Multi-mode tailing pond dam break digital twinning emergency deduction and decision optimization method

Abstract

The invention provides a multi-mode tailing pond dam break digital twinning emergency deduction and decision optimization method, which comprises the following steps of: firstly, constructing a high-precision three-dimensional scene model, performing simulation calculation on the high-precision three-dimensional tailing pond model based on a finite element model to obtain a dam body stress-strain field, and performing simulation calculation on the dam body stress-strain field based on ANSYS + reinforcement learning. And predicting a dam break probability distribution diagram by using the dam body stress-strain field and the current environment state, generating a rescue scheduling scheme based on the dam break probability distribution diagram and the resource distribution data, building a dynamic virtual deduction scene for exercise, and verifying and optimizing the rescue scheduling scheme according to exercise record data. And a final rescue scheduling scheme is obtained. According to the method, the ANSYS + reinforcement learning hybrid engine is utilized to reinforce the deduction precision, dam break path prediction, optimal resource scheduling and plan self-optimization closed loop are realized, the problems of data fusion distortion, deduction stiffness, decision lag and the like in the traditional technology are solved, and the method is particularly suitable for variable working conditions such as tailing pond seepage field sudden change and extreme weather.

Description

Technical Field

[0001] The present invention relates to the technical field of emergency management, and particularly to a multimodal digital twin emergency deduction and decision optimization method for tailings dam failures. Background Art

[0002] Traditional disaster drill methods mainly rely on on-site drills and tabletop deductions. On-site drills (such as fire extinguishing and earthquake evacuation) can provide real experiences, but they are usually difficult to organize, costly, limited in scale, and difficult to simulate complex or extreme disaster scenarios, with certain safety risks. Tabletop deductions mainly rely on paper-based plans, maps, and oral discussions, lacking interactivity, immersion, and realism, and are difficult to effectively train the emergency response and decision-making abilities of participants. With technological progress, technologies such as computer simulation, virtual reality (VR), and augmented reality (AR) have been introduced into the drill field, providing a more controllable, repeatable, and safe simulation environment. However, existing technical solutions still have significant deficiencies in aspects such as scene fidelity, multi-role collaborative linkage, dynamic assessment and feedback during the drill process, dynamic matching and optimization of plans, and low-cost large-scale deployment. In particular, the lack of an integrated solution that efficiently integrates multiple technical means, intelligently generates dynamic disaster situations, and achieves precise quantitative evaluation of drill effects limits the depth and breadth of drill effects. Therefore, there is an urgent need to develop more advanced, intelligent, efficient, and easy-to-promote disaster drill methods.

[0003] Currently, the existing technology mainly realizes emergency deduction through the following two types of solutions: Static parameter deduction system: It relies on preset parameters (such as fixed rainfall and dam height) for finite element simulation (such as ANSYS), but cannot respond to environmental changes in real time (such as sudden heavy rain).

[0004] Single-modal data-driven deduction: It generates a risk model based on a single data source (such as GNSS displacement monitoring), but does not solve the fusion accuracy problem of multi-source data (satellite images, LiDAR point clouds, sensors).

[0005] For example: The solution with the document number CN119248108A and the patent name Emergency Drill System Based on Big Data is used to collect on-site image and video data in real time using cameras, UAV devices, and mobile terminals, identify key information at the accident scene using image recognition algorithms, process and analyze the real-time data obtained from the data collection and recognition module, predict the development of emergency events, and make decisions and simulation tasks at different levels based on this.

[0006] The document number is CN118504399A, and the patent name is a solution for an emergency drill method and system based on digital twin and knowledge graph. Through the virtual environment constructed by digital twin technology, various environmental data of the site are collected in real time by an external parameter measurement device and then transmitted to the system. Through data processing and analysis, combined with the reasoning ability of the knowledge graph, various complex emergency scenarios and emergencies can be simulated, so as to comprehensively test and evaluate the feasibility and effectiveness of the emergency plan; the emergency plan can also be automatically optimized and improved according to the drill results, improving the efficiency and accuracy of dealing with emergencies.

[0007] As can be seen from the above, the existing technologies have the following defects: 1) The emergency drill system based on big data has the following defects: ① Through data-driven deduction, based on traditional big data analysis, the problem of multi-modal data size difference is not solved; ② In terms of interactivity, it relies on preset scenario deduction and has no user dynamic operation interface; ③ For general emergency scenario analysis, the industry adaptability to tailings ponds is poor; ④ Lack of dynamic modeling ability for scenarios (fusion of dynamic seepage field, phreatic line data and point cloud data of tailings pond dam break); 2) The emergency drill system based on digital twin and knowledge graph has the following defects: ① In the process of data fusion, there is also the problem of data scale difference; ② Knowledge graph reasoning generates an emergency decision thematic map through entity relationship analysis to optimize the emergency plan process, and there is a long response delay time; ③ The real-time coupling of physical simulation and large model prediction is not realized; ④ Lack of a multi-objective resource matching algorithm for emergency scenario problems. Summary of the Invention

[0008] In view of the above situation, the main purpose of the present invention is to propose a multi-modal digital twin emergency deduction and decision optimization method and system for tailings pond dam break to solve the above technical problems.

[0009] The present invention proposes a multi-modal digital twin emergency deduction and decision optimization method for tailings pond dam break, and the method includes the following steps: Step 1, collect remote sensing data, meteorological data and tailings pond monitoring data, and construct a high-precision three-dimensional scene model; Step 2, perform simulation calculations on the high-precision three-dimensional tailings pond model based on the finite element model, simulate the structural mechanical properties of the tailings pond, and obtain the dam body stress-strain field; Step 3: Based on the PPO reinforcement learning algorithm, use the dam stress-strain field and the current environmental state to predict the dam-break probability distribution map; Step 4: Generate a rescue dispatch plan based on the dam-break probability distribution map and the resource distribution data; Step 5: Build a VR scene based on the high-precision three-dimensional tailings pond model, and then import the dam-break probability distribution map and the resource distribution data into the VR scene to obtain a dynamic virtual deduction scene; Step 6: In the dynamic virtual deduction scene, use VR devices to conduct drills according to the rescue dispatch plan, and verify and optimize the rescue dispatch plan based on the drill record data to obtain the final rescue dispatch plan.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Advantage of dynamic scene generation: Breakthrough in accuracy achieved through multi-modal data fusion.

[0011] In the process of traditional data application, due to the differences in spatio-temporal reference and dimension between multi-modal data, the progress of data fusion is not good, which easily leads to fusion distortion. By using hierarchical adaptive normalization (AdaIN) to eliminate the dimension differences of sensors and aligning the coordinate systems with UTM projection, cross-modal GNSS and InSAR data fusion with millimeter-level error (<1mm) is realized, greatly improving the dam displacement monitoring accuracy, and the reliability for mutation scenarios is further enhanced; 2. Advantage of intelligent deduction: Real-time deduction closed-loop driven by dynamic interaction Traditional emergency deduction relies on preset scripts and cannot dynamically adjust key parameters (such as the width of the flood discharge channel ΔW), resulting in rigid deduction. The drag-and-drop user interaction method is adopted, and the drag operation triggers the update of the hybrid engine (ANSYS finite element calculation + ROM reduced-order model + reinforcement learning heat map warning generation) in real time, forming an "operation-deduction-feedback" closed-loop. By adopting this technical solution, the decision-making efficiency can be greatly improved, the response time for dam-break path prediction can be reduced, and at the same time, dynamic intervention of influencing parameters is supported, improving the flexibility of the system.

[0012] 3. Dynamic matching of plan risk and resources In the traditional way of dealing with risks, it generally relies on manual experience and does not quantify the risk probability and cost trade-off. Through the dynamic association of the risk heat map and the resource distribution, the NSGA-II multi-objective algorithm generates the Pareto optimal solution set, and based on the dam-break probability distribution (output of reinforcement learning) and the material position data, the loss and cost are refined and balanced.

[0013] The additional aspects and advantages of the present invention will be partially given in the following description, partially will become obvious from the following description, or can be understood through the embodiments of the present invention. Description of the Drawings

[0014] Figure 1 This is a flow chart of the multi-modal tailings dam breach digital twin emergency simulation and decision optimization method proposed in the present invention; Figure 2 This is the overall framework diagram of the multi-modal tailings dam breach digital twin emergency simulation and decision optimization method proposed in this invention. DETAILED DESCRIPTION

[0015] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0016] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0017] See also Figure 1 This embodiment provides a multi-modal tailings dam breach digital twin emergency simulation and decision optimization method, the method comprising the following steps: Step 1: Collect remote sensing data, meteorological data and tailings pond monitoring data to build a high-precision three-dimensional scene model; As a preferred embodiment of the present invention, collecting remote sensing data, meteorological data and tailings pond monitoring data, and constructing a high-precision three-dimensional scene model specifically includes the following steps: Obtain GNSS displacement meter data, piezometer data, weather station data, tailings pond LiDAR point cloud and satellite remote sensing images, align data streams by timestamp, and remove outliers; The tailings pond LiDAR point cloud is denoised, and the key point clouds of the tailings pond dam and surrounding surface are retained to obtain the denoised tailings pond LiDAR point cloud; The satellite remote sensing image and the denoised tailings pond LiDAR point cloud are aligned in spatial scale through bilinear interpolation to obtain a satellite image with a unified coordinate system. The WGS84 coordinates of the GNSS displacement meter data and the piezometer data are converted into the tailings pond LiDAR point cloud coordinate system through UTM projection to obtain the GNSS displacement meter data and the piezometer data in the unified coordinate system; The WGS84 coordinates of the geometrically corrected satellite image are converted to the UTM projection, and then aligned to the tailings pond LiDAR point cloud coordinate system through affine transformation, and cropped to obtain a satellite image with a unified coordinate system; Map the meteorological station data to the coordinate system of the tailings pond LiDAR point cloud through inverse distance weighted interpolation to obtain meteorological parameters in the unified coordinate system; Independently normalize each band of the satellite image in the unified coordinate system and retain the contrast of the surface coverage features to obtain a normalized satellite image; Normalize the elevation part of the denoised tailings pond LiDAR point cloud to obtain a normalized LiDAR elevation field; Calculate the mean and variance within the window of the GNSS displacement meter data, piezometer data, and meteorological parameters in the unified coordinate system respectively by using a sliding window method, and linearly map them to the interval [-1, 1] to obtain normalized GNSS displacement meter data, normalized seepage pressure distribution, and normalized meteorological parameters; Perform spatio-temporal alignment on the normalized LiDAR elevation field, normalized satellite image, normalized GNSS displacement meter data, normalized piezometer data, and normalized satellite image by using an improved Kalman filter to obtain spatio-temporally aligned fusion data, and construct a high-precision three-dimensional tailings pond model based on the spatio-temporally aligned fusion data.

[0018] As a preferred embodiment of the present invention, collecting remote sensing data, meteorological data, and tailings pond detection data, constructing a high-precision three-dimensional scene model specifically includes the following steps: performing spatio-temporal alignment on the normalized LiDAR elevation field, normalized satellite image, normalized GNSS displacement meter data, normalized piezometer data, and normalized satellite image by using an improved Kalman filter to obtain spatio-temporally aligned fusion data, and constructing a high-precision three-dimensional tailings pond model based on the spatio-temporally aligned fusion data specifically includes the following steps: Construct a state vector and a physical model according to the normalized GNSS displacement meter data, normalized seepage pressure distribution, and normalized meteorological parameters, and dynamically adjust the weights of the state vector and the physical model according to the sensor accuracy; Use the state vector and the physical model to predict the current dam displacement and pressure trend, and use the prediction results for correction. After the prediction-correction cycle is completed, obtain spatio-temporally aligned fusion data, and the spatio-temporally aligned fusion data includes millimeter-level displacement, continuous pressure distribution, and gridded meteorological parameters; Overlay the normalized LiDAR elevation field and the normalized GNSS displacement meter data, and generate a three-dimensional surface grid with centimeter-level resolution through Delaunay triangulation; Map the normalized piezometer data to the internal nodes of the three-dimensional surface network to construct a pore water pressure field; interpolate the normalized meteorological parameters to the three-dimensional surface grid to generate rainfall / wind speed distribution; based on the spatio-temporally aligned fusion data, adjust the grid node coordinates in real time to simulate the dam deformation, and obtain a high-precision three-dimensional tailings pond model integrating displacement, pressure, meteorology, and terrain.

[0019] Step 2: Perform simulation calculations on the high-precision three-dimensional tailings pond model based on the finite element model, simulate the structural mechanical properties of the tailings pond, and obtain the stress-strain field of the dam body; As a preferred embodiment of the present invention, performing simulation calculations on the high-precision three-dimensional tailings pond model based on the finite element model, simulating the structural mechanical properties of the tailings pond, and obtaining the stress-strain field of the dam body specifically includes the following steps: Extract the geometric contour of the dam body from the high-precision three-dimensional tailings pond model, distinguish materials, and load the constitutive model parameters of the materials for the materials to obtain a finite element geometric model; Use hexahedral mesh encryption for the potential plastic zone in the finite element geometric model, use tetrahedral meshes for the far-field foundation, and generate contact elements and target element paired meshes for the contact surface between the dam body and the water body to obtain a non-uniformly encrypted finite element mesh model; Process the rockfill body in the non-uniformly encrypted finite element mesh model using the displacement-pore pressure coupling equation, use the pore water pressure field as the volume force boundary condition, and apply displacement constraints to obtain a finite element solution model containing complete physical equations and boundary conditions; Input the high-precision three-dimensional tailings pond model as an external dynamic load into the finite element solution model containing complete physical equations and boundary conditions to obtain a full-order finite element model, and iteratively solve the full-order finite element model to obtain the critical reduction coefficient, the coordinates of the unstable slip surface, and the strain field evolution data; Map the full-order finite element model to a reduced-order model, encapsulate it into a callable module, and generate a stress-strain field distribution cloud map of the dam body according to the critical reduction coefficient, the coordinates of the unstable slip surface, and the strain field evolution data.

[0020] Step 3: Based on the PPO reinforcement learning algorithm, use the stress-strain field of the dam body and the current environmental state to predict the dam break probability distribution map; As a preferred embodiment of the present invention, based on the PPO reinforcement learning algorithm, using the stress-strain field of the dam body and the current environmental state to predict the dam break probability distribution map specifically includes the following steps: Based on the current environmental state and the current action, use the reduced-order model to output the stress-strain field distribution cloud map of the dam body, the distribution of the equivalent plastic strain zone, and the strength reduction coefficient. The current environmental state includes the water level elevation, the rainfall intensity, and the soil saturation degree, and the current action includes the flood discharge channel operation instruction and the personnel evacuation operation instruction; Construct a reward function with the magnitude of the strength reduction coefficient as the stability reward, an exponential penalty triggered when the area of the equivalent plastic strain zone distribution exceeds the threshold as the risk penalty, and the matching result of the flood discharge channel opening state and the personnel evacuation hot zone as the artificial rule reward to calculate the reward of the output result of the reduced-order model; Based on the Actor-Critic network architecture, the flood discharge channel opening, sub-dam reinforcement strength, and drainage pump start-stop frequency are used as actions. The Actor network generates online actions according to the current environmental state. The output results of the reduced-order model are converted into target actions for learning. The Critic network is used to predict the expected return based on the online actions, and the reward of the reduced-order model output results is used as the expected value for learning to guide the generation of online actions of the Actor network. After learning is completed, the optimal action is output. Random disturbances are superimposed on the optimal action to perform Monte Carlo simulation, and the frequency of dam breach events during the Monte Carlo simulation is counted to generate a two-dimensional probability distribution in time and space. The time-space two-dimensional probability distribution is interpolated to the grid nodes of the full-order finite element model, and the dam body stress-strain field distribution cloud map is superimposed for weighted rendering to obtain the dam break probability distribution thermal map.

[0021] Step 4: Generate a rescue dispatch plan based on the dam break probability distribution map and resource distribution data; As a preferred embodiment of the present invention, generating a rescue dispatching plan based on the dam break probability distribution map and resource distribution data specifically includes the following steps: The dam break probability distribution heat map is divided into risk areas of different levels according to the threshold; According to the obtained resource distribution data, the corresponding resource points are obtained, and the corresponding utility weights are given according to the resource categories; Based on the road network topology and real-time traffic conditions, resource points and risk areas are associated with the nearest path, and the estimated travel time is marked to obtain a risk-resource association matrix including risk level, resource type, path time and utility weight; A multi-objective optimization model was constructed with the optimization objectives of minimizing the probability of dam failure, minimizing the affected population density, minimizing the economic loss coefficient, minimizing the average arrival time from the resource point to the risk area, and maximizing the matching degree between the material distribution and the demand forecast, and with the total resource limit and time window limit as constraints. Based on the risk-resource association matrix, the multi-objective optimization model is iteratively optimized to obtain the Pareto optimal solution set; Calculate the comprehensive score of each solution in the Pareto optimal solution set based on economic losses, arrival time, and material distribution utilization; The Pareto optimal solution set is sorted according to the comprehensive score and divided into priority execution plans and backup plans to obtain a hierarchical rescue dispatch plan.

[0022] Step 5: Build a VR scene based on the high-precision three-dimensional tailings pond model, and then import the dam break probability distribution map and resource distribution data into the VR scene to obtain a dynamic virtual simulation scene; As a preferred embodiment of the present invention, a VR scene is built based on a high-precision three-dimensional tailings pond model, and then the dam break probability distribution map and resource distribution data are imported into the VR scene to obtain a dynamic virtual deduction scene, which specifically includes the following steps: Map the hierarchical rescue dispatch plan to the high-precision three-dimensional tailings pond model to generate a planned route; Add mechanical properties to the virtual objects of the high-precision three-dimensional tailings pond model, and align the scene time axes of all terminals based on the cloud synchronization server to obtain a dynamic virtual deduction scene.

[0023] Step 6: In the dynamic virtual deduction scene, use VR devices to conduct drills according to the rescue dispatch plan, and verify and optimize the rescue dispatch plan based on the drill record data to obtain the final rescue dispatch plan.

[0024] As a preferred embodiment of the present invention, in the dynamic virtual deduction scene, use VR devices to conduct drills according to the rescue dispatch plan, and verify and optimize the rescue dispatch plan based on the drill record data to obtain the final rescue dispatch plan, which specifically includes the following steps: Allocate VR devices to each staff member and determine the positions of the staff members in the dynamic virtual deduction scene; In the dynamic virtual deduction scene, generate the optimal path and a highlighted indication arrow based on the positions of the staff members and the planned route; The staff members move forward and perform drill operations according to the highlighted indication arrow and record the drill data; The commander makes real-time intervention and adjustment of the drill process based on the current drill process and the overall situation of the dispatch, and records the commander's command data; According to the drill data, obtain the VR device operation records, actual resource consumption, actual resource dispatch time, and dam break suppression effect, and assign manual labels to the commander's command data; Construct a multi-dimensional evaluation matrix based on the deviation rate of the actual resource dispatch time from the planned time, the matching of the actual resource consumption with the demand prediction, and the decline amplitude of the dam break probability after the deduction; Based on the multi-dimensional evaluation matrix, conduct evaluation and analysis through a causal reasoning engine, classify the executed rescue dispatch plan into a high-quality plan and a low-quality plan, and identify the core operations affecting the results during the drill to obtain the key decision points; Store the high-quality plan in the case knowledge base and use it as the final rescue dispatch plan; Obtain the key decision points of the low-quality plan, and optimize the parameters of the multi-objective optimization model according to the key decision points of the low-quality plan to realize parameter tuning using the low-quality plan as a negative example.

[0025] The architecture of the present invention is as Figure 2As shown in the figure, it includes a data acquisition layer, an intermediate processing layer, and an output linkage layer. The data acquisition layer is used for data acquisition, mainly including on-site devices such as GNSS displacement meters, piezometers, and weather stations. Remote sensing data includes InSAR satellites and UAV LiDAR. The output linkage layer mainly includes desktop and mobile terminals for display and operation, and an emergency device control instruction stream for sending instructions.

[0026] The intermediate processing layer mainly includes three key functions: 1) Dynamic scene: Collect device data and remote sensing data from the data acquisition layer and extract key features, such as deploying GNSS displacement meters (accuracy ±1mm), piezometers (range 0 - 2MPa), weather stations (wind speed / rainfall monitoring), etc. in the reservoir area. Use the Kalman filtering algorithm to align the spatio-temporal benchmarks of multi-modal data and construct a high-precision three-dimensional scene model.

[0027] 2) Intelligent deduction engine: Physical simulation core: Calculate the stress-strain field of the dam based on the finite element model; AI deduction core: Train an agent using the PPO reinforcement learning algorithm. The input parameters include real-time water level, rainfall intensity, and soil saturation, and the output is the probability distribution map of dam break.

[0028] 3) Plan self-optimization: Digital twin editor: Provide a calculator for the influence radius of dam break, such as dam height, reservoir capacity, geological correction coefficient, etc.; Closed-loop feedback mechanism: Reverse-optimize the reward function of the AI deduction core through the drill results, such as adjusting the weight coefficient of casualties.

[0029] In this embodiment, the present invention adopts a drag-and-drop data fusion interface. In the emergency drill of the tailings pond, the user drags the UAV LiDAR point cloud into the scene, and the system automatically aligns the GNSS displacement data and meteorological data to generate a three-dimensional model of the dynamic dam break path.

[0030] ① Visual drag-and-drop tool: The user drags satellite images, point clouds, and sensor data into the three-dimensional scene editor, triggering the feature alignment algorithm (spatio-temporal difference technology). ② Automatic feature recognition: The algorithm automatically extracts key features in the data (such as the edge of the secondary dam of the tailings pond and the crack displacement points) and updates the three-dimensional model in real time.

[0031] The overall business operation logic is as follows: 1) First, collect data through sensors.

[0032] Through the improved Kalman filtering algorithm: Introduce an adaptive instance normalization (AdaIN) module, and use the following formula to eliminate the scale difference of sensor data.

[0033] ; Among them, x represents GNSS displacement data; y represents the LiDAR point cloud coordinate system; 2) Cross-modal calibration process: Align the WGS84 coordinate system (GNSS) with the local grid coordinate system (LiDAR) through UTM projection transformation. The GNSS displacement data (with millimeter-level accuracy) and the LiDAR point cloud generate a three-dimensional grid through oblique photography modeling, and the real-time meteorological data is mapped to the grid nodes through an interpolation algorithm.

[0034] 3) Hybrid deduction engine ① Hybrid deduction engine: Physical simulation (ANSYS finite element model) simulates the structural mechanical properties of the tailings pond. The dam body uses hybrid interpolation (u-P hybrid formula) to handle incompressible deformation, and the CONTA174+TARGE170 equation is selected for the contact surface with water to simulate the potential slip surface. The potential plastic zone (dam slope, core wall) is encrypted to a resolution of 0.5 - 1 times the dam height, and the far-field foundation uses a gradually sparse grid to reduce the calculation pressure. The dynamic process of geological changes such as water level changes and earthquakes is simulated through the strength reduction method. Finally, the maximum principal stress, equivalent plastic strain zone, and reduction factor FS when the strength reduction does not converge are predicted based on the calculation results, and the visualization cloud map and prediction data are finally output. In addition, the simulation model is mapped into a reduced-order model for fast calculation and encapsulated into a callable simulation module. On-site personnel can customize input parameters and obtain simulation feedback immediately; ② User interaction drive: The reinforcement learning module predicts the risk probability distribution (such as the curve of dam break probability changing with time).

[0035] 4) Risk dynamic modeling ② User interaction drive: Adjust parameters (such as the reinforcement strength of the secondary dam, the width of the flood discharge channel) by dragging, and trigger the AI to dynamically update the risk heat map.

[0036] 5) Risk-resource dynamic matching ① Dynamic matching algorithm: Based on the risk probability distribution and resource distribution data (such as the location of emergency supplies, the number of rescue personnel), generate a Pareto optimal solution set.

[0037] ② Recommendation logic: Prioritize matching high-risk areas (such as the upstream secondary dam of the dam break) with the nearest resource points, and calculate the weights of response time and success rate.

[0038] 6) Closed-loop optimization mechanism ① Digital Twin Editor: The simulation results are fed back to the plan library, and users can modify the strategy through the drag-and-drop interface (sliders adjust the priority of evacuation routes, and hot spots select resource delivery areas). Construct an emergency decision-making model based on reinforcement learning, and optimize disaster avoidance path planning and rescue resource scheduling through algorithms such as deep networks. The system can evaluate the effects of different disposal plans in real time, rehearse the execution process of each plan in a virtual environment, and mark the optimal action route in an augmented reality way. Commanders coordinate the overall situation through a three-dimensional panoramic perspective, and rescue personnel wear VR equipment for immersive operation training. Force feedback equipment can simulate the operating resistance of rescue machinery, and cloud synchronization technology ensures the consistency of multi-terminal drill scenarios.

[0039] ② Self-optimization logic: The system records historical deduction data, optimizes parameter weights through genetic algorithms, generates a new version of the emergency plan, pushes it to the mobile command system, establishes an evaluation matrix including response timeliness, resource utilization and other dimensions, and analyzes key decision points in the exercise process through a causal reasoning engine. The system automatically generates a case knowledge base to drive the continuous optimization and update of the emergency plan, forming a closed loop of "exercise-evaluation-improvement".

[0040] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0041] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0042] In the description of this specification, the descriptions referring to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0043] The above-described embodiments merely represent several implementation manners of the present invention. The descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several variations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the appended claims.

Claims

1. A multi-modal tailings dam breach digital twin emergency simulation and decision optimization method, characterized in that: The method comprises the following steps: Step 1: Collect remote sensing data, meteorological data and tailings pond monitoring data to build a high-precision three-dimensional scene model; Step 2: Based on the finite element model, a high-precision three-dimensional tailings pond model is simulated and calculated to simulate the mechanical characteristics of the tailings pond structure and obtain the stress-strain field of the dam body; Step 3: Based on the PPO reinforcement learning algorithm, the dam body stress-strain field and the current environmental status are used to predict the dam break probability distribution map; Step 4: Generate a rescue dispatch plan based on the dam break probability distribution map and resource distribution data; Step 5: Build a VR scene based on the high-precision three-dimensional tailings pond model, and then import the dam break probability distribution map and resource distribution data into the VR scene to obtain a dynamic virtual simulation scene; Step 6: In the dynamic virtual simulation scenario, use VR equipment to conduct exercises according to the rescue dispatch plan, and verify and optimize the rescue dispatch plan based on the exercise record data to obtain the final rescue dispatch plan.

2. A multi-modal tailings dam breach digital twin emergency simulation and decision optimization method according to claim 1, characterized in that: In step 1, collecting remote sensing data, meteorological data and tailings pond detection data, and constructing a high-precision three-dimensional scene model specifically includes the following steps: Obtain GNSS displacement meter data, piezometer data, weather station data, tailings pond LiDAR point cloud and satellite remote sensing images, align data streams by timestamp, and remove outliers; The tailings pond LiDAR point cloud is denoised, and the key point clouds of the tailings pond dam and surrounding surface are retained to obtain the denoised tailings pond LiDAR point cloud; The satellite remote sensing image and the denoised tailings pond LiDAR point cloud are aligned in spatial scale through bilinear interpolation to obtain a satellite image with a unified coordinate system. The WGS84 coordinates of the GNSS displacement meter data and the piezometer data are converted into the tailings pond LiDAR point cloud coordinate system through UTM projection to obtain the GNSS displacement meter data and the piezometer data in the unified coordinate system; The WGS84 coordinates of the geometrically corrected satellite image are converted to the UTM projection, and then aligned to the tailings pond LiDAR point cloud coordinate system through affine transformation, and cropped to obtain a satellite image with a unified coordinate system; The meteorological station data is mapped to the tailings pond LiDAR point cloud coordinate system through inverse distance weighted interpolation to obtain the meteorological parameters of the unified coordinate system; Each band of the satellite image in the unified coordinate system is independently normalized, and the contrast of the surface coverage characteristics is retained to obtain a normalized satellite image; The elevation part of the tailings pond LiDAR point cloud after denoising is normalized to obtain the normalized LiDAR elevation field; The GNSS displacement meter data in the unified coordinate system, the osmometer data in the unified coordinate system, and the meteorological parameters in the unified coordinate system are respectively used to calculate the mean and variance in the window by using the sliding window method, and linearly mapped to the [-1,1] interval to obtain the normalized GNSS displacement meter data, normalized osmotic pressure distribution, and normalized meteorological parameters; The normalized LiDAR elevation field, normalized satellite images, normalized GNSS displacement meter data, normalized piezometer data, and normalized satellite images are aligned in space and time using an improved Kalman filter to obtain fused data that is aligned in space and time. Based on the fused data that is aligned in space and time, a high-precision three-dimensional tailings pond model is constructed.

3. A multimodal digital twin emergency deduction and decision optimization method for tailings dam failure according to claim 2, characterized in that, The specific steps for aligning the normalized LiDAR elevation field, normalized satellite images, normalized GNSS displacement meter data, normalized piezometer data, and normalized satellite images in space and time using an improved Kalman filter to obtain fused data that is aligned in space and time, and constructing a high-precision three-dimensional tailings pond model based on the fused data that is aligned in space and time are as follows: Construct a state vector and a physical model based on the normalized GNSS displacement meter data, normalized seepage pressure distribution, and normalized meteorological parameters, and dynamically adjust the weights of the state vector and the physical model according to the sensor accuracy. Use the state vector and the physical model to predict the current dam displacement and pressure trend, and use the prediction results for correction. After the prediction-correction cycle is completed, fused data that is aligned in space and time is obtained. The fused data that is aligned in space and time includes millimeter-level displacement, continuous pressure distribution, and gridded meteorological parameters. Overlay the normalized LiDAR elevation field and the normalized GNSS displacement meter data, and generate a three-dimensional surface grid with centimeter-level resolution through Delaunay triangulation. Map the normalized piezometer data to the internal nodes of the three-dimensional surface network to construct a pore water pressure field. Interpolate the normalized meteorological parameters to the three-dimensional surface grid to generate rainfall / wind speed distribution. Based on the fused data that is aligned in space and time, adjust the grid node coordinates in real time to simulate the deformation of the dam, and obtain a high-precision three-dimensional tailings pond model that integrates displacement, pressure, meteorology, and terrain.

4. A multimodal digital twin emergency deduction and decision-making optimization method for tailings pond dam break according to claim 3, characterized in that In step 2, perform simulation calculations on the high-precision three-dimensional tailings pond model based on a finite element model to simulate the structural mechanical properties of the tailings pond, and obtain the dam stress-strain field. The specific steps are as follows: Extract the dam geometric profile from the high-precision three-dimensional tailings pond model, distinguish materials, and load material constitutive model parameters for the materials to obtain a finite element geometric model. Use hexahedral mesh refinement for the potential plastic zone in the finite element geometric model, use tetrahedral meshes for the far-field foundation, and generate contact element and target element paired meshes for the contact surface between the dam and the water body to obtain a non-uniformly refined finite element mesh model. Process the rockfill body in the non-uniformly refined finite element mesh model using a displacement-pore pressure coupling equation, use the pore water pressure field as the volume force boundary condition, and apply displacement constraints to obtain a finite element solution model that includes complete physical equations and boundary conditions. Input the high-precision three-dimensional tailings pond model as an external dynamic load into the finite element solution model that includes complete physical equations and boundary conditions to obtain a full-order finite element model, and iteratively solve the full-order finite element model to obtain the critical reduction coefficient, the coordinates of the unstable slip surface, and the strain field evolution data. Map the full-order finite element model to a reduced-order model, package it into a callable module, and generate a cloud map of the dam stress-strain field distribution based on the critical reduction coefficient, the coordinates of the unstable slip surface, and the strain field evolution data.

5. A multi-modal tailings dam breach digital twin emergency simulation and decision optimization method according to claim 4, characterized in that: In step 3, based on the PPO reinforcement learning algorithm, using the dam stress-strain field and the current environmental state, predicting the dam-break probability distribution map specifically includes the following steps: Based on the current environmental state and the current action, using the reduced-order model to output the dam stress-strain field distribution contour map, the equivalent plastic strain zone distribution, and the strength reduction coefficient. The current environmental state includes the water level elevation, rainfall intensity, and soil saturation. The current action includes the flood discharge channel operation instruction and the personnel evacuation operation instruction; Taking the magnitude of the strength reduction coefficient as the stability reward, taking the exponential penalty triggered when the area of the equivalent plastic strain zone distribution exceeds the threshold as the risk penalty, and taking the matching result of the flood discharge channel opening state and the personnel evacuation hot zone as the artificial rule reward to construct the reward function to calculate the reward of the reduced-order model output result; Based on the Actor-Critic network architecture, taking the flood discharge channel opening degree, the sub-dam reinforcement strength, and the drainage pump start-stop frequency as the actions, the Actor network generates online actions according to the current environmental state; and converts the reduced-order model output result into the target action for learning. The Critic network is used to predict the expected return according to the online action and takes the reward of the reduced-order model output result as the expected value for learning to guide the generation of the Actor network online action; after learning is completed, the optimal action is output; Superimpose random perturbations on the optimal action for Monte Carlo simulation, count the occurrence frequency of the dam-break event during the Monte Carlo simulation process, and generate a spatio-temporal two-dimensional probability distribution; Interpolate the spatio-temporal two-dimensional probability distribution to the grid nodes of the full-order finite element model, and superimpose the dam stress-strain field distribution contour map for weighted rendering to obtain the dam-break probability distribution heat map.

6. A multimodal digital twin emergency deduction and decision optimization method for tailings pond dam break according to claim 5, characterized in that In step 4, based on the dam-break probability distribution map and the resource distribution data, generating the rescue scheduling plan specifically includes the following steps: Divide the dam-break probability distribution heat map into risk zones of different levels according to the threshold; According to the obtained resource distribution data, obtain the corresponding resource points, and assign corresponding utility weights according to the resource category; Based on the road network topology and real-time traffic conditions, associate the resource points and the risk zones by the shortest path, and mark the estimated travel time to obtain the risk-resource association matrix including the risk level, resource type, path time, and utility weight; Taking minimizing the dam-break probability, minimizing the affected population density, minimizing the economic loss coefficient, minimizing the average arrival time from the resource point to the risk zone, and maximizing the matching degree of the material distribution volume and the demand prediction as the optimization objectives, and taking the resource total limit and the time window limit as the constraint conditions to construct a multi-objective optimization model; Based on the risk-resource association matrix, iteratively optimize the multi-objective optimization model to obtain the Pareto optimal solution set; Calculate the comprehensive score of each solution in the Pareto optimal solution set according to the economic loss, arrival time magnitude, and material distribution utilization rate; Sort the Pareto optimal solution set according to the comprehensive score, divide it into the priority execution plan and the backup plan to obtain the hierarchical rescue scheduling plan.

7. A multimodal digital twin emergency deduction and decision optimization method for tailings reservoir dam break according to claim 6, characterized in that In step 5, a VR scene is built based on the high-precision three-dimensional tailings pond model, and then the dam break probability distribution map and resource distribution data are imported into the VR scene to obtain a dynamic virtual deduction scene, which specifically includes the following steps: Map the hierarchical rescue scheduling plan to the high-precision three-dimensional tailings pond model to generate a planned route; Add mechanical properties to the virtual objects in the high-precision three-dimensional tailings pond model, and align the scene time axes of all terminals based on the cloud synchronization server to obtain a dynamic virtual deduction scene.

8. A multimodal digital twin emergency deduction and decision optimization method for tailings pond dam break according to claim 7, characterized in that, In step 6, in the dynamic virtual deduction scene, use VR devices to conduct drills according to the rescue scheduling plan, and verify and optimize the rescue scheduling plan based on the drill record data to obtain the final rescue scheduling plan, which specifically includes the following steps: Allocate VR devices to each staff member and determine the positions of the staff members in the dynamic virtual deduction scene; In the dynamic virtual deduction scene, generate the optimal path based on the positions of the staff members and the planned route, and generate a highlighted indication arrow; The staff members move forward and perform drill operations according to the highlighted indication arrow, and record the drill data; The commander makes real-time intervention and adjustment to the drill process according to the current drill process and the overall situation of the scheduling, and records the commander's command data; According to the drill data, obtain the VR device operation records, actual resource consumption, actual resource scheduling time, and dam break suppression effect, and assign manual labels to the commander's command data; Construct a multi-dimensional evaluation matrix based on the deviation rate of the actual resource scheduling time from the planned time, the matching of the actual resource consumption with the demand prediction, and the decline rate of the dam break probability after the deduction; Based on the multi-dimensional evaluation matrix, conduct evaluation and analysis through a causal inference engine, classify the executed rescue scheduling plan into a high-quality plan and a low-quality plan, and identify the core operations that affect the results during the drill to obtain the key decision points; Store the high-quality plan in the case knowledge base and use it as the final rescue scheduling plan; Obtain the key decision points of the low-quality plan, and optimize the parameters of the multi-objective optimization model according to the key decision points of the low-quality plan to realize parameter tuning with the low-quality plan as a negative example.

Images