistic needs of emergency early warning for scenario simulation and decision response at the minute or even second level.

[0005] Weak model generalization ability: Models tuned for specific sites are difficult to apply directly to new scenarios with different terrain and material conditions, resulting in high reuse costs.

[0006] 2. Purely data-driven statistical and machine learning methods: These methods use historical disaster data or simulated data to train machine learning models (such as random forests and convolutional neural networks) to directly predict the flood range.

[0007] Fundamental flaws: The model is a "black box," and its predictions lack physical mechanism support, failing to describe the dynamic processes of fluid movement and terrain erosion. Its predictions heavily rely on the distribution of training data, exhibit poor generalization ability for "unseen scenarios" beyond the training set, and cannot provide key dynamic process information such as flow velocity and depth, resulting in low physical reliability.

[0008] Early physical information machine learning methods attempted to incorporate physical governing equations as constraints into neural network training (such as PINN). Key problems included inherent limitations such as "one network per scenario," difficulty in training convergence, high computational costs, and insufficient accuracy in solving highly discontinuous problems like dam-break surges. A universal, efficient, and reliable solution has yet to be developed.

[0009] To address this, a fluid boundary prediction method for instability and collapse in spoil heaps is proposed. Summary of the Invention

[0010] The purpose of this invention is to provide a fluid boundary prediction method for the instability and collapse of waste disposal sites, so as to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a fluid boundary prediction method for instability and collapse at a spoil heap, comprising the following steps: S1. Sensing and Dynamic Inversion: Based on real-time observation data obtained from a three-dimensional monitoring network deployed at the waste disposal site and downstream, rheological parameters reflecting the current state of the material are dynamically inverted. S2. Core Intelligent Simulation: The current terrain data and the rheological parameters are input into the pre-trained physical information neural operator, which performs forward calculations at the second level to predict the fluid motion field and terrain change field in the future period. S3. Assimilation and Rolling Deduction: The prediction results of S2 are assimilated with the new observation data acquired at the next moment to correct the system state and parameters. The corrected state is used as the new initial condition, and S1 to S3 are executed in a loop to achieve dynamic closed-loop prediction of the disaster evolution process.

[0012] Preferably, the three-dimensional monitoring network in S1 includes a GNSS receiver and a ground-based synthetic aperture radar for monitoring surface displacement, a high-definition and thermal imaging video monitoring station and a drone for monitoring surface flow fields, and a rain gauge and a piezometer for monitoring hydrological and meteorological conditions.

[0013] Preferably, the dynamic inversion in S1 specifically involves: constructing a differentiable physical forward model with the depth-averaged shallow water equation as its core; constructing a data fitting term based on the difference between the real-time observation data and the simulation results of the forward model, and combining it with a parameter regularization term to form an optimization objective function; and dynamically identifying the rheological parameter vector by solving for the minimum value of the optimization objective function, as shown in the following formula: ; in, For yield stress, Plastic viscosity, This is the turbulence coefficient.

[0014] Preferably, the physical information neural operator in S2 adopts an encoding-evolution-decoding architecture, and its mapping relationship is as follows: ; in, For operators, For the current terrain, For rheological parameters, , and These are the predicted future flow depth field, flow velocity field, and topographic change field, respectively.

[0015] Preferably, the encoder of the physical information neural operator is a graph neural network used to extract terrain features; the evolutionary layer operates in the Fourier domain, and its weight matrix is ​​subject to physical residual loss. Constraints, the The mass and momentum conservation equations are constructed based on the depth-averaged shallow water equation; the decoder is a fully connected network used to synchronously output the flow depth field, flow velocity field, and topographic change field.

[0016] Preferably, the data assimilation in S3 employs an ensemble Kalman filter algorithm.

[0017] Preferably, it also includes S4: during the rolling simulation process, a dynamic risk map, a visualization of the disaster evolution process, and a report on key disaster parameters are generated in real time based on the latest prediction results.

[0018] Preferably, the physical information neural operator in S2 is trained by generating a training dataset covering different failure scenarios using high-fidelity numerical simulation; jointly optimizing the operator with data fitting loss and physical residual loss, and embedding the fluid control equation as a constraint into the network weights.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention embeds the fluid dynamics control equations as physical constraints into a neural network and combines them with dynamic parameter inversion based on real-time observation. This ensures that the prediction results not only conform to physical laws but also reflect the dynamic changes in the material state, fundamentally solving the problems of inaccurate parameters and physical distortion of pure data models in traditional methods. 2. This invention utilizes a trained physical information neural operator to replace the traditional iterative CFD solution, shortening the simulation process, which takes hours or days, to seconds. This achieves a leap from "offline analysis" to "online real-time simulation," meeting the urgent need for minute-level response in emergency early warning. 3. By constructing a rolling closed loop of "perception-simulation-assimilation", the system can continuously integrate the latest observation data, automatically correct prediction biases and update model parameters, so that the prediction can be dynamically optimized along with the disaster process, significantly improving the adaptability and forecast reliability of complex evolution processes. Attached Figure Description

[0020] Figure 1 This is a flowchart of a fluid boundary prediction method for instability and collapse of a spoil heap, as proposed in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only 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 are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides a technical solution for predicting fluid boundary instability and collapse in spoil heaps: the method is based on an intelligent early warning system that integrates a "sensing and dynamic inversion layer", a "core intelligent simulation layer" and an "assimilation and rolling deduction layer".

[0023] 1. System Construction 1.1 Perception and Dynamic Inversion Layer This layer is the system's data input and status diagnosis module, responsible for converting multi-source, heterogeneous real-time monitoring data into reliable physical model input parameters, solving the problem of high parameter uncertainty in traditional methods.

[0024] (1) Three-dimensional monitoring network: A three-dimensional monitoring network consisting of the following devices will be deployed on the surface of the target spoil heap, at potential breaches, and at key downstream sections: Surface displacement monitoring: 10-15 high-precision GNSS receivers (positioning accuracy ≤ ±3mm) are evenly deployed on the surface of the spoil heap dam, and 2-3 ground-based synthetic aperture radars (GB-SAR, measurement distance resolution ≤ 0.1m) are deployed on the mountains on both sides of the dam to cover the key slip surfaces and potential instability areas of the dam, and to capture the three-dimensional displacement field and deformation rate of the dam surface in real time.

[0025] Surface flow field monitoring: Deploy one set of high-definition + thermal imaging dual-mode video monitoring stations (resolution ≥1080P, frame rate ≥30fps) at 50m upstream, 100m downstream, and 200m downstream of potential breaches in the spoil heap. Simultaneously, deploy one drone equipped with a high-definition camera (endurance ≥60 minutes, cruising speed ≤5m / s) for low-altitude inspection every 30 minutes, using particle image velocimetry (PIV) technology to obtain the real-time location of the debris flow front. Surface velocity field and flow depth distribution .

[0026] Hydrological and meteorological monitoring: Install one tipping bucket rain gauge (measurement accuracy ≤ ±0.2 mm) at the top, middle and 500 m downstream of the spoil heap. Deploy a total of 6-8 piezometers (measurement range 0-1 MPa, accuracy ≤ ±0.5% FS) at different depths (5 m, 10 m, 15 m) of the dam body to monitor disaster triggering factors such as rainfall intensity and pore water pressure in real time.

[0027] All monitoring data is transmitted back to the central server in real time via a dedicated low-latency, high-bandwidth data transmission link. The raw data undergoes time-series alignment (time synchronization accuracy ≤1s) and outlier removal (using...). Criteria), missing value completion (based on linear interpolation), forming a standardized observation dataset that is spatiotemporally synchronized. .

[0028] (2) Dynamic inversion module for rheological parameters: This module is deployed on GPU nodes of a central server, and its core is a physically-enhanced online optimization model. Its goal is to optimize at every time step... Based on observation data Dynamically identify the rheological parameter vector that best reflects the current state of the material. (in For yield stress, Plastic viscosity, (e.g., turbulence coefficient).

[0029] The inversion process reduces to solving the following constrained optimization problem: in: The goal of the optimization process is to find a way to optimize the objective function. The parameter with the smallest value ; To optimize the objective function, its value consists of two parts: the data fitting residuals and the parameter regularization term; This is a differentiable physical forward model with the depth-averaged shallow water equation as its core, inputting the current terrain. and parameters It can quickly simulate and output the flow field. ; For a moment The set of observational data actually obtained through monitoring the network; The square of the weighted Euclidean norm, It is the covariance matrix of the observation error, and its inverse is... As weights, they are obtained through statistical analysis of historical monitoring data. The initial values ​​are set as a diagonal matrix, with the diagonal elements being the reciprocals of the variance of each monitoring indicator's error. This is a regularization term used to constrain the smoothness of parameter changes and avoid overfitting; The regularization coefficient is determined through cross-validation and ranges from 0.001 to 0.01. It is used to balance the importance of "data fitting accuracy" and "prior reasonableness of parameter solutions".

[0030] An efficient solution is achieved using a gradient descent algorithm based on automatic differentiation (learning rate set to 0.001, momentum factor (momentum coefficient = 0.9) is used to accelerate convergence. The convergence condition is set as follows: the change in the objective function value ≤ 10⁻⁶ over 5 consecutive iterations, or the number of iterations reaches 1000, at which point a dynamically updated parameter vector is output). .

[0031] 1.2 Core Intelligent Simulation Layer This layer is the system's "simulation engine," and at its core is a pre-trained Physics-Informed Neural Operator (PINO). It is used to replace traditional time-consuming CFD solvers and achieve forward modeling of fluid-terrain coupling processes in seconds.

[0032] (1) Neural operator architecture: Operator The architecture adopts an "encoding-evolution-decoding" approach, with the following mapping relationship: where is the input (terrain point cloud + rheological parameters), and is the output (future flow field + terrain change rate field). The specific structure is as follows: in, The physical information neural operator is the "intelligent simulation engine" of the entire prediction method. These represent the weight parameters of the trained neural network. Operator in The result triplet of the one-step prediction output within the time limit: For predicted future moments The depth field of the flow, For predicted future moments The velocity vector field, For prediction in The topographic change field that occurs within a time period (erosion is negative, deposition is positive).

[0033] Encoder: Employs a graph neural network (GNN) to process irregularly discrete elevation point clouds. The system takes the spatial topological relationship (constructed by the K-nearest neighbor algorithm, K=10) as input, extracts multi-level terrain geometric features through 3 layers of graph attention layers (each layer has 256 hidden units), and outputs a feature vector with a dimension of 256.

[0034] Evolutionary Layer: Computation is performed in the Fourier space, with a core of three layers of Fourier Neural Operators (FNOs). Each layer contains one Fourier transform module and one linear transform module. Linear transform weight matrix. Guided by the physical principles of the discrete operators in the shallow water equations, i.e., through physical residual loss. Apply constraints: in: The rate of change of flow depth over time; The divergence of fluid mass flux; The mass conservation equation (continuity equation) requires that the increase in fluid volume within any infinitesimal element must equal the difference between the inflow and outflow rates within that element, allowing the neural network to predict... and By satisfying this equation as much as possible, we ensure that the prediction results do not violate the law of conservation of mass. The rate of change of flow velocity over time; This is convective acceleration, caused by non-uniformity of the velocity field; It is the acceleration due to gravity; The gradient of the total head (flow depth plus topographic elevation) is the pressure term that drives the fluid motion; The bed shear stress is the frictional force that hinders fluid movement; its expression is related to the rheological parameters. and flow rate Related; The fluid density is taken as 2650 kg / m³, determined based on the physical properties of the waste disposal site materials.

[0035] The decoder consists of four fully connected layers (with 512, 256, 128, and 64 hidden units respectively), which map the features from the Fourier domain back to physical space and output the future data synchronously through multiple fully connected layers. The depth field of time Flow velocity field and the rate of change of topography The topographic changes are calculated based on a differentiable empirical erosion / deposition model, ensuring consistent coupling between hydrodynamics and geomorphic evolution. (2) Offline training and deployment: Data generation: High-fidelity coupled simulation software based on the material point method (MPM) was used to generate a simulation dataset covering a wide range of failure scenarios. : Triggering mechanisms: covering three categories: heavy rainfall (rainfall intensity 50-200 mm / h), earthquake (magnitude 4-7), and human disturbance; Material condition: Yield stress range 500-5000 Pa, plastic viscosity range 10-100 Pa·s, moisture content range 10%-30%; Terrain conditions: slope range 10°-45°, elevation difference range 50-300m, including typical terrain such as canyons, gentle slopes, and obstacles.

[0036] The final dataset contains 10,000 samples, each containing input (terrain point cloud + rheological parameters) and output (flow field + terrain change rate field), and is divided into training set, validation set and test set in a 7:2:1 ratio.

[0037] Phased training: Pre-training: Minimizing data fitting loss This allows the operator to learn the basic input-output mapping. The Adam optimizer is used with a learning rate of 0.001 and 100 training epochs. Pre-training stops when the validation set loss does not decrease for 10 consecutive epochs.

[0038] Physical Constraint Refinement: Total Loss of Joint Optimization in: For data fitting loss; This is the physical residual loss; The boundary condition loss is defined as follows: the boundary conditions include solid wall boundary (velocity = 0) and open boundary (pressure = 0). The loss function is the MSE of the predicted value and the actual boundary condition at the boundary. ; ; .

[0039] A learning rate decay strategy was adopted (the learning rate was halved every 20 rounds), and 200 training rounds were conducted until the ensemble loss was verified to converge.

[0040] Model solidification: Verify performance compliance on independent test sets (Nash efficiency coefficient) After that, the model is converted into a lightweight inference engine and deployed on a cloud GPU server.

[0041] 1.3 Assimilation and Rolling Deduction Layer This layer is the "feedback and control center" of the system, responsible for using the constantly arriving new observation data to correct model predictions, achieve rolling improvement in forecast accuracy, and solve the problem that static predictions are not adapted to the evolution of disasters.

[0042] (1) Data assimilation algorithm: An ensemble Kalman filter (EnKF) is used, with an ensemble size of 50. The specific steps are as follows: Forecast step: at time It possesses a set of system states (flow field, terrain). By using the core intelligent simulation layer to perform forward predictions on each set member, a prediction set is obtained. .

[0043] Analysis step: When New observational data at time Upon arrival, calculate the observation estimate corresponding to each forecast state. The state is analyzed by updating the EnKF equation. This means making optimal corrections to prediction biases.

[0044] Status Update: Analyze the status As the optimal estimate of the system state at the current moment, it is used to update the terrain database. (Using PostgreSQL+PostGIS spatial database for storage, supporting fast query and update of terrain data), and fed back to the perception layer as part of the input for the next round of parameter inversion.

[0045] (2) Rolling closed-loop mechanism: This layer controls a fixed prediction period. (For example, 5 minutes). Within each cycle, the system executes a closed loop of "positive prediction → waiting / receiving observations → data assimilation → state / parameter update → next round of prediction," making the prediction no longer static but dynamically optimized as the disaster evolves and data accumulates. The specific steps are as follows: Positive prediction: Based on the current state and parameters, the core intelligent simulation layer predicts the evolution of the disaster in the next 5 minutes; Data reception: Continuously receive real-time data transmitted back from the monitoring network during the prediction period, preprocess it and wait for assimilation; Data assimilation: At the end of the prediction period, the EnKF algorithm is activated to fuse the prediction results with real-time observation data and correct the system state; Status / parameter update: Update the terrain database with the corrected system status, and trigger the rheological parameter inversion of the perception layer to obtain parameters adapted to the latest disaster status; Next round of prediction: Input the updated terrain, parameters and initial conditions into the core intelligent simulation layer to start the next cycle of prediction.

[0046] 2. Work Process The workflow for the system's online early warning phase is as follows: Step S1: Event triggering and initialization.

[0047] The real-time data processing module of the monitoring network continuously analyzes the data. When any of the following instability symptoms are met, the system automatically switches from monitoring mode to the highest level of early warning mode: The displacement rate of the dam surface monitored by GNSS is ≥5 mm / h for three consecutive cycles (cumulative 90s); The deformation of the area monitored by GB-SAR is ≥10mm; The video surveillance footage identified the breach (using the YOLOv8 target detection algorithm, the breach identification accuracy is ≥95%). The pore water pressure monitored by the piezometer has an increase rate of ≥0.01 MPa / h for two consecutive cycles.

[0048] Convergence Initial high-precision terrain at any given time and flow field observations based on preliminary analysis of monitoring data .

[0049] Step S2: First round of dynamic parameter inversion and prediction.

[0050] The perception and dynamic inversion layer is initiated based on The initial optimal rheological parameters were obtained by inversion within tens of seconds. .Will and Input core intelligent simulation layer, physical information neural operator Complete the future within 3-5 seconds Evolution prediction (e.g., 10 minutes), output flow field and topographic changes The system updates the terrain accordingly: .

[0051] Step S3: Rolling assimilation and closed-loop deduction.

[0052] exist time: New data assimilation: New observational data Return. The assimilation and rolling inference layer initiates the EnKF algorithm, combining the predicted flow field from step S2 with... Fusion, generating pairs Optimal estimation of the system state (flow field, topography) at any given time. Correcting prediction bias.

[0053] Parameter update: At the same time, and Feedback is sent to the perception layer, triggering a new round of rheological parameter inversion to obtain updated parameters. This is to reflect possible changes in the state of the material (such as a decrease in yield stress due to dilution with water).

[0054] Next round of predictions: Assimilation and correction of terrain (Already used) Updated), new parameters Together with the corrected initial flow field conditions, the core intelligent simulation layer is restarted to predict... The evolution of time.

[0055] Step S4: Loop execution and decision output.

[0056] Repeat step S3 for a fixed period. The "prediction-assimilation-update" closed loop is executed repeatedly until any of the following stopping conditions are met: The monitoring network has not detected any signs of instability for 10 consecutive cycles, and the core intelligent simulation layer predicts the flow field intensity (average flow velocity ≤ 0.5 m / s, average flow depth ≤ 0.3 m) or the emergency command system issues an alarm cancellation command; In each cycle, the system synchronously generates and updates dynamic risk maps, disaster evolution animations, and reports on key parameters (such as peak flow and arrival time of the leading edge) based on the latest prediction results, and intelligently matches evacuation routes and response plans for the emergency command system.

[0057] Working Principle: A novel simulation-assimilation closed-loop system is constructed, characterized by "physical mechanism-driven, real-time data correction, and efficient parallel computation." Its core utilizes a "surrogate model"—a physical information neural operator—to transform the time-consuming numerical solution process of partial differential equations in traditional CFD into a second-level functional mapping query based on a neural network. Simultaneously, dynamic parameter inversion transforms the most uncertain inputs (rheological parameters) in the model into "state variables" based on real-time observations, and data assimilation technology continuously merges model predictions with real observations, constantly correcting the system state. These three interconnected stages overcome the bottlenecks of traditional methods in terms of efficiency, accuracy, and dynamic adaptability, achieving a fundamental leap from "static estimation" to "second-level dynamic prediction" of the collapse flow boundary in spoil heaps.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fluid boundary prediction method for instability and collapse in spoil heaps, characterized in that: Includes the following steps: S1. Sensing and Dynamic Inversion: Based on real-time observation data obtained from a three-dimensional monitoring network deployed at the waste disposal site and downstream, rheological parameters reflecting the current state of the material are dynamically inverted. S2. Core Intelligent Simulation: The current terrain data and the rheological parameters are input into the pre-trained physical information neural operator, which performs forward calculations at the second level to predict the fluid motion field and terrain change field in the future period. S3. Assimilation and Rolling Deduction: Assimilate the prediction results of S2 with the new observation data acquired at the next moment, correct the system state and parameters, and use the corrected state as the new initial condition to cyclically execute steps S1 to S3 to achieve dynamic closed-loop prediction of the disaster evolution process.

2. The fluid boundary prediction method for instability and collapse of a spoil heap according to claim 1, characterized in that: The three-dimensional monitoring network in S1 includes a GNSS receiver and ground-based synthetic aperture radar for monitoring surface displacement, a high-definition and thermal imaging video monitoring station and drone for monitoring surface flow fields, and rain gauges and piezometers for monitoring hydrological and meteorological conditions.

3. The fluid boundary prediction method for instability and collapse at a spoil heap according to claim 1, characterized in that: The dynamic inversion in S1 specifically involves: constructing a differentiable physical forward model with the depth-averaged shallow water equation as its core; constructing a data fitting term based on the difference between the real-time observation data and the simulation results of the forward model, and combining it with a parameter regularization term to form an optimization objective function; and dynamically identifying the rheological parameter vector by solving for the minimum value of the optimization objective function, as shown in the following formula: ； in, For yield stress, Plastic viscosity, This is the turbulence coefficient.

4. The fluid boundary prediction method for instability and collapse of a spoil heap according to claim 1, characterized in that: The physical information neural operator in S2 adopts an encoding-evolution-decoding architecture, and its mapping relationship is as follows: ； in, For operators, For the current terrain, For rheological parameters, , and These are the predicted future flow depth field, flow velocity field, and topographic change field, respectively.

5. The fluid boundary prediction method for instability and collapse of a spoil heap according to claim 4, characterized in that: The encoder of the physical information neural operator is a graph neural network, used to extract terrain features; The evolutionary layer operates in the Fourier domain, and its weight matrix is ​​subject to physical residual loss. Constraints, the The mass and momentum conservation equations are constructed based on the depth-averaged shallow water equation; The decoder is a fully connected network used to synchronously output the flow depth field, flow velocity field, and terrain change field.

6. The fluid boundary prediction method for instability and collapse at a spoil heap according to claim 1, characterized in that: The data assimilation in S3 employs an ensemble Kalman filter algorithm.

7. The fluid boundary prediction method for instability and collapse of a spoil heap according to claim 1, characterized in that: It also includes S4: During the rolling simulation process, dynamic risk maps, disaster evolution visualization, and key disaster parameter reports are generated in real time based on the latest prediction results.

8. The fluid boundary prediction method for instability and collapse of a spoil heap according to claim 1, characterized in that: The physical information neural operator in S2 is trained in the following way: a training dataset covering different failure scenarios is generated using high-fidelity numerical simulation; the operator is jointly optimized using data fitting loss and physical residual loss, and the fluid control equation is embedded as a constraint into the network weights.

Images