Intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion

Abstract

The application discloses a method for intelligent identification of submerged oil pollution based on multi-source heterogeneous data fusion, comprising collecting multi-source monitoring data and meteorological data of a submerged oil area, and preprocessing the multi-source monitoring data and the meteorological data; adopting a stereoscopic perception system to perform space-time registration optimization on the multi-source monitoring data and the meteorological data to obtain spatial alignment data, and performing deep multi-modal fusion on the spatial alignment data to obtain fusion data; adopting a graph neural network to simulate oil cluster migration along a flow based on the fusion data to obtain a simulated oil cluster migration path, and performing oil cluster migration game analysis based on an actual oil cluster migration path and the simulated oil cluster migration path to obtain an oil cluster movement migration rule; and constructing an intelligent identification model for submerged oil pollution based on physical information embedding based on the oil cluster movement migration rule, inputting to-be-identified data into the intelligent identification model for submerged oil pollution, and outputting an identification result.

Description

Technical Field

[0001] This invention relates to the field of intelligent identification technology for submerged oil pollution, and in particular to an intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion. Background Technology

[0002] Submerged oil pollution poses a serious threat to marine ecosystems, fishery resources, and coastal economies. Accurately identifying the extent and migration trend of this pollution is crucial for pollution control. Currently, the identification of submerged oil pollution faces multiple challenges: the marine environment is complex and variable, oil blobs are easily affected by ocean currents, water temperature, and other factors, and their submerged nature makes it difficult for traditional monitoring methods to fully capture their spatial distribution; multi-source monitoring data (such as remote sensing images, sensor data, and ship trajectories) are heterogeneous, with inconsistent spatiotemporal benchmarks and significant differences in data modes, making direct integration and utilization difficult; existing identification methods mostly rely on a single data source or static model, lacking a deep characterization of the dynamic migration patterns of oil blobs, resulting in insufficient identification accuracy and real-time performance.

[0003] With the increase in marine development activities, the risk of oil spills continues to rise, making the need for rapid and accurate identification of submerged oil pollution increasingly urgent. Developing identification methods that can integrate multi-source heterogeneous data and fuse physical mechanisms with intelligent algorithms to address issues such as poor data synergy and weak dynamic prediction capabilities has become an important topic in the field of marine environmental protection, and is of significant practical importance for improving the efficiency of pollution emergency response and reducing ecological and environmental losses. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Multi-source monitoring data and meteorological data of the submerged oil area are collected, and the multi-source monitoring data and meteorological data are preprocessed; the multi-source monitoring data includes optical remote sensing images, SAR remote sensing images, sensor data, AIS ship trajectory data, marine numerical model data, hydrological data and the actual migration path of the oil clump; A three-dimensional sensing system is used to perform spatiotemporal registration optimization on the multi-source monitoring data and the meteorological data to obtain spatially aligned data, and deep multimodal fusion is performed on the spatially aligned data to obtain fused data; A graph neural network is used to simulate the migration of oil masses along the flow in the fused data to obtain the simulated oil mass migration path. Based on the actual oil mass migration path and the simulated oil mass migration path, a game analysis of oil mass migration is performed to obtain the oil mass movement and migration law. Based on the movement and migration patterns of the oil clumps, a smart identification model for submerged oil pollution based on physical information embedding is constructed. The data to be identified is input into the smart identification model for submerged oil pollution, and the identification results are output.

[0006] Furthermore, a method for spatiotemporal registration and optimization of the multi-source monitoring data and the meteorological data using a three-dimensional sensing system includes: The original timestamps of each sensor are collected and converted to UTC standard time. For missing timestamp data, linear interpolation is used to complete them. Initial time synchronization is achieved using Kalman filtering, and a time synchronization error state vector is defined. State transition matrix ,in Here is the state transition matrix. The sampling interval is... Due to time deviation, For transpose, The process noise at time k, Let k be the state vector at time k-1; Particle swarm optimization error correction is employed, and an observation vector is constructed based on the GPS second pulse signal. The time deviation is corrected iteratively using Kalman filtering, expressed as: ; in The observation noise at time k, Let k be the time deviation at time k. The Kalman gain at time k, For the observation matrix, Let k be the observation vector at time k. Let be the optimal state vector at time k. Let be the optimal state vector at time k-1. This is a predicted value for the current state. Minimizing the time synchronization error is transformed into an optimization problem, expressed as: ; in Let be the reference time for the i-th sensor. Let be the original timestamp of the i-th sensor. The number of sensors; The particle swarm is initialized with the fitness function being the sum of squared time deviations. The particle velocity and position are iteratively updated until convergence. Based on the optimized time deviations, the sonar, electromagnetic, and optical data are timestamped to control the time difference between different modes of data within the same monitoring area to within 5 seconds. Unify multi-source data into the WGS84 coordinate system to achieve spatial consistency of air, space, sea and underwater three-dimensional data, and achieve preliminary alignment of data at different layers based on a common geographic reference using the least squares method. Introducing prior knowledge of the ocean dynamic field to construct a spatial transformation model, the expression is: ; in Rotation matrix It is a translation vector. This represents the spatial offset caused by the flow velocity field. For the spatial coordinates of the target image, These are the corresponding spatial coordinates of the source image. Outliers are removed by random sampling consensus algorithm. Based on the SLAM 3D reconstruction results of AUV, the oil clump contour of underwater multispectral image, the oil droplet scattering characteristics of forward-looking sonar, and the oil fingerprint sampling points of seabed observation network are integrated into the same 3D grid to form a three-dimensional representation of the spatial distribution of submerged oil and output as spatially aligned data.

[0007] Furthermore, the method for obtaining fused data by performing deep multimodal fusion on the spatially aligned data includes: An improved U-Net architecture is used to extract image features from optical remote sensing images, SAR images, and AUV underwater multispectral images: multi-scale textures are captured through an Encoder-Decoder structure to obtain feature maps; and global average pooling is used to compress the feature maps into image vectors. The time-series data of the smart buoy sensor and the sonar echo sequence are learned by BiLSTM network to learn the time-series dependencies: forward LSTM captures historical trends, backward LSTM captures future associations, outputs hidden state sequences, and dynamically weights time step features through self-attention mechanism to output time-series vectors. A trajectory graph structure is constructed based on AIS ship trajectory data: ship position points are used as nodes and temporal adjacency relationships are used as edges to form a ship trajectory graph. A graph attention network is used to calculate the attention weights between nodes, aggregate neighborhood information, and output trajectory vectors. Based on the velocity field and hybrid layer depth data output by the ROMS / HYCOM model, the physical parameters are encoded into a motion constraint graph: nodes represent spatial grid cells, edge weights are velocity gradients, hybrid layer depth is used as node features, graph convolutional networks extract dynamic field features, and output constraint vectors. The image vector, time series vector, trajectory vector, and constraint vector are concatenated into a modal query matrix. Calculate the intermodal similarity matrix : ; in For image vectors, For time series vectors, For trajectory vectors, For constraint vectors, For encoding dimensions, The modal number; Using the constraint vector as the guiding vector, the modal weights are adjusted through a gating mechanism, as expressed in the following expression: ; in Let a be the weight of the a-th mode. Let a be the learnable parameter. It is the sigmoid activation function; The modal features are weighted and summed according to their modal weights to generate a fused feature matrix. The original modal features are then added to the fused features to obtain the deep fused feature matrix, expressed as follows: ; in For deep fusion of feature matrices, For fusion feature matrix; The deep fusion features are mapped into a fusion vector through a fully connected layer. The fusion vector includes spatial features, temporal features, and dynamic features. The spatial features are the image texture and location coordinates of the suspected oil blobs. The temporal features are the changing trends of environmental parameters monitored by the sensors. The dynamic features are the probability distribution of oil blobs migration under the constraints of the ocean dynamic field.

[0008] Furthermore, a method for simulating the migration path of oil patches by using a graph neural network to simulate the flow migration of oil patches in the fused data includes: The continuous oil mass is discretized into N oil mass units. Each unit is regarded as an independent moving entity and is denoted as a set. The unit division is based on the initial shape of the oil mass, and the size range is 5-50m. Using oil patch units as nodes, the feature vector of each node contains physical properties and environmental interactions; physical properties include location coordinates, density, viscosity, and initial concentration; environmental interactions include current flow velocity, water temperature, and salinity. An undirected graph is constructed based on spatial proximity, where edges exist between node pairs whose distance is less than or equal to the radius of influence of the spread; edge features include Euclidean distance. diffusion coefficient ,in Reference diffusion coefficient For temperature sensitivity coefficient, Let be the diffusion coefficient between the i-th node and the j-th node. Let the water temperature be at the i-th node. For reference temperature; Using message passing, aggregation, and updating as the framework of a graph neural network, the message propagation is calculated based on the nodes' neighbors: ; in For the i-th node and its neighbors The news For the j-th neighbor of the i-th node, For bias vectors, This is the weight matrix. For activation function, Let be the edge between the i-th node and the j-th node. For the i-th node; We use a weighted summation to aggregate all neighbor messages of a node, expressed as: ; in For the total neighbor messages of the i-th node, For the radius of diffusion influence, For nodes The set of neighbors; The three-dimensional current velocity is obtained from the ocean dynamic field data by bilinear interpolation based on the node location. By combining the diffusion message and the velocity field, motion updates are performed, and the new position of the node is calculated: ; in For nodes longitude, For nodes latitude, For nodes depth, This represents the eastward flow velocity component. For the northward flow velocity component, The vertical velocity component is... For nodes The updated three-dimensional spatial coordinates For time step, To influence the weighting of diffusion, For the longitudinal component of the total neighbor message, For the latitudinal component of the total neighbor messages, For the vertical component of the total neighbor messages, For the settling velocity of the oil clump, It is the acceleration due to gravity. Let be the radius of the oil droplet. The viscosity of seawater, The density of seawater, For oil density, This is the oil droplet shape factor; The expression for updating node concentrations based on the oil droplet volatilization and emulsification processes is as follows: ; in The attenuation coefficient is... For nodes concentration, For the updated node concentration, For the concentration of the j-th node, This is a temperature correction factor; The undirected graph is input into the graph neural network and extrapolated over multiple time steps. The node positions of all time steps are integrated to form a simulated oil mass migration path.

[0009] Furthermore, the method for obtaining the movement and migration patterns of the oil clumps includes: An oil clump migration simulator based on graph neural networks is used as the simulation method, where the policy space is the model parameter level, including ocean dynamic field weights, diffusion coefficients, and settling rates. The actual migration path observation system of the oil plume is taken as the actual party, where the strategy space is the observation error distribution. ,in It is the covariance matrix; The profit function is defined with the objective of minimizing the spatiotemporal deviation between the simulated path and the actual path: ; in This is a sequence of spatial coordinates for simulating the path. This is the spatial coordinate sequence of the actual path. For the time series of the simulation path, This is the time series of the actual path. The weighting coefficients for the root mean square error are: The weighting coefficients for the mean absolute error. These are the simulation parameters; A multi-layered distributed fuzzy dominant strategy iterative algorithm is adopted to find the Pareto-Nash equilibrium through bidirectional iteration of the outer loop game situation and the inner loop individual strategy. The specific steps are as follows: randomly generate the initial simulation parameters. Set an observation error threshold; fix the actual observation error, and update the simulation parameters through particle swarm optimization, expressed as: ,in For inertial weights, , The learning factor for particle swarm optimization. For the individual optimal parameters, The parameters are globally optimal. Let be the parameter vector of the simulation scheme at the u-th iteration. Let be the parameter vector of the simulation scheme at the (u+1)th iteration. Let be the velocity vector of the particle at the u-th iteration. , Use random numbers within the range [0,1]; dynamically adjust the observation error weights based on the deviation between the simulated path and the actual path. Construct a multi-objective optimal membership matrix to quantify the game payoffs of different parameter combinations. The expression is as follows: ,in To provide a given observation error The maximum value of the utility values ​​of all the options. To provide a given observation error The minimum utility value of all possible solutions. Let i be the utility value of the i1th scheme under the j1st observation error condition. The optimal membership function. The element in the i1th row and j1st column of the optimal membership matrix for a multi-objective system; when the parameter update amount in continuous iterations satisfies Or the game payout reaches a preset threshold. When the iteration stops, the iteration is stopped. This is the convergence threshold; The final output is the calibrated simulation parameters. This includes the ocean dynamic field coupling coefficient, oil clump diffusion coefficient, and settling rate correction factor; The movement characteristics of oil masses were inverted through game equilibrium solutions: regularity function. ,in, It is a spatiotemporal function used to predict the migration path of oil clumps under different environmental conditions.

[0010] Furthermore, the method for constructing an intelligent identification model for submerged oil pollution based on physical information embedding according to the aforementioned oil mass movement and migration patterns includes: The intelligent identification model for submerged oil pollution adopts a three-layer architecture: data-driven, physically constrained, and dynamically evolving. It integrates acoustic scattering mechanisms and neural differential equations. The core of the model comprises two parts: a physically constrained loss function and continuous dynamic system modeling. Specifically, the structure includes an input layer, a feature fusion layer, a physical embedding layer, and an output layer. The input layer receives spatiotemporally aligned multimodal data, forming a 128-dimensional feature vector. The feature fusion layer uses a Cross-Modal Transformer architecture, dynamically allocating modal weights through an attention mechanism to output a fused feature matrix. The physical embedding layer embeds acoustic scattering mechanism equations and neural differential equations to constrain static identification accuracy and dynamic migration processes, respectively. The output layer outputs a heatmap of submerged oil pollution probability and a migration trajectory prediction curve. The acoustic scattering mechanism equation is embedded as a regularization term into the traditional cross-entropy loss function to form a hybrid loss function, expressed as follows:

[0011]

[0012]

[0013]

[0014] in For cross-entropy loss, For physical regularization, The weight coefficients for the regularization term. For a mixed loss function, This is the sonar echo attenuation coefficient. In order to receive the echo amplitude, For the amplitude of the transmitted signal, To detect distance, Labels exist for oil clumps. To predict probabilities for the model, The attenuation coefficient predicted by the model. These are theoretical values ​​calculated based on differences in acoustic impedance. The total number of training samples, The sonar signal attenuation coefficient in ocean water; The oil sluice deposition process is modeled as a continuous dynamic system, and the NODE architecture is used to describe the migration trajectory of the oil sluice in the ocean dynamic field: the location, volume, and density of the oil sluice are taken as state variables, and the transpose of the concatenation forms the state variable vector; based on the ocean dynamic field and the physical properties of the oil sluice, a state evolution equation is established, the expression of which is: ,in Let be the velocity field vector at time k. Let k be the state variable vector at time k. For model parameters, The function is a nonlinear function parameterized by a neural network; the Runge-Kutta method is used to solve the differential equation, with a time step of 10 minutes, and the migration trajectory of the oil patch is iteratively calculated over the next 72 hours. Construct training and validation sets, freeze the neural differential equation module, train only the feature fusion layer and physical constraint loss function, and iterate 50 times using the Adam optimizer; unfreeze the neural differential equation module, and adjust the learning rate using a step decay method: the initial learning rate is 1e. -4 The learning rate decreases by 0.5 every 20 rounds until it drops to 5e. -5 The process remains unchanged, and training continues using a multi-task loss function for 100 iterations. A transfer learning strategy is introduced, using weights pre-trained on the MADOS dataset to initialize the feature extraction network. The oil patch density output by the model is then adjusted. Constraints were imposed to ensure that the submersion conditions were met. The simulated trajectory was compared with the actual AUV observation trajectory, and the weights of the ocean dynamic field were dynamically adjusted through game analysis. A thermal map of the spatial distribution of submerged oil, a prediction of the oil clump migration trajectory in the next 72 hours, and a confidence level of the pollution probability were obtained. The spatial distribution heat map of submerged oil, the predicted migration trajectory of oil clumps in the next 72 hours, and the confidence level of pollution probability are output as the identification results.

[0015] The beneficial effects of this invention are: This invention is a smart identification method for submerged oil pollution based on multi-source heterogeneous data fusion. Compared with existing technologies, this invention has the following technical advantages: This invention integrates multi-source heterogeneous data with physical information embedding technology to achieve accurate and intelligent identification of submerged oil pollution, offering significant advantages over existing technologies: It optimizes spatiotemporal registration through a three-dimensional perception system, unifying the spatiotemporal benchmark of multi-source data, resolving data heterogeneity issues, and improving the effectiveness of data fusion; it combines graph neural networks and game theory analysis to invert oil clump migration patterns, accurately depicting the dynamic movement characteristics of oil clumps; and it constructs a three-layer architecture identification model that integrates acoustic scattering mechanisms and neural differential equations, balancing data-driven approaches and physical constraints, outputting pollution heat maps, migration trajectories, and confidence levels, significantly improving identification accuracy and predictive capabilities, and providing scientific and efficient technical support for emergency prevention and control of submerged oil pollution. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion according to the present invention. Detailed Implementation

[0017] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0018] The present invention provides an intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Multi-source monitoring data and meteorological data of the submerged oil area are collected, and the multi-source monitoring data and meteorological data are preprocessed; the multi-source monitoring data includes optical remote sensing images, SAR remote sensing images, sensor data, AIS ship trajectory data, marine numerical model data, hydrological data and the actual migration path of the oil clump; In the actual assessment, a certain nearshore sea area (122°30′~122°45′E, 37°10′~37°25′N) was taken as the research object. The nearshore sea area was polluted by submerged oil caused by a cargo ship fuel oil leak. The leaked oil was marine diesel, and the initial leak volume was about 80t. The oil slick migrated southwestward under the action of ocean currents and tides. Some oil bodies became submerged due to seawater mixing and particle adsorption. The submerged depth was mainly distributed in the sea area of ​​5~20m. The experimental monitoring period was 72 hours, covering the entire process of oil slick diffusion, submersion and migration. The multi-source monitoring data collected in this experiment were all measured data from the sea area. Meteorological data came from real-time monitoring at a nearshore meteorological station: optical remote sensing images were acquired once every 6 hours, with core monitoring parameters including oil slick reflectivity, suspected pollution area outline, and sea surface roughness; SAR remote sensing images were acquired once every 12 hours, with core monitoring parameters including low-albedo oil film areas, oil slick area, and morphology; sensor data were acquired once every 5 minutes, with core monitoring parameters including water temperature, salinity, dissolved oxygen, sonar echo amplitude, and oil fingerprint concentration; AIS vessel trajectory data were acquired once every 30 seconds, with core monitoring parameters including... The parameters monitored include the location, speed, and heading of the leaking vessel and surrounding vessels; ocean numerical model data is collected once per hour, with core monitoring parameters being the three-dimensional velocity field (eastward / northward / vertical), mixing layer depth, and ocean current eddy distribution; hydrological data is collected once every two hours, with core monitoring parameters being seawater density, viscosity, water depth, and tidal height; the actual migration path of the oil slick is collected once per hour, with core monitoring parameters being the three-dimensional spatial coordinates of the oil slick, the outline of the submerged oil slick, and concentration distribution; meteorological data is collected once per hour, with core monitoring parameters being wind speed, wind direction, air pressure, air temperature, and precipitation. Data preprocessing: Radiometric calibration, atmospheric correction, and geometric correction were performed on remote sensing images to remove noise interference from clouds, waves, etc., and suspected oil clump areas were extracted; outlier removal (using the 3σ principle) and missing value completion (linear interpolation) were performed on sensor data, and sonar echo signals were converted into standardized amplitude values; trajectory smoothing was performed on AIS ship trajectory data, outlier waypoints were removed, and the core motion trajectory of the leaking ship was retained; the spatiotemporal reference of all data was unified: timestamps were initially converted to UTC standard time, and spatial coordinates were initially converted to the WGS84 coordinate system to prepare for subsequent spatiotemporal registration and optimization; A three-dimensional sensing system is used to perform spatiotemporal registration optimization on the multi-source monitoring data and the meteorological data to obtain spatially aligned data, and deep multimodal fusion is performed on the spatially aligned data to obtain fused data; A graph neural network is used to simulate the flow migration of oil masses in the fused data to obtain the simulated oil mass migration path. Based on the actual oil mass migration path and the simulated oil mass migration path, a game analysis of oil mass migration is performed to obtain the oil mass movement and migration law. Based on the movement and migration patterns of the oil clumps, a smart identification model for submerged oil pollution based on physical information embedding is constructed. The data to be identified is input into the smart identification model for submerged oil pollution, and the identification results are output.

[0019] In this embodiment, the method for spatiotemporal registration optimization of the multi-source monitoring data and the meteorological data using a three-dimensional sensing system includes: The original timestamps of each sensor are collected and converted to UTC standard time. For missing timestamp data, linear interpolation is used to complete them. Initial time synchronization is achieved using Kalman filtering, and a time synchronization error state vector is defined. State transition matrix ,in Here is the state transition matrix. The sampling interval is... Due to time deviation, For transpose, The process noise at time k, Let k be the state vector at time k-1; Particle swarm optimization error correction is employed, and an observation vector is constructed based on the GPS second pulse signal. The time deviation is corrected iteratively using Kalman filtering, expressed as:

[0020] in The observation noise at time k, Let k be the time deviation at time k. The Kalman gain at time k, For the observation matrix, Let k be the observation vector at time k. Let be the optimal state vector at time k. Let be the optimal state vector at time k-1. This is a predicted value for the current state. Minimizing the time synchronization error is transformed into an optimization problem, expressed as:

[0021] in Let be the reference time for the i-th sensor. Let be the original timestamp of the i-th sensor. The number of sensors; The particle swarm is initialized with the fitness function being the sum of squared time deviations. The particle velocity and position are iteratively updated until convergence. Based on the optimized time deviations, the sonar, electromagnetic, and optical data are timestamped to control the time difference between different modes of data within the same monitoring area to within 5 seconds. Unify multi-source data into the WGS84 coordinate system to achieve spatial consistency of air, space, sea and underwater three-dimensional data, and achieve preliminary alignment of data at different layers based on a common geographic reference using the least squares method. Introducing prior knowledge of the ocean dynamic field to construct a spatial transformation model, the expression is:

[0022] in Rotation matrix It is a translation vector. This represents the spatial offset caused by the flow velocity field. For the spatial coordinates of the target image, These are the corresponding spatial coordinates of the source image; Outliers are removed by random sampling consensus algorithm. Based on the SLAM 3D reconstruction results of AUV, the oil blobs outline of underwater multispectral images, the oil droplet scattering features of forward-looking sonar, and the oil fingerprint sampling points of the seabed observation network are integrated into the same 3D grid to form a three-dimensional representation of the spatial distribution of submerged oil and output as spatially aligned data. In the actual evaluation, the particle swarm was initialized with 30 particles, a dimensional sensor count, a search space of [-10, 10] seconds, and a time deviation corresponding to one sensor for each dimension. Particle velocity and position were iteratively updated with an inertial weight of 0.7 and a learning factor of [missing information]. 1.49; 3D mesh: resolution 0.1m × 0.1m × 0.1m; Time synchronization: Kalman filter initial time synchronization: sampling interval is 5 minutes, process noise variance is set to 0.01, and observation noise variance is set to 0.05; Particle swarm optimization error correction: number of particles 30, search space [-10, 10] seconds, inertia weight 0.7, convergence after 20 iterations; Processing results: the time difference of sonar, electromagnetic and optical data in the same monitoring area is controlled within 3.2 seconds, which meets the requirement of 5 seconds or less; Spatial alignment: Using the WGS84 coordinate system as a reference, preliminary alignment of air-space-sea-submerged data was completed using the least squares method, with residuals controlled within 0.5m. A spatial transformation model was constructed by introducing prior knowledge of the dynamic field of the study area, and outliers were removed using a random sampling consensus algorithm (removal rate of approximately 8%). Based on the SLAM 3D reconstruction results of AUV, a 0.1m×0.1m×0.1m 3D grid was constructed, integrating the oil clump contours from underwater multispectral images, the oil droplet scattering characteristics from forward-looking sonar, and the oil fingerprint sampling points from the seabed observation network into the grid to form a three-dimensional representation of the spatial distribution of submerged oil, and outputting spatially aligned data.

[0023] In this embodiment, the method for obtaining fused data by performing deep multimodal fusion on the spatially aligned data includes: An improved U-Net architecture is used to extract image features from optical remote sensing images, SAR images, and AUV underwater multispectral images: multi-scale textures are captured through an Encoder-Decoder structure to obtain feature maps; and global average pooling is used to compress the feature maps into image vectors. The time-series data of the smart buoy sensor and the sonar echo sequence are learned by BiLSTM network to learn the time-series dependencies: forward LSTM captures historical trends, backward LSTM captures future associations, outputs hidden state sequences, and dynamically weights time step features through self-attention mechanism to output time-series vectors. A trajectory graph structure is constructed based on AIS ship trajectory data: ship position points are used as nodes and temporal adjacency relationships are used as edges to form a ship trajectory graph. A graph attention network is used to calculate the attention weights between nodes, aggregate neighborhood information, and output trajectory vectors. Based on the velocity field and hybrid layer depth data output by the ROMS / HYCOM model, the physical parameters are encoded into a motion constraint graph: nodes represent spatial grid cells, edge weights are velocity gradients, hybrid layer depth is used as node features, graph convolutional networks extract dynamic field features, and output constraint vectors. The image vector, time series vector, trajectory vector, and constraint vector are concatenated into a modal query matrix. Calculate the intermodal similarity matrix :

[0024] in For image vectors, For time series vectors, For trajectory vectors, For constraint vectors, For encoding dimensions, The modal number; Using the constraint vector as the guiding vector, the modal weights are adjusted through a gating mechanism, as expressed in the following expression:

[0025] in Let a be the weight of the a-th mode. Let a be the a-th learnable parameter. It is the sigmoid activation function; The modal features are weighted and summed according to their modal weights to generate a fused feature matrix. The original modal features are then added to the fused features to obtain the deep fused feature matrix, expressed as follows:

[0026] in For deep fusion of feature matrices, For fusion feature matrix; The deep fusion features are mapped to fusion vectors through a fully connected layer. The fusion vectors include spatial features, temporal features, and dynamic features. Spatial features are the image texture and location coordinates of the suspected oil blobs. Temporal features are the changing trends of environmental parameters monitored by the sensors. Dynamic features are the probability distribution of oil blobs migration under the constraint of the ocean dynamic field. In practical evaluation, feature extraction was performed as follows: Image features (optical / SAR / AUV multispectral): Improved U-Net architecture, with 5 layers for both the encoder and decoder, outputting a 256-dimensional image vector after global average pooling; Temporal features (sensor / sonar echo): BiLSTM network, with 128 hidden layer units and 8 self-attention mechanism heads, outputting a 256-dimensional temporal vector; Trajectory features (AIS ships): Graph attention network, with the number of nodes equal to the total number of ship location points (1200 in total), 3 neighborhood aggregation layers, outputting a 256-dimensional trajectory vector; Constraint features (ocean dynamic field): Graph convolutional network, with 500×500 grid nodes, edge weights equal to the current velocity gradient, outputting a 256-dimensional constraint vector. Modality fusion: Four types of vectors are concatenated to form a 256×4 modality query matrix M, and the inter-modality similarity matrix S is calculated. Using the constraint vector as the guiding vector, the weights of each modality are adjusted through a sigmoid gating mechanism: image modality 0.3, temporal modality 0.4, trajectory modality 0.1, and constraint modality 0.2. After weighted summation, the result is added to the original modal features to obtain a deep fusion feature matrix, which is then mapped to a 1024-dimensional fusion vector through a fully connected layer, containing the spatial, temporal, and dynamic features of the oil patch.

[0027] In this embodiment, the method for simulating the migration path of oil patches by using a graph neural network to simulate the flow migration of oil patches from the fused data includes: The continuous oil mass is discretized into N oil mass units. Each unit is regarded as an independent moving entity and is denoted as a set. The unit division is based on the initial shape of the oil mass, and the size range is 5-50m. Using oil patch units as nodes, the feature vector of each node contains physical properties and environmental interactions; physical properties include location coordinates, density, viscosity, and initial concentration; environmental interactions include current flow velocity, water temperature, and salinity. An undirected graph is constructed based on spatial proximity, where edges exist between node pairs whose distance is less than or equal to the radius of influence of the spread; edge features include Euclidean distance. diffusion coefficient ,in Reference diffusion coefficient For temperature sensitivity coefficient, Let be the diffusion coefficient between the i-th node and the j-th node. Let the water temperature be at the i-th node. For reference temperature; Using message passing, aggregation, and updating as the framework of a graph neural network, the message propagation is calculated based on the nodes' neighbors: ; in For the i-th node and its neighbors The news For the j-th neighbor of the i-th node, For bias vectors, This is the weight matrix. For activation function, Let be the edge between the i-th node and the j-th node. For the i-th node; We use a weighted summation to aggregate all neighbor messages of a node, expressed as:

[0028] in For the total neighbor messages of the i-th node, For the radius of diffusion influence, For nodes The set of neighbors; The three-dimensional current velocity is obtained from the ocean dynamic field data by bilinear interpolation based on the node location. By combining the diffusion message and the velocity field, motion updates are performed, and the new position of the node is calculated:

[0029] in For nodes longitude, For nodes latitude, For nodes depth, This represents the eastward flow velocity component. This represents the northward flow velocity component. The vertical velocity component is... For nodes The updated three-dimensional spatial coordinates For time step, To influence the weighting of diffusion, For the longitudinal component of the total neighbor message, For the latitudinal component of the total neighbor messages, For the vertical component of the total neighbor messages, For the settling velocity of the oil clump, It is the acceleration due to gravity. Let be the radius of the oil droplet. The viscosity of seawater, The density of seawater, For oil density, This is the oil droplet shape factor; The expression for updating node concentrations based on the oil droplet volatilization and emulsification processes is as follows:

[0030] in The attenuation coefficient is... For nodes concentration, For the updated node concentration, For the concentration of the j-th node, This is a temperature correction factor; The undirected graph is input into the graph neural network and extrapolated over multiple time steps. The node positions of all time steps are integrated to form a simulated oil mass migration path. In actual assessments, the fused data includes the current field and mixing layer depth output from the ocean numerical model; the initial state of the oil clump: location coordinates, density, viscosity, and initial concentration; and environmental parameters: water temperature, salinity, and wind speed. The number of modalities is 4, corresponding to four modalities: image, time series, trajectory, and physical constraint; the encoding dimension of each modal vector is uniformly 256 dimensions. Oil patch units are divided based on the initial morphology of the oil patch and the contaminated area: the contaminated area is less than 1 km². 2 At that time, the unit size is 5~10m; the polluted area is greater than or equal to 1km². 2 and less than 10km 2 At that time, the unit size is 10~30m; the polluted area is greater than or equal to 10km². 2 At that time, the unit size is taken as 30~50m; In this experiment, the initial contaminated area of ​​the oil plume was approximately 5 km². 2 According to the method requirements, the continuous oil patch was discretized into 800 oil patch units, with a unit size of 20m. The core implementation steps and parameters are as follows: 1. Construct an undirected graph: with oil patch units as nodes, the node features include physical attributes (location coordinates, diesel density 840kg / m³). 3 The study considered the interaction between the following parameters: viscosity 3.5 mPa·s, initial concentration 1200 mg / L, and environmental parameters (measured water temperature 18℃, salinity 32‰, three-dimensional flow velocity). The diffusion influence radius was set to 50 m, and the edge features included Euclidean distance and diffusion coefficient (reference diffusion coefficient was 0.01 m). 2 / s, temperature sensitivity coefficient is 0.02℃ -12. Graph Neural Network Derivation: Message Passing: The weight matrix has a dimension of 256×256, the bias vector has a dimension of 256×1, and the activation function is ReLU; Message Aggregation: The diffusion influence weight is 0.3, and the total neighbor messages of the node are obtained after weighted summation; Motion Update: The time step is 10 minutes, the droplet radius is 50μm, the droplet shape coefficient is 0.8, and the gravitational acceleration is 9.8m / s². 2 The calculated settling velocity of the oil clump was 0.002 m / s; concentration update: attenuation coefficient was 0.001 h. -1 The temperature correction factor is 0.01℃. -1 Considering the diesel volatilization and emulsification process, the node concentration was updated; 3. Simulation results: Through 36 time steps (72 hours), the simulated oil clump migration path was output, covering the planar diffusion and vertical submergence process of the oil clump, simulating the overall migration of the oil clump in the southwest direction, with the submergence depth gradually increasing to 10~25m.

[0031] In this embodiment, the method for obtaining the movement and migration patterns of the oil clump includes: An oil clump migration simulator based on graph neural networks is used as the simulation method, where the policy space is the model parameter level, including ocean dynamic field weights, diffusion coefficients, and settling rates. The actual migration path observation system of the oil plume is taken as the actual party, where the strategy space is the observation error distribution. ,in It is the covariance matrix; The profit function is defined with the objective of minimizing the spatiotemporal deviation between the simulated path and the actual path:

[0032] in This is a sequence of spatial coordinates for simulating the path. This is the spatial coordinate sequence of the actual path. For the time series of the simulation path, This is the time series of the actual path. The weighting coefficients for the root mean square error are: The weighting coefficients for the mean absolute error. These are the simulation parameters; A multi-layered distributed fuzzy dominant strategy iterative algorithm is adopted to find the Pareto-Nash equilibrium through bidirectional iteration of the outer loop game situation and the inner loop individual strategy. The specific steps are as follows: randomly generate the initial simulation parameters. Set an observation error threshold; fix the actual observation error, and update the simulation parameters through particle swarm optimization, expressed as: ,in For inertial weights, , The learning factor for particle swarm optimization. For the individual optimal parameters, The parameters are globally optimal. Let be the parameter vector of the simulation scheme at the u-th iteration. Let be the parameter vector of the simulation scheme at the (u+1)th iteration. Let be the velocity vector of the particle at the u-th iteration. , Use random numbers within the range [0,1]; dynamically adjust the observation error weights based on the deviation between the simulated path and the actual path. Construct a multi-objective optimal membership matrix to quantify the game payoffs of different parameter combinations. The expression is as follows: ,in To provide a given observation error The maximum value of the utility values ​​of all the options. ) is for a given observation error The minimum utility value of all possible solutions. Let i be the utility value of the i1th scheme under the j1st observation error condition. The optimal membership function. The element in the i1th row and j1st column of the optimal membership matrix for a multi-objective system; when the parameter update amount in continuous iterations satisfies Or the game payout reaches a preset threshold. When the iteration stops, the iteration is stopped. This is the convergence threshold; The final output is the calibrated simulation parameters. This includes the ocean dynamic field coupling coefficient, oil clump diffusion coefficient, and settling rate correction factor; The movement characteristics of oil masses were inverted through game equilibrium solutions: regularity function. ,in, It is a spatiotemporal function used to predict the migration path of oil clumps under different environmental conditions; In actual evaluation, the convergence threshold is 10. -5 Using a graph neural network migration simulator as the simulater and the actual migration path of oil clumps measured by AUVs as the actual player, a multi-layer distributed fuzzy dominant strategy iterative algorithm was employed to complete the game analysis. The core parameters and pattern extraction results are as follows: 1. Game settings: Simulator strategy space: Ocean dynamic field weights (0.4~0.8), diffusion coefficients (0.008~0.012m) 2 / s), settling rate (0.0015~0.0025m / s); the actual strategy space covariance matrix is ​​a 3×3 diagonal matrix, with diagonal elements set to 0.5; the profit function: root mean square error weight is 0.7, mean absolute error weight is 0.3, with the goal of minimizing spatiotemporal bias; 2. Iterative optimization: particle swarm optimization parameters: inertia weight is 0.6, 1.5 The iteration count was 1.5; the number of iterations was 30 for the outer loop game situation and 50 for the inner loop individual strategy. The parameter update satisfied the convergence condition on the 45th iteration, at which point the iteration stopped; 3. Pattern extraction results: output calibrated simulation parameters: ocean dynamic field coupling coefficient 0.65, oil clump diffusion coefficient 0.0095m. 2 / s, settling rate correction factor 1.12; the inversion yielded the following oil slick movement and migration patterns: the migration velocity of submerged oil slicks in this sea area is positively correlated with the eastward current velocity (correlation coefficient 0.89), and the submerging depth is negatively correlated with the depth of the seawater mixing layer (correlation coefficient -0.78). In the study area with a water temperature of 15~20℃ and a salinity of 30~35‰, the diffusion area of ​​diesel oil slicks increases by approximately 0.12 km² per hour. 2 The vertical diving depth increases by 0.3~0.5m per hour.

[0033] In this embodiment, the method for constructing an intelligent identification model for submerged oil pollution based on physical information embedding according to the movement and migration patterns of the oil clump includes: The intelligent identification model for submerged oil pollution adopts a three-layer architecture: data-driven, physically constrained, and dynamically evolving. It integrates acoustic scattering mechanisms and neural differential equations. The core of the model comprises two parts: a physically constrained loss function and continuous dynamic system modeling. Specifically, the structure includes an input layer, a feature fusion layer, a physical embedding layer, and an output layer. The input layer receives spatiotemporally aligned multimodal data, forming a 128-dimensional feature vector. The feature fusion layer uses a Cross-Modal Transformer architecture, dynamically allocating modal weights through an attention mechanism to output a fused feature matrix. The physical embedding layer embeds acoustic scattering mechanism equations and neural differential equations to constrain static identification accuracy and dynamic migration processes, respectively. The output layer outputs a heatmap of submerged oil pollution probability and a migration trajectory prediction curve. The acoustic scattering mechanism equation is embedded as a regularization term into the traditional cross-entropy loss function to form a hybrid loss function, expressed as follows:

[0034]

[0035]

[0036]

[0037] in For cross-entropy loss, For physical regularization, The weight coefficients for the regularization term. For a mixed loss function, This is the sonar echo attenuation coefficient. In order to receive the echo amplitude, For the amplitude of the transmitted signal, To detect distance, Labels exist for oil masses. To predict probabilities for the model, The attenuation coefficient predicted by the model. These are theoretical values ​​calculated based on differences in acoustic impedance. The total number of training samples, The sonar signal attenuation coefficient in ocean water; The oil sluice deposition process is modeled as a continuous dynamic system, and the NODE architecture is used to describe the migration trajectory of the oil sluice in the ocean dynamic field: the location, volume, and density of the oil sluice are taken as state variables, and the transpose of the concatenation forms the state variable vector; based on the ocean dynamic field and the physical properties of the oil sluice, a state evolution equation is established, the expression of which is: ,in Let be the velocity field vector at time k. Let k be the state variable vector at time k. For model parameters, The function is a nonlinear function parameterized by a neural network; the Runge-Kutta method is used to solve the differential equation, with a time step of 10 minutes, and the migration trajectory of the oil patch is calculated iteratively over the next 72 hours. Construct training and validation sets, freeze the neural differential equation module, train only the feature fusion layer and physical constraint loss function, and iterate 50 times using the Adam optimizer; unfreeze the neural differential equation module, and adjust the learning rate using a step decay method: the initial learning rate is 1e. -4 The learning rate decreases by 0.5 every 20 rounds until it drops to 5e. -5 The process remains unchanged, and training continues using a multi-task loss function for 100 iterations. A transfer learning strategy is introduced, using weights pre-trained on the MADOS dataset to initialize the feature extraction network. The oil patch density output by the model is then adjusted. Constraints were imposed to ensure that the submersion conditions were met. The simulated trajectory was compared with the actual AUV observation trajectory, and the weights of the ocean dynamic field were dynamically adjusted through game analysis. A thermal map of the spatial distribution of submerged oil, a prediction of the oil clump migration trajectory in the next 72 hours, and a confidence level of the pollution probability were obtained. The spatial distribution heat map of submerged oil, the predicted migration trajectory of oil clumps in the next 72 hours, and the confidence level of pollution probability are output as the identification results. In practical assessments, multimodal data includes sonar echo signals, optical / SAR remote sensing features, and ocean dynamic field parameters. Sonar echo signals are used to obtain topological charge changes, optical / SAR remote sensing features are used to obtain the gray-level co-occurrence matrix and sea surface roughness, and ocean dynamic field parameters are used to obtain current velocity and mixing layer depth. Sonar data has a weight of 0.4, remote sensing data a weight of 0.3, and dynamic field data a weight of 0.3. The sonar signal attenuation coefficient in the ocean water is calculated from hydrological data (water temperature, salinity, and depth). Cross-entropy loss is used to measure the difference between the model's predicted probability and the true label. The physical regularization term is constructed based on the acoustic scattering mechanism equation to constrain the consistency between the model output and physical laws. The optimal value was determined through verification experiments; Training set: Contains over 100,000 samples, covering sonar echoes, remote sensing images, and corresponding oil clump migration trajectories under different oil types and sea states; Validation set: Real-world data from an oil spill accident, spanning 200km. 2 The sea area is monitored continuously for 48 hours; the oil types include crude oil and diesel; the sea conditions are temperature 5-30℃ and salinity 28-35‰; the Adam optimizer's learning rate is 1e. -4 ; The multi-task loss function is spatial recognition loss + trajectory prediction loss, and the diving condition is... Submersion occurred at times. Based on the maximum density setting of seawater under normal marine salinity (28-35‰) and water temperature (5-30℃), the oil body is ensured to meet the sinking and submerging characteristics and sink to the underwater non-suspended area; The MADOS dataset is a public dataset for monitoring marine oil spills and submerged oil pollution. It contains multi-source heterogeneous monitoring data and oil clump migration annotation data. The multi-source heterogeneous monitoring data includes remote sensing, sonar, and sensor data. Input layer: Receives spatiotemporally aligned multimodal data to form a 128-dimensional feature vector; Feature fusion layer: Cross-Modal Transformer architecture with 16 attention heads and 512 hidden layer dimensions; Physical embedding layer: Embeds acoustic scattering mechanism equations (the sonar signal medium attenuation coefficient is calculated from measured hydrological data to be 0.005 dB / m), uses the NODE architecture to describe the oil clump settling process, solves the differential equations using the Runge-Kutta method, and has a time step of 10 minutes; Output layer: Outputs a heatmap of the probability of submerged oil pollution (resolution 0.1 m) and a 72-hour migration trajectory prediction curve; The weight of the regularization term was determined to be 0.3 through verification experiments. The weight of spatial recognition loss in the multi-task loss function was 0.6, and the weight of trajectory prediction loss was 0.4. Dataset: The training set consists of 100,000+ measured samples of marine oil spills (including diesel, crude oil, and other oil types); the validation set consists of 48 hours of measured data (200km) of oil spill events in the study area. 2(Sea area); Training strategy: Use the MADOS dataset to pre-train weights to initialize the feature extraction network; first freeze the NODE module, then iterate the Adam optimizer for 50 rounds (initial learning rate 1e). -4 Then unfreeze the NODE module and adjust the learning rate using a step decay method (decaying by 0.5 every 20 rounds, down to 5e). -5 (Continue to iterate for 100 rounds); Submerged condition constraint: Model output oil clump density The time was determined to be submerged, which matched the maximum density of seawater in the study area; The 72-hour data to be identified from the oil spill in this study area was input into the trained intelligent identification model, which output three core identification results. The model's accuracy was verified by matching the actual monitoring results with a 92.5% accuracy: 1. Spatial distribution heat map of submerged oil: The heat map clearly depicts the three-dimensional spatial distribution of submerged oil blobs in the study area. The submersion depth of the core pollution area is 5-25m, and the area of ​​the high pollution probability zone (>90%) is approximately 3.8km². 2 The deviation from the measured pollution area is less than 0.3 km. 2 2. Prediction of oil mass migration trajectory in the next 72 hours: The oil mass is predicted to continue migrating southwestward. The polluted area will expand in 72 hours, with a maximum migration distance of about 18km and a submersion depth of 15-30m. The spatiotemporal RMSE of the measured migration trajectory is 1.2km. 3. Pollution probability confidence: The pollution probability confidence of the entire monitoring area is above 85%, and the confidence of the core polluted area reaches 95%, providing a highly reliable decision-making basis for pollution emergency prevention and control.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of submerged oil pollution based on multi-source heterogeneous data fusion, characterized in that, Includes the following steps: Multi-source monitoring data and meteorological data of the submerged oil area are collected, and the multi-source monitoring data and meteorological data are preprocessed; the multi-source monitoring data includes optical remote sensing images, SAR remote sensing images, sensor data, AIS ship trajectory data, marine numerical model data, hydrological data and the actual migration path of the oil clump; A three-dimensional sensing system is used to perform spatiotemporal registration optimization on the multi-source monitoring data and the meteorological data to obtain spatially aligned data, and deep multimodal fusion is performed on the spatially aligned data to obtain fused data; A graph neural network is used to simulate the migration of oil masses along the flow in the fused data to obtain the simulated oil mass migration path. Based on the actual oil mass migration path and the simulated oil mass migration path, a game analysis of oil mass migration is performed to obtain the oil mass movement and migration law. Based on the movement and migration patterns of the oil clumps, a smart identification model for submerged oil pollution based on physical information embedding is constructed. The data to be identified is input into the smart identification model for submerged oil pollution, and the identification results are output.

2. The intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion according to claim 1, characterized in that, A method for spatiotemporal registration and optimization of the multi-source monitoring data and the meteorological data using a three-dimensional sensing system includes: The original timestamps of each sensor are collected and converted to UTC standard time. For missing timestamp data, linear interpolation is used to complete them. Initial time synchronization is achieved using Kalman filtering, and a time synchronization error state vector is defined. State transition matrix ,in Here is the state transition matrix. The sampling interval is... Due to time deviation, For transpose, The process noise at time k, Let k be the state vector at time k-1; Particle swarm optimization error correction is employed, and an observation vector is constructed based on the GPS second pulse signal. The time deviation is corrected iteratively using Kalman filtering, expressed as: ； in The observation noise at time k, Let k be the time deviation at time k. The Kalman gain at time k, For the observation matrix, Let k be the observation vector at time k. Let be the optimal state vector at time k. Let be the optimal state vector at time k-1. This is a predicted value for the current state. Minimizing the time synchronization error is transformed into an optimization problem. The particle swarm is initialized with the fitness function being the sum of squared time deviations. The particle velocity and position are iteratively updated until convergence. Based on the optimized time deviation, the sonar, electromagnetic, and optical data are timestamped to control the time difference of different modal data within the same monitoring area to within 5 seconds. Unify multi-source data into the WGS84 coordinate system to achieve spatial consistency of air, space, sea and underwater three-dimensional data, and achieve preliminary alignment of data at different layers based on a common geographic reference using the least squares method. A spatial transformation model is constructed by introducing prior knowledge of the ocean dynamic field. Outliers are eliminated by random sampling consensus algorithm. Based on the SLAM three-dimensional reconstruction results of AUV, the oil clump contours of underwater multispectral images, the oil droplet scattering features of forward-looking sonar, and the oil fingerprint sampling points of the seabed observation network are integrated into the same three-dimensional grid to form a three-dimensional representation of the spatial distribution of submerged oil and output as spatially aligned data.

3. The intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion according to claim 1, characterized in that, A method for obtaining fused data by performing deep multimodal fusion on the spatially aligned data includes: An improved U-Net architecture is used to extract image features from optical remote sensing images, SAR images, and AUV underwater multispectral images: multi-scale textures are captured through an Encoder-Decoder structure to obtain feature maps; and global average pooling is used to compress the feature maps into image vectors. The time-series data of the smart buoy sensor and the sonar echo sequence are learned by BiLSTM network to learn the time-series dependencies: forward LSTM captures historical trends, backward LSTM captures future associations, outputs hidden state sequences, and dynamically weights time step features through self-attention mechanism to output time-series vectors. A trajectory graph structure is constructed based on AIS ship trajectory data: ship position points are used as nodes and temporal adjacency relationships are used as edges to form a ship trajectory graph. A graph attention network is used to calculate the attention weights between nodes, aggregate neighborhood information, and output trajectory vectors. Based on the velocity field and hybrid layer depth data output by the ROMS / HYCOM model, the physical parameters are encoded into a motion constraint graph: nodes represent spatial grid cells, edge weights are velocity gradients, hybrid layer depth is used as node features, graph convolutional networks extract dynamic field features, and output constraint vectors. The image vector, time series vector, trajectory vector, and constraint vector are concatenated into a modal query matrix. Calculate the intermodal similarity matrix ; The constraint vector is used as the guiding vector, and the weights of each mode are adjusted through a gating mechanism. The modal features are weighted and summed according to the modal weights to generate a fusion feature matrix. The original modal features are added to the fusion features to obtain a deep fusion feature matrix. The deep fusion features are mapped into a fusion vector through a fully connected layer. The fusion vector includes spatial features, temporal features, and dynamic features. The spatial features are the image texture and location coordinates of the suspected oil blobs. The temporal features are the changing trends of environmental parameters monitored by the sensors. The dynamic features are the probability distribution of oil blobs migration under the constraints of the ocean dynamic field.

4. The intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion according to claim 1, characterized in that, A method for simulating the migration path of oil patches by using a graph neural network to simulate the flow migration of the fused data includes: The continuous oil mass is discretized into N oil mass units. Each unit is regarded as an independent moving entity and is denoted as a set. The unit division is based on the initial shape of the oil mass, and the size range is 5-50m. Using oil patch units as nodes, the feature vector of each node contains physical properties and environmental interactions; physical properties include location coordinates, density, viscosity, and initial concentration; environmental interactions include current flow velocity, water temperature, and salinity. An undirected graph is constructed based on spatial proximity, where edges exist between node pairs whose distance is less than or equal to the radius of influence of the spread; edge features include Euclidean distance. diffusion coefficient ,in Reference diffusion coefficient For temperature sensitivity coefficient, Let be the diffusion coefficient between the i-th node and the j-th node. Let i be the water temperature at the i-th node. For reference temperature; Based on the framework of message passing, aggregation, and updating in graph neural networks, the message is propagated based on the neighbors of the nodes. We employ a weighted summation method to aggregate all neighbor messages of a node, and then use bilinear interpolation to obtain the three-dimensional current velocity from the ocean dynamic field data based on the node's location. The motion is updated by combining the diffusion message and the velocity field, and the new position of the node is calculated. The node concentration is updated based on the oil pulverization and emulsification process. The undirected graph is input into the graph neural network for multi-time step extrapolation. The node positions of all time steps are integrated to form a simulated oil pulverization migration path.

5. The intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method for obtaining the movement and migration patterns of the oil clumps includes: An oil clump migration simulator based on graph neural networks is used as the simulation method, where the policy space is the model parameter level, including ocean dynamic field weights, diffusion coefficients, and settling rates. Using the actual migration path observation system of the oil patch as the actual party, and with the goal of minimizing the spatiotemporal deviation between the simulated path and the actual path, a benefit function is defined. A multi-layered distributed fuzzy dominant strategy iterative algorithm is adopted to find the Pareto-Nash equilibrium through bidirectional iteration of the outer loop game situation and the inner loop individual strategy. The specific steps are as follows: randomly generate the initial simulation parameters. Set an observation error threshold; fix the actual observation error, and update the simulation parameters through particle swarm optimization, expressed as: ,in For inertial weights, , The learning factor for particle swarm optimization. For the individual optimal parameters, The parameters are globally optimal. Let be the parameter vector of the simulation scheme at the u-th iteration. Let be the parameter vector of the simulation scheme at the (u+1)th iteration. Let be the velocity vector of the particle at the u-th iteration. , Use random numbers within the range [0,1]; dynamically adjust the observation error weights based on the deviation between the simulated path and the actual path. Construct a multi-objective optimal membership matrix to quantify the game payoffs of different parameter combinations. The expression is as follows: ,in To provide a given observation error The maximum value of the utility values ​​of all the options. To provide a given observation error The minimum utility value of all possible solutions. Let i be the utility value of the i1th scheme under the j1st observation error condition. The optimal membership function. The element in the i1th row and j1st column of the optimal membership matrix for a multi-objective system; when the parameter update amount in continuous iterations satisfies Or the game payout reaches a preset threshold. When the iteration stops, the iteration is stopped. This is the convergence threshold; The final output is the calibrated simulation parameters. This includes the ocean dynamic field coupling coefficient, oil clump diffusion coefficient, and settling rate correction factor; The movement characteristics of oil masses were inverted through game equilibrium solutions: regularity function. ,in, It is a spatiotemporal function used to predict the migration path of oil clumps under different environmental conditions.

6. The intelligent identification method for submerged oil pollution based on multi-source heterogeneous data fusion according to claim 1, characterized in that, A method for constructing an intelligent identification model for submerged oil pollution based on the movement and migration patterns of the oil clumps, using embedded physical information, includes: The intelligent identification model for submerged oil pollution adopts a three-layer architecture of data-driven, physical constraint, and dynamic evolution, integrating acoustic scattering mechanism and neural differential equations. The core includes two parts: physical constraint loss function and continuous dynamic system modeling. The specific structure includes an input layer, a feature fusion layer, a physical embedding layer, and an output layer. The acoustic scattering mechanism equation is embedded as a regularization term into the traditional cross-entropy loss function to form a hybrid loss function; The oil sluice deposition process is modeled as a continuous dynamic system, and the NODE architecture is used to describe the migration trajectory of the oil sluice in the ocean dynamic field: the location, volume, and density of the oil sluice are taken as state variables, and the transpose of the concatenation forms the state variable vector; based on the ocean dynamic field and the physical properties of the oil sluice, a state evolution equation is established, the expression of which is: ,in Let be the velocity field vector at time k. Let k be the state variable vector at time k. For model parameters, The function is a nonlinear function parameterized by a neural network; the Runge-Kutta method is used to solve the differential equation, with a time step of 10 minutes, and the migration trajectory of the oil patch is iteratively calculated over the next 72 hours. Construct training and validation sets, freeze the neural differential equation module, train only the feature fusion layer and physical constraint loss function, and iterate 50 times using the Adam optimizer; unfreeze the neural differential equation module, and adjust the learning rate using a step decay method: the initial learning rate is 1e. -4 The learning rate decreases by 0.5 every 20 rounds until it drops to 5e. -5 The process remains unchanged, and training continues using a multi-task loss function for 100 iterations. A transfer learning strategy is introduced, using weights pre-trained on the MADOS dataset to initialize the feature extraction network. The oil patch density output by the model is then adjusted. Constraints were imposed to ensure that the submersion conditions were met. The simulated trajectory was compared with the actual AUV observation trajectory, and the weights of the ocean dynamic field were dynamically adjusted through game analysis. A thermal map of the spatial distribution of submerged oil, a prediction of the oil clump migration trajectory in the next 72 hours, and a confidence level of the pollution probability were obtained. The spatial distribution heat map of submerged oil, the predicted migration trajectory of oil clumps in the next 72 hours, and the confidence level of pollution probability are output as the identification results.

Images