Intelligent Early Warning Method and System for Marine Risks Based on Digital Twin

Abstract

This invention discloses a digital twin-based intelligent early warning method and system for marine risks, relating to the field of intelligent risk early warning. It includes: acquiring multi-source heterogeneous marine data of a target sea area to construct an initial digital twin marine environment model; using an AI proxy model to perform real-time dynamic calibration of the initial digital twin marine environment model, generating a dynamically updated high-fidelity virtual ocean image; and simulating the evolution process of at least one marine risk event based on the high-fidelity virtual ocean image and generating preliminary early warning information. This invention achieves a continuous mutual calibration closed loop between the digital twin and the real marine environment by introducing an adversarial comparison mechanism between actual post-disaster observation data and simulation results, and by automatically triggering model topology reconstruction based on deviation thresholds. This enables the model to possess long-term adaptive evolution capabilities, effectively overcoming the problem of traditional models' accuracy decreasing over time.

Description

Technical Field

[0001] This invention relates to the field of intelligent risk early warning, specifically to a marine risk intelligent early warning method and system based on digital twins. Background Technology

[0002] Intelligent early warning of marine risks (such as storm surges, marine heat waves, shore crashes, and rift currents) is of great significance for coastal disaster prevention and mitigation, marine engineering safety, and maritime navigation assurance. With the rapid development of "AI + Ocean" technology, digital twin technology is gradually being introduced into the field of marine environmental modeling and risk prediction, enabling dynamic mapping and simulation of the real ocean state by constructing a virtual ocean mirror. Existing technologies have already attempted to combine multi-source heterogeneous marine data (such as real-time observation data, remote sensing inversion data, and numerical simulation data) with machine learning models to construct marine digital twin systems for auxiliary early warning of typical marine risk events such as storm surge overtopping simulations and wave anomaly identification.

[0003] Currently, existing marine risk early warning schemes mainly adopt the following technical approaches: First, numerical forecasting methods based on physical models, such as calculating storm surge rise and wave field evolution using ocean dynamics models; second, AI forecasting methods based on pure data-driven approaches, such as using recurrent neural networks or graph neural networks to extrapolate sea surface temperature, ocean currents, and other elements over time; and third, schemes that initially integrate digital twins and AI proxy models, simulating single or a few types of risk events by constructing static or semi-dynamic virtual marine environments. Some advanced schemes also introduce reinforcement learning to adjust model parameters online, or use generative adversarial networks to perform discriminative calibration of the simulated state field, aiming to improve early warning accuracy.

[0004] However, existing technologies still have the following significant shortcomings in practical applications: the structure of most digital twin ocean models is fixed and static in advance, and they cannot autonomously evolve their topology based on the long-term evolution or sudden anomalies of the real ocean environment, resulting in a significant decrease in model accuracy when actual observations and simulation results continue to deviate; existing early warning methods generally lack the ability to infer the causal path of risk propagation, and are prone to misjudging spurious correlation signals as risk triggering factors, outputting false early warning information with high confidence, causing false alarms or missed alarms; the lack of a continuous, adversarial closed-loop mutual calibration mechanism between the virtual model and the real ocean environment makes it difficult for the digital twin to maintain a high degree of consistency with the dynamic characteristics of the real ocean during long-term operation, limiting the long-term reliability and adaptability of the early warning system. Summary of the Invention

[0005] Based on this, the purpose of this invention is to provide a marine risk intelligent early warning method and system based on digital twins, in order to solve the technical problems existing in the current marine risk early warning technology, such as the static structure of digital twin models that cannot evolve autonomously, the tendency to misjudge spurious correlation signals as risk triggering factors leading to false alarms or missed alarms, and the lack of a continuous adversarial closed-loop mutual calibration mechanism between the virtual model and the real marine environment, which makes it difficult to maintain high accuracy and high reliability in long-term operation.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method and system for intelligent early warning of marine risks based on digital twins.

[0007] The first aspect of this application provides a digital twin-based intelligent early warning method for marine risks. The method includes: acquiring multi-source heterogeneous marine data of a target sea area and constructing an initial digital twin marine environment model; using an AI agent model to perform real-time dynamic calibration of the initial digital twin marine environment model to generate a dynamically updated high-fidelity virtual marine image; simulating the evolution process of at least one marine risk event based on the high-fidelity virtual marine image and generating primary early warning information; introducing a causal inference model to invert the risk propagation path of the primary early warning information, identifying key triggering factors and risk transmission chains, and outputting causally verified risk early warning information; and conducting an adversarial comparison between actual post-disaster observation data and the simulation results, driving the structural self-evolution of the digital twin model based on the comparison results to form a continuous mutual calibration closed loop between the digital twin and the real marine environment.

[0008] A second aspect of this application provides a digital twin-based intelligent early warning system for marine risks. The system includes: a data acquisition and initial modeling module for acquiring multi-source heterogeneous marine data and constructing an initial digital twin marine environment model; an AI dynamic calibration module for real-time dynamic calibration and generating a high-fidelity virtual marine image; a risk simulation and primary early warning module for simulating the evolution of marine risk events and outputting primary early warning information; a causality verification module for performing risk propagation path inversion and key trigger factor identification on the primary early warning information, and outputting causally verified risk early warning information; and a virtual-real closed-loop mutual calibration module for adversarially comparing actual post-disaster observation data with simulation results to drive the self-evolution of the digital twin model structure.

[0009] In summary, the present invention has the following main beneficial effects:

[0010] This invention introduces an adversarial comparison mechanism between actual post-disaster observation data and simulation results, and automatically triggers model topology reconstruction based on deviation thresholds. This achieves a continuous mutual calibration loop between the digital twin and the real marine environment, enabling the model to have long-term adaptive evolution capabilities and effectively overcoming the problem of traditional models' accuracy decreasing over time. Addressing the shortcomings of existing technologies that easily misjudge spurious correlation signals as risk triggering factors, leading to false alarms or missed alarms, this invention innovatively introduces a structural causal model and counterfactual reasoning to invert risk propagation paths and verify causal consistency of primary early warning information. This accurately identifies dominant risk factors, eliminates spurious correlation early warning signals, and outputs early warning results with clear causal relationships and quantified uncertainty, significantly improving the accuracy and interpretability of early warnings. Furthermore, by establishing a confidence arbitration mechanism between the AI ​​proxy model and the causal inference model, this invention forcibly triggers local high-resolution resimulation when causal consistency is insufficient. Combined with an automatic switching strategy for ensemble forecast models, this significantly enhances the system's robustness and reliability against complex marine risk events, thus providing a more scientific and reliable intelligent early warning method for coastal disaster prevention and mitigation and marine engineering safety. Attached Figure Description

[0011] Figure 1 This is the control flowchart of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] The embodiments of the present invention will now be described.

[0014] A digital twin-based intelligent early warning method for marine risks, such as Figure 1 As shown, including,

[0015] I. Overall Methodology and Flow

[0016] Multi-source heterogeneous ocean data is acquired to construct an initial digital twin ocean environment model. In this embodiment, the AI ​​proxy model (GNN+VAE) is used to simulate or approximate the dynamic behavior of the initial digital twin ocean environment model to achieve rapid extrapolation. A dynamic calibration module composed of Ensemble Kalman Filtering (EnKF) and Reinforcement Learning (DDPG) is used to dynamically calibrate key parameters of the AI ​​proxy model and the underlying physical model in real time based on real-time observation data, thereby generating a dynamically updated high-fidelity virtual ocean image. The evolution of ocean risk events is simulated and primary early warning information is generated. A causal inference model is introduced to invert risk propagation paths and verify causality. Actual post-disaster observation data and simulation results are compared adversarially to drive the self-evolution of the digital twin model structure, forming a closed-loop mutual calibration between the virtual and real worlds.

[0017] II. Data Acquisition and Initial Modeling

[0018] Real-time observation data:

[0019] Six seabed observation base stations were deployed, located at water depths of 10 meters, 12 meters, 15 meters, 20 meters, 25 meters, and 30 meters, respectively. Each base station is equipped with: a pressure-type tide gauge (model: Sea-BirdSBE26plus), with a sampling frequency of 1Hz and a pressure measurement accuracy of ±0.05% of full scale, which translates to a tide height accuracy of ±0.01 meters (m); and an acoustic Doppler current profiler (ADCP, model: RDIWorkHorse), with a sampling frequency of 1Hz and a current velocity measurement accuracy of ±0.5% of the measured value, which is ±0.005 meters per second (m / s).

[0020] Remote sensing inversion data:

[0021] Data is sourced from the Sentinel-3 satellite, with a transit cycle of approximately 3 days and a spatial resolution of 300 meters. Sea surface temperature (SST, in degrees Celsius, °C) and sea surface height anomaly (SLA, in meters, m) are extracted. Historical reanalysis data is derived from the FES2014 global tidal and storm surge model database, spanning from 1990 to 2022, with a spatial resolution of 1 / 16° (approximately 6.96 km) and a temporal resolution of 1 hour. Extracted parameters include storm surge rise (in meters, m). Ocean model simulation data is generated by the locally running FVCOM (Unstructured Grid Finite Volume Ocean Model), with a nearshore grid resolution of 50 meters and an offshore resolution of 500 meters, and a time step of 2 seconds. Data transmission and synchronization strategy: Real-time observation data is transmitted to the central server every 600 seconds (10 minutes) via a 4G network; remote sensing and reanalysis data are synchronized daily at 00:00 (UTC+8).

[0022] Construction of the initial digital twin marine environment model:

[0023] Define a spatially discretized grid for the target sea area. An unstructured triangular grid is used, with a total number of elements in the horizontal direction. Each is vertically divided into The governing equations of the model for the sigma coordinate layer are as follows.

[0024] Continuity equation (mass conservation):

[0025] ;

[0026] in, The height of the free surface relative to the mean sea level (in meters, m), at the initial moment. It is obtained by subtracting the local mean sea level from the measured tide level.

[0027] Time (unit: seconds, s).

[0028] and The vertical average flow velocities (in meters per second, m / s) in the x and y directions are given by the historical average flow field.

[0029] Total water depth (unit: meters, m). ,in The static water depth (unit: meters, m) is obtained by interpolation from multibeam echo sounding data.

[0030] The momentum equation for direction:

[0031] ;

[0032] in, Coriolis parameter (unit: Hertz, Hz). Earth's rotational angular velocity. Radius per second, average latitude of the target sea area ,

[0033] The acceleration due to gravity (unit: meters per second squared, m / s²) is expressed using international standard values.

[0034] For sea surface wind stress Directional component (unit: N / m²). The calculation formula is: ,in, Kilograms per cubic meter is the standard atmospheric density. This is the drag coefficient (dimensionless), applicable to situations where the wind speed at 10 meters is less than 25 m / s; The wind speed at a height of 10 meters above the sea surface (unit: meters per second, m / s) was obtained from the on-site meteorological station or reanalysis data. For wind speed at Components in direction, for example when (pure easterly wind), then .

[0035] (where is the unit of seawater density: kilograms per cubic meter, kg / m³), and represents a typical value for nearshore areas.

[0036] For the bottom frictional stress at Component in direction (unit: N / m²). The calculation formula is: ,in, The coefficient of friction at the bottom (dimensionless), an empirical value. Let be the near-bottom current velocity (unit: meters per second, m / s) at a height of 1 meter above the seabed, calculated and output by the model. For example, when ,but .

[0037] is the horizontal eddy viscosity coefficient (unit: square meters / second, m² / s), estimated using the Smagorinsky parameterization.

[0038] III. AI Agent Model and Real-time Dynamic Calibration

[0039] The architecture of the AI ​​agent model is a fusion of graph neural network (GNN) and variational autoencoder (VAE).

[0040] Graph neural network unit: treats the grid as a graph , It is a set of nodes (125,680 nodes in total). This is the set of edges connecting adjacent nodes. The feature vector of each node i. include Dimension .

[0041] The graph convolution operation is defined as:

[0042] ;

[0043] in, Let i be the hidden state (dimensionless) of node i at level l.

[0044] Let i be the set of neighboring nodes of node i.

[0045] Let i be the degree of node i (number of neighbors, dimensionless).

[0046] The learnable weight matrix of the l-th layer (dimensions) The ADAM optimizer (learning rate 0.001) was applied to historical marine data (30 years of data from 1990 to 2020, with a time resolution of 1 hour and a total of approximately 262,800 samples). , (Training)

[0047] () is the ReLU activation function.

[0048] Variational autoencoder unit: converts the spatial feature sequence output by the GNN into a single unit. (Dimension 128) Input VAE, encoder output latent variable mean and variance The loss function is:

[0049] ;

[0050] in, The input is a multi-source heterogeneous data matrix.

[0051] These are latent variables (32 dimensions, dimensionless).

[0052] and The distribution of the encoder and decoder.

[0053] It is a standard normal prior.

[0054] Let KL divergence be a dimensionless value.

[0055] The VAE training data is the same as before, with 200 training rounds and a batch size of 64.

[0056] Real-time dynamic calibration and high-fidelity virtual ocean image generation: Dynamic calibration employs a rolling assimilation method with a time window of 1 hour (3600 seconds). Let the k-th assimilation period be... Observation vector Ensemble Kalman Filtering (EnKF) is used, and the number of set members is... Analyzing the field calculations:

[0057] ;

[0058] in, For observation operators (mapping model states to observation locations, with units depending on variable type, e.g., tide level is mapped to unit 1).

[0059] The Mann gain matrix is ​​calculated using the following formula:

[0060] ;

[0061] in, The forecast error covariance matrix (units vary with the variable, for example, the tide level covariance is in square meters, m²) is calculated from the set members.

[0062] Let's define the observation error covariance matrix (with the same units). For example, if the standard deviation of the tide observation error is 0.05 meters, then... square meters.

[0063] The dynamic calibration process is performed every 10 minutes (600 seconds), and the resulting new state field forms a high-fidelity virtual ocean mirror.

[0064] Reinforcement learning optimization parameter update strategy: Defining the state space of reinforcement learning The mean of the current model prediction error (Unit: meters) and variance (Unit: square meters). Movement space Adjusting the turbulence closure coefficient (Dimensionless, default value 0.09, adjustment range 0.05~0.15) and bottom friction coefficient (Dimensionless, default 0.0025, adjustment range 0.001~0.005). Reward function: Where MAE is the mean absolute error (in meters) and RMSE is the root mean square error (in meters).

[0065] Deep Deterministic Policy Gradient (DDPG) was used. Both the Actor and Critic networks were 3 fully connected layers (256–128–64 neurons). Training was conducted online for 1000 time steps (10 minutes per step, approximately 7 days in total), with a learning rate of 0.0001 for Actor and 0.001 for Critic. The optimized parameters were updated every 6 hours.

[0066] IV. Risk Simulation and Initial Early Warning

[0067] Based on a high-fidelity virtual ocean mirror, scenario simulations were conducted to address the risk of storm surge breaching. A meteorological forcing field was set: the typhoon's path passed through the target sea area, with a central pressure of 950 hPa, a maximum wind speed radius of 50 km, and a center movement speed of 5 m / s.

[0068] Hydrodynamic response calculation: The same momentum and continuity equations as the initial model are used, and the boundary conditions are given by meteorological forcing. The maximum water level rise at the grid points is calculated. (Unit: meters), of which For grid indexing. Extract the normal profile of the nearshore protective dike location (linear, 12.3 km in length) to obtain the overtopping height:

[0069] ;

[0070] in, The coordinates of the arc length along the direction of the embankment (unit: meters).

[0071] The elevation of the levee crest (in meters) is given by LiDAR measurement data.

[0072] This represents the maximum water level at the corresponding point (unit: meters).

[0073] The total simulation time is 72 hours (259,200 seconds), and the time step is... Seconds. Calculate and output the water level field and flow velocity field at each time step.

[0074] Generate preliminary early warning information: Risk level: based on maximum overtopping height. The risk levels are divided into: no risk (0~0.1 meters), low risk (0.1~0.3 meters), medium risk (0.3~0.8 meters), and high risk (>0.8 meters).

[0075] Area affected: The total length of a continuous area where the overtopping height is greater than 0 meters.

[0076] Expected time of occurrence: from the onset of typhoon impact to the first appearance of the typhoon The time interval.

[0077] The purpose of this step is to output preliminary early warning information.

[0078] V. Causal Inference Models and Causal Verification

[0079] Structural Causal Models and Counterfactual Reasoning: Defining the Set of Causal Graph Nodes ,in, Pressure difference at the center of the typhoon (Unit: hectopascals) Standard atmospheric pressure is taken as 1013 hPa; This refers to the lowest atmospheric pressure during a typhoon (unit: hectopascals). For example, when... When the pressure is 950 hPa, Hectopascals.

[0080] The local wind speed is W (unit: meters per second, m / s), obtained from the weather station.

[0081] Astronomical tide level (Unit: meters, m), obtained through harmonic analysis forecasting.

[0082] Storm surge increase B (unit: meters, m).

[0083] Structural equations: , , , .

[0084] in, , , , All are causal coefficients, estimated using linear regression on a historical dataset (60 historical typhoon events). The unit is m / hPa. Dimensionless (Dimensionless).

[0085] , , , This is an independent noise term (unit dependent).

[0086] Counterfactual reasoning: setting up actual events Wind speed Counterfactual question: If , 08, 2, , ,but How much has changed? Calculate:

[0087] ;

[0088] ;

[0089] Pick ,but , , Marginal causality ,and Consistent.

[0090] Identifying dominant risk factors and eliminating spurious correlations: Calculating the causal contribution rate of each factor:

[0091] ;

[0092] in For example, the difference of standard variables. , , Substituting into , , ,sum Rounded to three decimal places, the contribution rates are 31.1%, 19.4%, and 49.5% respectively. Therefore, the dominant factor is... That is the astronomical tide level.

[0093] Causality verification and confidence arbitration: Defining causal consistency score:

[0094] ;

[0095] in: The score is based on causal consistency.

[0096] This represents the number of times the causal path is sampled.

[0097] The water increase (unit: meters) simulated by the causal model.

[0098] Water level increase for actual observation or digital twin simulation (unit: meters).

[0099] The standard deviation of the uncertainty in water increase (unit: meter) is set to 0.1 meters.

[0100] Set threshold ,like The system automatically triggers a local high-resolution re-simulation mechanism, with the following specific steps:

[0101] (1) Automatic identification of re-simulation regions

[0102] Spatial range determination: Based on the spatial location of nodes with causal consistency scores below the threshold in the causal graph, the contribution gradient of each grid cell to causal inconsistency is calculated through backpropagation of the graph neural network, and the continuous region where the grid cells with the top 5% absolute gradient values ​​are located is determined as the resimulation focus area.

[0103] Region expansion: Based on the outer rectangle of the focal region, expand outwards by a distance equal to twice the diameter of the grid cells in each direction to form the final resimulation subregion. For example, if the focal region spans 2 km × 1.5 km horizontally, the resimulation subregion will expand to 6 km × 5.5 km.

[0104] Time window determination: Starting from the time point t0 when the causal inconsistency first occurs, the time window extends forward to the end of the risk event or no more than 24 hours later, and backward by 1 hour as a buffer period, forming the resimulation time window. .

[0105] (2) Local mesh encryption strategy

[0106] Spatial resolution enhancement: Within the re-simulated sub-region, the side lengths of the original unstructured triangular mesh are refined to 1 / 5 to 1 / 10 of the original length. Specifically:

[0107] The original near-shore grid resolution was 50 meters, which was reduced to 5-10 meters after densification.

[0108] The original offshore grid resolution was 500 meters, which was reduced to 50-100 meters after densification;

[0109] The encrypted transition zone is set with three layers of gradient mesh, with the side length of each layer increasing by a factor of 1.5 to ensure numerical stability.

[0110] Vertical layer densification: In the resimulated sub-region, the original 10 sigma coordinate layers are increased to 20 layers, with an additional 5 layers densified in the 0-5 meter water layer below the free surface (layer thickness is distributed exponentially, with the surface layer being the thinnest) to more finely depict the storm surge and wave breaking process.

[0111] Time step adjustment: based on the minimum mesh size after encryption. The time step is adaptively adjusted according to the CFL condition:

[0112] ;

[0113] in Take a value of 0.5 to 0.8. This represents the minimum side length of the encrypted grid (in meters). The acceleration due to gravity is 9.81 m / s². This represents the maximum water depth (in meters) within the sub-region. Typically, when rice, Rice time, It takes approximately 0.5 to 0.8 seconds, which is significantly more encrypted than the original 2-second step size.

[0114] (3) Boundary conditions and initial field treatment

[0115] Boundary condition transfer: The output variables (water level, flow velocity, temperature, salinity, etc.) of the original coarse mesh model at the boundary of the resimulated sub-region are used as boundary conditions of the resimulated sub-model through linear interpolation or inverse distance weighted interpolation, ensuring physical consistency between the sub-model and the global model.

[0116] Initial field nesting: All state variables of the resimulation sub-region at time t0-1h are extracted from the original model output, interpolated and mapped to the refined mesh, serving as the initial conditions for the resimulation. For high-frequency components generated by interpolation, a 3-point smoothing filter is used to eliminate numerical noise.

[0117] Atmospheric forcing field refinement: If meteorological forcing (such as typhoon wind field) exists in the resimulation area, the original coarse-resolution forcing field is replaced with high-resolution reanalysis data (spatial resolution 1~3 km, temporal resolution 10 minutes), and re-interpolated according to the densified grid nodes.

[0118] When causal confidence ,in Let be the standard normal cumulative distribution function, when When the system automatically switches to ensemble forecast mode, it generates 30 ensemble members (using different turbulence parameters and initial disturbances) and outputs the member mean and standard deviation as the final warning.

[0119] (4) Re-simulation execution and result fusion

[0120] Independent sub-model calculation: Under a refined mesh and time step, the same governing equations (continuity equation and momentum equation) and physical parameterization scheme as the original model are used to independently perform re-simulation calculations, with a duration not exceeding 24 hours.

[0121] Resulting interpolation and smooth stitching:

[0122] The calculation results (water level, flow velocity, etc.) on the refined grid of the resimulated sub-region are mapped back to the original grid resolution through area-weighted averaging or inverse distance interpolation;

[0123] Within the transition zone extending 3-5 mesh layers outward from the boundary of the resimulated sub-region, the following fusion formula is used to achieve smooth stitching:

[0124] ;

[0125] in The weighting coefficients within the transition zone increase linearly from 0 at the boundary to 1 at the inner boundary; It represents any state variable (such as water level).

[0126] Consistency check: After the resimulation is completed, the causal consistency score C is recalculated. If C is still greater than the threshold, the above encryption and resimulation process is repeated (no more than 3 times); if it still fails after 3 times, the current result is output along with a large uncertainty interval (twice the standard deviation) as a warning.

[0127] (5) Computational resource scheduling guarantee

[0128] The system reserves 20% of computing resources (GPU / CPU core count) to respond to local re-simulation trigger requests;

[0129] The computation time for a single local resimulation must not exceed 30% of the original model simulation time; otherwise, the encryption rate will be automatically reduced (from 1 / 5 to 1 / 3) to ensure the timeliness of the early warning.

[0130] If the resimulation process is interrupted or times out, it will automatically revert to the original coarse-grid simulation results and mark "Local resimulation not completed, confidence level reduced" in the warning message.

[0131] VI. Virtual-Real Closed-Loop Mutual Verification and Model Self-Evolution

[0132] Adversarial comparison: Generative Adversarial Networks (GANs) are used, where: the generator Input the current state field of the digital twin model (dimension) Output corrected state field .

[0133] Discriminator Input the actual observation field Alternatively, the generator outputs the discrimination probability (0~1).

[0134] Training loss:

[0135] ;

[0136] ;

[0137] in, The dimensionless equilibrium coefficient, The L2 norm squared (unit: squared variable). Both the generator and discriminator are 5-layer convolutional neural networks (64, 128, 256, 256, 128 filters), and the training data consists of 1000 matching time points (the entire field at each time point), using the ADAM optimizer with a learning rate of 0.0002.

[0138] Model structure self-evolution: Define a sustained deviation threshold: Over 72 consecutive time windows (each window is 1 hour, i.e., 3 days), the average error is: If the error exceeds a certain threshold in meters, and at least 50 windows have errors exceeding this threshold, then structural self-evolution is triggered.

[0139] Adding a layer: Add a fifth layer after the existing four layers of the GNN. The new layer has 256 neurons, using small random numbers. Initialize weights.

[0140] Pruned layer: If the average activation level of a certain layer is... If the training continues for more than 200 steps, the layer is deleted.

[0141] Shared Case Library: Case Library Storage Structure: After each typhoon event, a storage vector is created.

[0142] ;

[0143] New events and existing events are matched using Euclidean distance (variables are first standardized by Z-score, dimensionless), and the three most similar events are selected, with their average warning deviation being calculated. The initial warning is then corrected. For example, if the standardized deviations of a new event from events 1, 2, and 3 in the database are +0.12m, -0.03m, and +0.03m respectively, then... Add rice to the initial warning.

[0144] VII. Generation of Risk Transmission Map

[0145] With directed weighted graph express:

[0146] The nodes include "typhoon low pressure", "increased wind stress", "water level rise", "overtopping", and "expansion of inundation area".

[0147] : Directed edge, such as "typhoon low pressure" → "increased wind stress".

[0148] Edge weights, for example, causal strength. Value (after dimensionless processing) or time delay (unit: seconds).

[0149] To enable interactive queries between users and the visualized map, the system employs a WebGL-based event listening and data binding mechanism in the front-end graphical user interface. The specific implementation is as follows:

[0150] First, in generating a directed weighted graph At that time, the system assigns a unique identifier to each node and each directed edge, and structurally binds this identifier with the following attributes: the name of the physical quantity corresponding to the node or edge, the causal strength coefficient, the time lag, the risk contribution rate ranking, and the associated original data index. The above binding information is stored in the front-end graphical data structure, forming a "visual element-data" mapping table.

[0151] Secondly, in the front-end rendering engine, each node is rendered as an interactive vector graphic object (such as a circle or rounded rectangle), and a mouse click event listener function is registered. When a user clicks a node with a mouse or touchscreen, the event listener function captures the unique identifier of the clicked object and then retrieves the corresponding attribute data in the mapping table. Based on the retrieval results, the system displays the following information in structured text or table format in the sidebar, floating tooltip, or modal dialog box of the current interface:

[0152] The node represents the name of the risk factor or environmental state.

[0153] This node serves as the location in the risk transmission chain (initial triggering factor, intermediate state, final area of ​​impact).

[0154] The causal strength coefficient of this node to the next level of risk;

[0155] The ranking of the dominant factors in the overall risk transmission chain of this node (such as "major contributing factor" or "minor contributing factor").

[0156] If the node has a time lag attribute, the average lag time (in minutes or hours) relative to the previous node will also be displayed.

[0157] In addition, the directed edge system also supports click query: when a user clicks on a directed edge connecting two nodes, an information window pops up on the interface, displaying the name of the causal path represented by the directed edge, the causal strength coefficient and the time lag, and optionally displaying an example of the counterfactual reasoning result corresponding to the edge.

[0158] To enhance the intuitiveness and user understanding of the interaction, the system synchronously changes the visual style of a node when it is clicked (e.g., adding a highlight halo, enlarging the node size, or darkening the color), and highlights all downstream paths originating from that node in the graph, while also indicating the direction of risk propagation with dynamic arrows. The calculation and rendering of this interactive response process are entirely completed on the front end, with a response time of no more than 100 milliseconds, eliminating the need for secondary requests to the server and ensuring the real-time performance and smoothness of user queries.

[0159] VIII. System Integration and Operation Examples

[0160] Integrate the above modules into a distributed system:

[0161] Data acquisition module: Based on the Kafka stream processing platform, latency ≤ 2 seconds.

[0162] Model building module: runs on 4 NVIDIA A100 GPUs (40GB of video memory each).

[0163] AI dynamic calibration module: executes once every 10 minutes, taking ≤3 minutes.

[0164] Risk simulation module: A 72-hour simulation takes about 45 minutes (125,680 grid cells, 2-second step).

[0165] Causality verification module: The time taken for a single causal inference is ≤2 seconds.

[0166] Virtual-Real Closed-Loop Mutual Calibration Module: GAN retraining is triggered once every 24 hours and lasts for 1 hour.

[0167] Feedback learning module: Actual observation data (such as actual tide gauge measurements) are entered into the database within 12 hours after the typhoon, compared with simulation results, and the parameters of the digital twin model are updated.

[0168] Execution Example: An analogy to the impact of Typhoon Haiyan on a target sea area. The system was launched 48 hours before the typhoon's landfall. The initial digital twin model was based on the UTC status at 00:00 on August 15th. After acquiring multi-source data, the AI ​​proxy model completed dynamic calibration. The risk simulation module output a preliminary warning: maximum overtopping height 0.65 meters, medium risk. Causal inference model calculations... The reading was below the threshold of 1.2, triggering a localized re-simulation with enhanced resolution. After correction, the maximum overtopping height was 0.71 meters. The final warning was issued at 06:00 UTC on August 17th, with an actual observed overtopping height of 0.69 meters and an error of 0.02 meters. The causal confidence level was... After the typhoon, adversarial comparison inputs real data and simulation results into the GAN, reducing the discriminator loss from 0.45 to 0.12, updating the model parameters, and adding the event record to the event library.

[0169] The steps of the methods or algorithms described in this application can be directly embedded in hardware, software units executed by a processor, or a combination of both. Exemplarily, a storage medium can be connected to a processor so that the processor can read information from and write information to the storage medium. Optionally, the storage medium can also be integrated into the processor. Optionally, the processor and the storage medium can also be located in different components within a terminal. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0170] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative examples of this application and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A digital twin-based intelligent early warning method for marine risks, characterized in that, include: Acquire multi-source heterogeneous marine data of the target sea area and construct an initial digital twin marine environment model; The initial digital twin marine environment model is dynamically calibrated in real time using an AI agent model to generate a dynamically updated virtual ocean image. Based on the virtual ocean mirror, the evolution process of at least one marine risk event is simulated and a primary early warning message is generated; A causal inference model is introduced to invert the risk propagation path of the primary early warning information, identify key triggering factors and risk transmission chains, and output risk early warning information that has been causally verified. The actual post-disaster observation data is compared adversarially with the simulation results. Based on the comparison results, the structure of the digital twin model is driven to evolve, forming a continuous mutual calibration closed loop between the digital twin and the real marine environment.

2. The intelligent early warning method for marine risks based on digital twins according to claim 1, characterized in that: The multi-source heterogeneous ocean data includes at least: real-time observation data acquired by seabed observation base stations, remote sensing inversion data acquired by remote sensing satellites, reanalysis data acquired by historical reanalysis databases, and simulation data generated by ocean numerical models; wherein, the real-time observation data includes tidal level data and current velocity data, the remote sensing inversion data includes sea surface temperature and sea surface height anomalies, the reanalysis data includes storm surge data, and the simulation data is generated by an unstructured grid finite volume ocean model.

3. The intelligent early warning method for marine risks based on digital twins according to claim 1, characterized in that: The causal inference model employs a structural causal model combined with graph-based counterfactual reasoning to identify dominant risk factors and eliminate spurious correlation warning signals in the following manner: Construct a structural causal graph containing multiple environmental factor nodes and risk event nodes, and assign structural equations to the causal relationships between each node; The marginal causal effects of various environmental factors on the target marine risk are estimated based on historical typhoon event datasets. Calculate the causal contribution rate for each environmental factor, which is the proportion of the sum of the marginal causal effects of the factor and its standard deviation among all factors. Identify at least one environmental factor with the highest causal contribution rate as the dominant risk factor; Warning signals that do not have a causal path verified by a structural causal model and only have statistical correlation are judged as spurious correlation warning signals and removed from the warning information.

4. The intelligent early warning method for marine risks based on digital twins according to claim 1, characterized in that, Also includes: A visualized risk propagation map of the risk transmission chain is generated. The map, in the form of a directed weighted graph, dynamically displays the complete spatiotemporal propagation path of risk from at least one initial triggering factor node, through at least one intermediate environmental state node, to at least one final affected area node. The weights of the directed edges between nodes include at least a causal strength coefficient or a time lag. The map dynamically displays the attenuation or amplification process of risk intensity at each node, and the attenuation or amplification process is visualized through changes in node color depth or node size.

5. The intelligent early warning method for marine risks based on digital twins according to claim 1, characterized in that: The AI ​​agent model adopts an architecture that integrates graph neural networks and variational autoencoders; An architecture that integrates graph neural networks (GNNs) and variational autoencoders (VAEs) is adopted. Definition of graph convolution operation: ； in, Let i be the hidden state (dimensionless) of node i at level l. Let i be the set of neighboring nodes. Let i be the degree of node i (number of neighbors, dimensionless). Let be the learnable weight matrix of the l-th layer. ( ) represents the ReLU activation function; The loss function of the variational autoencoder unit is: ； in, Given a multi-source heterogeneous data matrix as input, These are latent variables (32 dimensions, dimensionless). and For the distribution of encoders and decoders, For the standard normal prior, Let KL divergence be a dimensionless property. The AI ​​agent model and the causal inference model are equipped with a confidence arbitration mechanism: when the causal inference model determines that the causal consistency score of the current primary warning information is lower than a preset threshold, the digital twin model is forcibly triggered to perform a local high-resolution resimulation in the corresponding spatiotemporal region.

6. A marine risk intelligent early warning system based on digital twins, characterized in that, include: The data acquisition and initial modeling module is used to acquire multi-source heterogeneous marine data and build an initial digital twin marine environment model; The AI ​​dynamic calibration module is used for real-time dynamic calibration and generation of virtual ocean images; The risk simulation and primary early warning module is used to simulate the evolution of marine risk events and output primary early warning information; The causality verification module is used to perform risk propagation path inversion and key trigger factor identification on the primary early warning information, and output risk early warning information that has been verified by causality. The virtual-real closed-loop mutual calibration module is used to conduct adversarial comparisons between actual post-disaster observation data and simulation results, driving the self-evolution of the digital twin model structure.

7. A marine risk intelligent early warning system based on digital twins according to claim 6, characterized in that: The virtual-real closed-loop mutual calibration module includes a generative adversarial network, in which the generator is used to output the corrected digital twin state field, and the discriminator is used to distinguish between the simulated state field and the real observed state field. Through adversarial training, the digital twin model is forced to approximate the real ocean dynamics characteristics.

8. A marine risk intelligent early warning system based on digital twins according to claim 6, characterized in that: The causality verification module is also used to output the uncertainty range and causal confidence of each early warning event, and automatically switch to ensemble forecast mode when the causal confidence is lower than the lower limit, so as to replace the single deterministic early warning with the multi-member forecast results.

9. A marine risk intelligent early warning system based on digital twins according to claim 6, characterized in that: It also includes a risk propagation graph generation module, which is used to dynamically draw the complete propagation link of risk from triggering to disaster in the form of a directed weighted graph.

10. A marine risk intelligent early warning system based on digital twins according to claim 6, characterized in that: The AI ​​dynamic calibration module and the causal verification module share a case library. The case library stores the causal chain structure and digital twin model response characteristics of typical historical risk events, which are used to quickly compare and match new events and correct initial warning deviations.

Images