A marine nuclear pollution diffusion prediction method based on multi-modal data fusion

Abstract

The application relates to a marine nuclear pollution diffusion prediction method based on multi-modal data fusion, which comprises the following steps: constructing a marine nuclear pollution data set, wherein the marine nuclear pollution data set comprises marine environment data and marine nuclear pollution data; introducing graph structure modeling based on the marine nuclear pollution data set, aggregating node features through graph convolution, and distributing cross-modal attention weight to the features; inputting the distributed features into a ConvNeXt V2 for feature extraction to generate a feature map, and constructing a diffusion initialization network based on nuclear pollution features; constructing a diffusion evolution network, combining continuously acquired marine environment data and marine nuclear pollution data on the basis of the diffusion initialization network, obtaining nuclear pollution diffusion prediction results at different times through three diffusion evolutions, generating a prediction graph, and calculating the centroid trajectory of the nuclear pollution concentration distribution after each diffusion prediction; and realizing automatic and high-precision prediction of marine nuclear pollution diffusion, and significantly improving the prediction efficiency and reliability.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and marine environmental technology, specifically to a method for predicting the spread of marine nuclear pollution based on multimodal data fusion. Background Technology

[0002] Marine nuclear pollution spread prediction falls under the field of marine environmental protection and disaster early warning technology, and is an important component of nuclear safety management. With the widespread application of nuclear energy, nuclear pollution accidents (such as nuclear power plant leaks and nuclear waste leaks) pose an increasing threat to marine ecosystems and human health. Accurately predicting the spread trend and impact range of marine nuclear pollution is of great significance for developing emergency response measures, protecting the marine environment, and reducing ecological and economic losses.

[0003] Traditional methods for predicting the spread of marine nuclear contamination primarily rely on numerical simulations and physical models, such as diffusion equation models and particle tracking models based on ocean current dynamics. These methods calculate the diffusion paths and concentration distribution of nuclear contaminants by simulating ocean currents, wind fields, and hydrological conditions in the marine environment. However, traditional methods have certain limitations in practical applications. First, the marine environment is highly complex and dynamic; the spatiotemporal variations of factors such as ocean currents, wind fields, temperature, and salinity are difficult to model accurately, resulting in limited prediction accuracy. Second, traditional models typically require numerous prior assumptions and parameter adjustments, leading to high computational complexity and making them unsuitable for real-time prediction. Furthermore, traditional methods often struggle to capture complex temporal and spatial characteristics when dealing with nonlinear diffusion processes and multi-source data fusion, resulting in insufficient reliability and robustness of the prediction results.

[0004] In recent years, deep learning technology has demonstrated strong potential in environmental monitoring and disaster prediction. Deep learning methods, by constructing multi-layered neural networks, can automatically extract complex spatial and temporal features from large-scale data and achieve high-precision predictions through end-to-end training. Existing research has attempted to apply deep learning technology to marine pollution diffusion prediction, for example, using convolutional neural networks (CNNs) to extract spatial features of the marine environment or using long short-term memory networks (LSTMs) to capture the temporal patterns of pollution diffusion. Furthermore, graph neural networks (GNNs), by constructing a graph structure of the marine environment and pollution sources, capture spatial topological relationships and dynamic interaction features, significantly improving the ability to fuse multimodal data (such as ocean current velocity, temperature, and nuclide concentration) in complex marine environments. GNNs, through dynamic updates of nodes and edges, can effectively simulate the spatial dependence and nonlinear propagation characteristics of nuclear pollution diffusion. However, when processing multimodal data, GNNs still face challenges such as strong intermodal heterogeneity and low feature fusion efficiency. To address this issue, a cross-modal attention mechanism has been introduced. By adaptively assigning weights to different modalities (such as satellite remote sensing data and buoy sensor data), the semantic association and feature fusion efficiency between modalities are enhanced, thereby improving the model's adaptability to dynamic marine environments and its prediction accuracy.

[0005] However, existing deep learning-based methods for predicting the spread of marine nuclear pollution still have some shortcomings. On the one hand, existing methods rarely consider the overall movement trend of nuclear pollutants, making it difficult to accurately describe the direction and speed of spread. On the other hand, the models are not adaptable enough to the dynamic changes in marine environmental data, and their prediction accuracy and generalization ability in complex scenarios need to be improved. Predicting the spread of nuclear pollution requires determining the centroid, and the traditional centroid trajectory method may be affected by noisy data or nonlinear diffusion, leading to trajectory deviations.

[0006] Therefore, in order to address the problems of low prediction accuracy, high computational complexity, and insufficient capture of temporal features in the existing technology for predicting the spread of marine nuclear pollution, it is urgent to propose a new method that can make full use of deep learning technology, combine graph neural networks and cross-modal attention mechanisms, integrate the spatiotemporal characteristics of marine environmental data and the diffusion patterns of nuclear pollutants, and achieve efficient and accurate prediction of the spread of marine nuclear pollution. Summary of the Invention

[0007] The purpose of this invention is to address the problems of traditional marine nuclear pollution diffusion prediction methods, which mainly rely on numerical simulation and statistical analysis, resulting in long computation time, low prediction accuracy, and difficulty in real-time response. Furthermore, deep learning methods based on a single data source are not adaptable to complex marine environments. This invention proposes a marine nuclear pollution diffusion prediction method based on multimodal data fusion. By integrating multi-source marine environmental data, deep learning algorithms, and an improved centroid trajectory algorithm, it achieves high-precision, real-time prediction of nuclear pollution concentration distribution and diffusion paths.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for predicting the spread of marine nuclear pollution based on multimodal data fusion, comprising the following steps:

[0009] A marine nuclear pollution dataset is constructed and divided into a training set and a test set. The marine nuclear pollution dataset is time-step encoded, and a time-series code is generated for each time step. The marine nuclear pollution dataset includes marine environmental data and marine nuclear pollution data.

[0010] Based on a marine nuclear pollution dataset, graph-structured modeling is introduced, node features are aggregated through graph convolution, and cross-modal attention weights are assigned to the features;

[0011] The assigned features are input into ConvNeXt V2 for feature extraction to generate feature maps, and a diffusion initialization network based on nuclear contamination features is constructed.

[0012] A diffusion evolution network was constructed. Based on the initial diffusion network, and combined with continuously acquired marine environmental data and marine nuclear pollution data, nuclear pollution diffusion prediction results at different times were obtained through three diffusion evolutions. A prediction map was generated, and the centroid trajectory of the nuclear pollution concentration distribution after each diffusion prediction was calculated.

[0013] Analyze the prediction charts, compare the prediction results, and determine the diffusion evolution trend.

[0014] Preferably, the marine environmental data includes ocean current velocity, temperature, salinity, sea surface wind speed, wind direction, and multispectral images; and for densely populated island areas, data at a depth of 0-200m are recorded to capture changes in eddies and wind fields; for open ocean areas, satellite remote sensing is used to cover day and night data.

[0015] The marine nuclear contamination data includes the collection of nuclide concentrations using high-sensitivity gamma-ray detectors and underwater robots. Samples are taken hourly near nuclear power plants, every 3 hours during the activity period of nuclear submarines / aircraft carriers, and daily during the non-activity period.

[0016] Preferably, the marine nuclear pollution dataset is preprocessed, including:

[0017] Preprocess the dataset and interpolate to fill in cloud-obscured areas in the satellite remote sensing data; formula;

[0018]

[0019] Among them, t m t represents the current time step. prev Indicates the previous time step, t next Indicates the next time step, h m w represents length.m z represents the width. m Indicates depth;

[0020] Data augmentation: Simulate island blocking effect (ocean current velocity ±10%) for densely populated island areas; add Gaussian noise (η=0.05) for open ocean areas; simulate sudden high concentration areas (+50%) and concentration gradient changes in nuclear contamination data;

[0021] Normalization: Min-Max normalization was applied to the marine environmental data.

[0022]

[0023] Where x is the value of the original data point, x max and x min These are the minimum and maximum values ​​in the dataset, x. norm It is the normalized value;

[0024] The concentration of nuclear contamination was logarithmically transformed and then normalized using Min-Max.

[0025] c norm =log(1+c);

[0026] Where c is the original nuclear contamination concentration value, c norm It is the value after logarithmic transformation;

[0027] Time step encoding: Generates time embeddings for each time step to support subsequent time encoding in the iTransformer module; uses sine / cosine function form.

[0028]

[0029] Where TE(Δt,2i) and TE(Δt,2i+1) represent the components encoded at time steps, Δt is the time step, i is the encoding dimension index, and D is the encoding dimension (set to 64); a temporal coding tensor with shape (Δt,D) is generated.

[0030] Preferably, in the graph-structured modeling, nodes include marine environment node i and nuclear pollution data node j. Node characteristics are the parameter values ​​of the corresponding spatiotemporal point (t, h, w, z). Edges include connections between nodes of the same modality based on spatial proximity, and connections between nodes of different modalities reflecting the impact of ocean currents on pollution diffusion. The formula for dynamically updating edge weights is:

[0031] w ij (t)=α·Corr(v i (t),c j (t))+β·Phys(v i (t),cj (t))

[0032] Among them, w ij (t) represents the dynamic edge weight between node i and node j at time t, v i (t) represents the characteristics of the marine environment, c j (t) represents the characteristics of nuclear contamination, Corr is the correlation between ocean current and concentration, Phys is based on the advection-diffusion equation, and α and β are learnable parameters;

[0033] Graph convolution aggregation node features:

[0034]

[0035] Where σ is the ReLU activation function, N(i) represents the set of neighbors of node i, and W (l) This represents the learnable weight matrix of the l-th layer. This represents the characteristics of neighbor node j at layer l;

[0036] Preferably, ConvNeXt V2 is used to extract features from the input data to generate a feature map with shape (Δt, H, W, Z, C), where the formula for ConvNeXt V2 is:

[0037] Y = X + λ·Conv 1×1 (GELU(Conv 1×1 (GRN(LN(DC 7×7 (X))))));

[0038] Where GELU is the activation function, GRN is the global response normalization, LN is the layer normalization, and DC is the activation function. 7×7 It employs 7x7 depthwise separable convolutions, with independent convolutions for each channel to extract spatial features. GRN is used for global response normalization, and λ is a learnable scaling factor to adjust the feature representation channel by channel. Residual connections are used to ensure the stability of multimodal data and to handle the high dimensionality and noise characteristics of marine nuclear pollution data.

[0039] Preferably, the diffusion evolution network includes a global diffusion module and two time-coded iTransformer modules; the three-stage diffusion evolution includes:

[0040] Based on the initial distribution input into the global diffusion module, the first temporal coding diffusion evolution is performed to obtain the first diffusion evolution result;

[0041] Based on the results of the first diffusion evolution, the feature map and time step encoding are input into the first temporal coding iTransformer module to perform the second temporal coding diffusion evolution, and the second diffusion evolution result is obtained.

[0042] Based on the results of the second diffusion evolution and the feature map input into the second temporal coding iTransformer module, a third temporal coding diffusion evolution is performed to obtain the final diffusion prediction result.

[0043] Preferably, in the second timing-coded iTransformer module:

[0044] The Douglas-Puk algorithm is used to simplify the concentration distribution and filter out key points in high-concentration areas.

[0045] At key points, priority is given to matching the diffusion trend of high-concentration areas, and Euclidean distance is used to calculate the nearest neighbor matching between key points.

[0046] Preferably, a particle filtering algorithm is introduced to simulate the probability distribution of nuclear contamination diffusion using the Monte Carlo method:

[0047] Particle initialization is performed, with initial weights based on the nuclear contamination concentration distribution;

[0048] Perform state updates; update particle positions based on ocean environmental data.

[0049] x i (t+1)=x i (t)+v current (t)·Δt+η,

[0050] Where η is Gaussian noise, x(t) represents the state at time t, and v current (t) represents the rate of nuclear contamination diffusion at the current time t, where Δt is the time step;

[0051] Update the weights based on the nuclide decay model:

[0052] w i (t+1)=w i (t)·e -λΔt ·p(c i (t))

[0053] Where w(t) is the edge weight at time t, λ is the decay constant, and p(c i ) is the concentration likelihood function;

[0054] Resampling is performed based on particle weights, removing low-weight particles and retaining high-weight particles.

[0055] The centroid is calculated based on particle distribution, and the centroid trajectory is calculated by combining Kalman filtering with the nuclear contamination concentration distribution after each diffusion prediction.

[0056]

[0057] Where xi,j y i,j ,z i,j , where is the coordinate of network point (i,j), and c is the coordinate of the point (i,j). i,j Let t represent the nuclear contamination concentration value at network point (i,j), and t represent the stage of the diffusion prediction.

[0058] Preferably, for the diffusion evolution network, a training dataset is used for training, and a multi-task loss function is introduced; the deviation between the predicted concentration and the actual concentration is calculated.

[0059] The multi-task loss function is: L total =αL concentration +βL centroid ;in:

[0060] Concentration distribution loss:

[0061]

[0062] Centroid trajectory loss:

[0063]

[0064] α and β are weighting coefficients, c pred (t,x,y,z), c true (t,x,y,z) represent the predicted and actual nuclear contamination concentrations, respectively; x pred (t), y pred (t),z pred (t) and x true (t), y true (t),z true (t) represents the predicted and actual centroid positions, respectively; Δt is the time step, and H·W·Z is the spatial network size.

[0065] Preferably, key monitoring areas for nuclear contamination are identified, and different data collection frequencies are set according to the expected rate of diffusion and monitoring needs; each collection ensures consistency of data conditions.

[0066] Use monitoring equipment to collect data from a predetermined monitoring area, ensuring that the location and parameters of the monitoring equipment are consistent, and mark and archive the collected data by date.

[0067] Beneficial effects: 1. This invention enhances the model's adaptability to complex terrain and dynamic environmental changes through customized data acquisition (such as island high-density buoys and ocean satellite remote sensing) and data augmentation (such as simulating island blocking effects and Gaussian noise).

[0068] 2. This invention introduces graph-structured modeling and cross-modal attention mechanisms to integrate multimodal data (such as ocean current velocity, temperature, and nuclide concentration) to enhance the predictive ability for complex marine environments; making it applicable to pollution diffusion simulation in different regions, and optimizing spatial resolution and local aggregation effects.

[0069] 3. This invention achieves high-precision prediction of marine nuclear pollution spread by fusing deep learning technology (ConvNeXtV2 and iTransformer modules) with multimodal data, and supports real-time updates, significantly improving the timeliness of emergency response.

[0070] 4. This invention introduces the centroid trajectory method of particle filtering and the concentration-priority matching algorithm, which can not only predict the concentration distribution, but also accurately track the diffusion path and trend, providing a more comprehensive scientific basis for risk assessment and environmental protection. Attached Figure Description

[0071] Figure 1 This is a flowchart of the marine nuclear pollution diffusion prediction method based on multimodal data fusion, as described in this invention.

[0072] Figure 2 This is a diagram of the ConvNeXtV2 network structure of the present invention;

[0073] Figure 3 This is a diagram of the global diffusion network structure based on the diffusion initialization network of this invention;

[0074] Figure 4 This is a diagram of the diffusion evolution network method of the present invention; Detailed Implementation

[0075] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one method for predicting the spread of marine nuclear pollution based on multimodal data fusion, or several specific implementations of this invention, and does not strictly limit the scope of protection specifically claimed by this invention.

[0076] The technical solution adopted in this invention is as follows: Figure 1 As shown, a method for predicting the spread of marine nuclear pollution based on multimodal data fusion includes the following steps:

[0077] S1: After obtaining marine environmental data and marine nuclear pollution data and preprocessing them, we obtain the initial dataset of marine nuclear pollution. The dataset is then divided into a training set and a test set.

[0078] S11: Customize the acquisition of marine environmental data and marine nuclear pollution data based on the characteristics of different sea areas and pollution sources, ensuring data coverage of multiple scenarios and pollution sources.

[0079] Marine environmental data: Ocean current velocity, temperature, and salinity are collected using buoys and acoustic Doppler current meters; sea surface wind speed and direction are obtained using ASCAT scatterometers; multispectral images are acquired using satellite remote sensing, and high-density buoys (one every 50km) are deployed in densely populated island areas to record depth data from 0 to 200m to capture changes in eddies and wind fields; for open ocean areas, satellite remote sensing is prioritized to cover day and night data.

[0080] Marine nuclear contamination data: Nuclide concentrations (0-200m depth) are collected using high-sensitivity gamma-ray detectors and underwater robots (ROVs). Samples are taken hourly near nuclear power plants, and the sampling frequency is adjusted according to the activity intensity in areas where nuclear submarines / aircraft carriers are active (once every 3 hours during activity periods and once a day during non-activity periods).

[0081] S12: Preprocess the dataset and interpolate and fill in cloud-obscured areas in the satellite remote sensing data, formula;

[0082]

[0083] Among them, t m t represents the current time step. prev Indicates the previous time step, t next Indicates the next time step, h m w represents length. m z represents the width. m Indicates depth;

[0084] Data augmentation: Simulate island blocking effect (ocean current velocity ±10%) for densely populated island areas; add Gaussian noise (η=0.05) for open ocean areas; simulate sudden high concentration areas (+50%) and concentration gradient changes in nuclear contamination data;

[0085] Normalization: Min-Max normalization was applied to the marine environmental data.

[0086]

[0087] Where x is the value of the original data point, x max and x min These are the minimum and maximum values ​​in the dataset, x. norm It is the normalized value;

[0088] The concentration of nuclear contamination was logarithmically transformed and then normalized using Min-Max.

[0089] c norm =log(1+c);

[0090] Where c is the original nuclear contamination concentration value, c norm It is the value after logarithmic transformation;

[0091] Time step encoding: Generates time embeddings for each time step to support subsequent time encoding in the iTransformer module; uses sine / cosine function form.

[0092]

[0093] Where TE(Δt,2i) and TE(Δt,2i+1) represent the components encoded at each time step, Δt is the time step, i is the encoding dimension index, and D is the encoding dimension (set to 64); a temporal coding tensor with shape (Δt,D) is generated.

[0094] S13: Split the training set and test set in a 7:3 ratio. Input shape (Δt,H,W,Z), label is concentration distribution (Δt,H,W,Z), and add time-series encoding (Δt,D).

[0095] S2: After graph structured modeling and cross-modal attention, ConvNeXtV2 is used to extract features and construct a diffusion initialization network based on nuclear pollution features. That is, by extracting the assigned features to generate feature maps, the initial state of nuclear pollution diffusion is initialized, providing a stable initial distribution for the subsequent diffusion evolution network.

[0096] refer to Figure 2 As shown:

[0097] S21: Introducing graph-structured modeling, nodes include marine environmental and nuclear pollution data, node features are parameter values ​​at corresponding spatiotemporal points (t, h, w, z), edges include connections between nodes of the same modality based on spatial proximity, and connections between nodes of different modalities reflecting the impact of ocean currents on pollution diffusion; the dynamically updated edge weight formula is:

[0098] w ij (t)=α·Corr(v i (t),c j (t))+β·Phys(v i (t),c j (t));

[0099] Among them, w ij (t) represents the dynamic edge weight between node i (marine environmental data) and node j (nuclear pollution data) at time t, v i (t) represents the characteristics of the marine environment, c j (t) represents the characteristics of nuclear contamination, Corr is the correlation between ocean current and concentration, Phys is based on the advection-diffusion equation, and α and β are learnable parameters;

[0100] Aggregate node features through graph convolution;

[0101]

[0102] Where σ is the ReLU activation function, N(i) represents the set of neighbors of node i, and W (l) This represents the learnable weight matrix of the l-th layer. This represents the characteristics of neighbor node j at layer l;

[0103] S22: Cross-Modal Attention is used to assign cross-modal attention weights to features, optimizing the fusion efficiency of multimodal data (such as ocean currents and nuclide concentrations). The formula is as follows:

[0104]

[0105] Among them, Q m Let K be the query vector for modality m. n V n Let d be the key and value vector of mode n. k Indicates the dimension of the key vector;

[0106] The assigned features are then input into the subsequent ConvNeXtV2;

[0107] S23: Use ConvNeXtV2 to extract features from the input data and generate a feature map with shape (T,H,W,Z,C). The formula for ConvNeXtV2 is:

[0108] Y = X + λ·Conv 1×1 (GELU(Conv 1×1 (GRN(LN(DC 7×7 (X))))));

[0109] Where GELU is the activation function, GRN is the global response normalization, LN is the layer normalization, and DC is the activation function. 7×7 It uses 7x7 depthwise separable convolution, with independent convolution for each channel to extract spatial features. GRN is used for global response normalization, λ is a learnable scaling factor, and the feature representation is adjusted channel by channel. Residual connections are used to ensure the stability of multimodal data and to handle the high dimensionality and noise characteristics of marine nuclear pollution data.

[0110] S3: Construct a diffusion evolution network. The diffusion evolution network mainly consists of a global diffusion module and two time-coded iTransformer modules. Based on S2, the diffusion evolution network is trained using a training dataset. At the same time, particle filtering is introduced to optimize the centroid trajectory and calculate the nuclear contamination diffusion centroid.

[0111] Reference Figure 3 , Figure 4 Specifically:

[0112] S31: Construct a diffusion evolution network, mainly composed of a global diffusion module and two time-coded iTransformer modules. These modules will be used to perform different stages of diffusion evolution and integrate time-coded and concentration-first matching algorithms.

[0113] The formula for iTransformer is:

[0114] Y = MLP(LN(H) L-1 +FFN(LN(H L-1 +Self-Attn(MLP embed (X T )))))) T ;

[0115] Where MLP stands for Multilayer Perceptron, LN for Layer Normalization, and H... L-1 The hidden state of layer L(L-1) represents the output of the previous layer. FFN stands for Feedforward Neural Network, Self-Attn represents Self-Attention, and MLP represents the hidden state of layer L(L-1). embed This represents the embedding layer of a multilayer perceptron, where X is the input;

[0116] First, the time series data of each variable is embedded into a D-dimensional token representation using a multilayer perceptron (MLP) to form the initial features. Where N is the number of tokens and D is the feature dimension; the time series of each variable is used as an independent token, the attention mechanism models multivariate correlations in the variable dimension, and the feedforward network (FFN) processes the time dimension;

[0117] S32: Input the initial distribution into the global diffusion module to obtain the first diffusion evolution result;

[0118] S33: Input the first diffusion evolution result, the feature map output by the feature extraction network ConvNeXt V2, and the time step encoding into the temporal coding iTransformer module to perform the second temporal coding diffusion evolution, and obtain the second diffusion evolution result;

[0119] S34: Input the second diffusion result and feature map into the second temporal coding iTransformer module to obtain the final diffusion prediction result;

[0120] The second diffusion result is reduced in dimensionality by 1×1 convolution and fed into the second iTransformer module along with the ConvNeXt V2 feature map. The hidden state dimension is set to 128 and the output shape is (T,H,W,Z,128). The iTransformer output is mapped back to the concentration distribution space through 1×1 convolution and transposed convolution to generate the final diffusion result. The second temporally encoded iTransformer module further optimizes the diffusion result and combines it with a concentration-first matching algorithm.

[0121] First, the concentration distribution is simplified using the Douglas-Peucker Algorithm to identify key points in high-concentration areas (grid points with concentration values ​​exceeding 0.8). At these key points, the diffusion trend of high-concentration areas is prioritized, and Euclidean distance is used to calculate the nearest neighbor matching between key points.

[0122]

[0123] Where p i ,p j and p k Here are the coordinates of the key point, (x i ,y i ),(x j ,y j ) and (z i -z j () represents the corresponding position;

[0124] A particle filtering algorithm is introduced to simulate the probability distribution of nuclear contamination diffusion using the Monte Carlo method. First, particle initialization is performed, with initial weights based on the nuclear contamination concentration distribution. Then, state updates are performed, updating particle positions based on environmental data such as ocean currents and wind speed.

[0125] x i (t+1)=x i (t)+v current (t)·Δt+η,

[0126] Where η is Gaussian noise, x(t) represents the state at time t, and v current (t) represents the rate of nuclear contamination diffusion at the current time t, where Δt is the time step;

[0127] Update the weights based on the nuclide decay model:

[0128] w i (t+1)=w i (t)·e -λΔt ·p(c i (t));

[0129] Where w(t) is the edge weight at time t, λ is the decay constant, and p(c i ) is the concentration likelihood function;

[0130] Resampling is performed based on particle weights, low-weight particles are removed and high-weight particles are retained. The centroid is calculated based on the particle distribution, and the centroid trajectory is post-processed using Kalman filtering to reduce jitter.

[0131] S35: Train the diffusion evolution network using the training dataset, introducing a multi-task loss function; calculate the deviation between the predicted concentration and the actual concentration, with the concentration distribution loss being:

[0132]

[0133] Calculate the centroid trajectory loss between the predicted centroid and the true centroid:

[0134]

[0135] Finally, calculate the total loss:

[0136] L total =αL concentration +βL centroid ;

[0137] Where α and β are weighting coefficients, which are defaulted to 0.5 and 0.5, respectively. pred (t,x,y,z), c true (t,x,y,z) represent the predicted and actual nuclear contamination concentrations, respectively; x pred (t), y pred (t),z pred (t) and x true (t), y true (t),z true (t) represents the predicted and actual centroid positions, respectively; Δt is the time step, and H·W·Z is the spatial network size;

[0138] S4: Input the test dataset into the trained diffusion evolution network of step S3, obtain the result of nuclear pollution diffusion prediction through three diffusion evolutions, and calculate the centroid trajectory;

[0139] The centroid trajectory is calculated for the nuclear contamination concentration distribution after each diffusion prediction, using the following formula:

[0140]

[0141] Where x i,j y i,j ,z i,j , where is the coordinate of network point (i,j), and c is the coordinate of the point (i,j). i,jLet t represent the nuclear contamination concentration value at network point (i,j), and t represent the stage of the diffusion prediction.

[0142] S5: Continuously acquire marine environmental and nuclear pollution data, execute step S4, obtain nuclear pollution diffusion prediction results at different times, generate and analyze prediction maps, compare prediction results, and determine diffusion evolution trends.

[0143] S51: Identify key monitoring areas for nuclear contamination and set different data collection frequencies based on the expected rate of spread and monitoring needs; ensure consistency of data conditions for each collection.

[0144] S52: Use monitoring equipment to collect data from pre-determined monitoring areas to ensure that the data accurately reflects subtle changes in nuclear contamination; keep the location and parameters of the monitoring equipment consistent, and mark and archive the collected data by date for easy tracking and retrieval;

[0145] S53: Perform step S4 on each collected data to predict nuclear contamination spread and generate a prediction map;

[0146] S54: Based on continuous monitoring data and forecast maps, assess the evolution trend of nuclear contamination spread, determine whether there are signs of accelerated spread, and organize further on-site verification and emergency response if a significant expansion of the spread range or the emergence of new high-concentration areas are found.

[0147] The predicted distribution of nuclear contamination diffusion in the region is shown in the plots before and after one time step. The left side represents the initial diffusion state, and the right side represents the evolved diffusion state. In both plots, the gray areas represent the nuclear contamination concentration distribution, with the concentration gradually increasing from light gray to dark gray, indicating that the diffusion range and intensity expand over time. The data shows that the initial high-concentration area expands after evolution, with the peak concentration increasing from 0.8 to 1.2.

[0148] As discussed above, traditional methods for predicting marine nuclear pollution primarily rely on numerical simulation and statistical analysis, which suffer from problems such as long computation time, low prediction accuracy, and difficulty in real-time response. Furthermore, deep learning methods based on single data sources have limitations in adapting to complex marine environments. This invention proposes a marine nuclear pollution diffusion prediction method based on multimodal data fusion. By fusing multi-source marine environmental data and a centroid trajectory optimization algorithm, it achieves high-precision, real-time prediction of nuclear pollution concentration distribution and diffusion paths. This invention utilizes customized collection of marine environmental data (such as ocean current velocity, temperature, and salinity) and nuclear pollution data (such as nuclide concentration), combined with preprocessing operations such as data cleaning, enhancement, and temporal coding, to achieve efficient fusion of multimodal data. This significantly improves the accuracy and robustness of the prediction results, outperforming single-data-source models. By employing graph-structured modeling and a cross-modal attention mechanism to integrate multimodal data, it dynamically captures the spatiotemporal interaction between ocean currents, wind fields, and nuclide concentrations, optimizing spatial resolution and local clustering effects, thus enhancing adaptability to complex marine environments. To address the shortcomings of traditional deep learning models like CNN and LSTM in capturing temporal features and describing overall movement trends, this invention introduces ConvNeXt V2 to extract high-dimensional spatial features, utilizes the iTransformer module combined with temporal coding to model the time dimension, and optimizes the centroid trajectory through particle filtering and Kalman filtering, significantly improving the ability to capture dynamic evolution trends of diffusion and the accuracy of path prediction. These techniques effectively solve the problems of poor adaptability to single data sources, low prediction accuracy, and insufficient capture of temporal features, providing a scientific basis and technical support for on-site verification and emergency response to marine nuclear pollution diffusion prediction.

[0149] The embodiments of the present invention have been described in detail above with reference to the examples. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A method for predicting the spread of marine nuclear pollution based on multimodal data fusion, characterized in that, Includes the following steps: A marine nuclear pollution dataset is constructed and divided into a training set and a test set. The marine nuclear pollution dataset is time-step encoded, and a time-series code is generated for each time step. The marine nuclear pollution dataset includes marine environmental data and marine nuclear pollution data. The marine environmental data includes ocean current velocity, temperature, salinity, sea surface wind speed, wind direction, and multispectral images; for densely populated island areas, data at depths of 0-200m are recorded to capture changes in eddies and wind fields; for open ocean areas, satellite remote sensing is used to cover day and night data. The marine nuclear contamination data includes the collection of nuclide concentrations using high-sensitivity gamma-ray detectors and underwater robots. Samples are taken hourly near nuclear power plants, every 3 hours during the activity period of nuclear submarines / aircraft carriers, and daily during non-activity periods. Based on a marine nuclear pollution dataset, graph-structured modeling is introduced, node features are aggregated through graph convolution, and cross-modal attention weights are assigned to the features; The assigned features are input into ConvNeXt V2 for feature extraction to generate feature maps, and a diffusion initialization network based on nuclear contamination features is constructed. A diffusion evolution network was constructed. Based on the diffusion initialization network, combined with continuously acquired marine environmental data and marine nuclear pollution data, the nuclear pollution diffusion prediction results at different times were obtained through three diffusion evolutions, a prediction map was generated, and the centroid trajectory of the nuclear pollution concentration distribution after each diffusion prediction was calculated. as well as Analyze the prediction map, compare the prediction results, and determine the diffusion evolution trend; A particle filter algorithm is introduced to simulate the probability distribution of nuclear contamination diffusion using the Monte Carlo method: Particle initialization is performed, with initial weights based on the nuclear contamination concentration distribution; Perform state updates; update particle positions based on ocean environmental data. ； Where η is Gaussian noise, This represents the state at time t. This represents the rate of nuclear contamination spread at the current time t. For time step; Update the weights based on the nuclide decay model: ； in, Let λ be the edge weight at time t, and λ be the decay constant. Let be the concentration likelihood function; Resampling is performed based on particle weights, removing low-weight particles and retaining high-weight particles. The centroid is calculated based on particle distribution, and the centroid trajectory is calculated by combining Kalman filtering with the nuclear contamination concentration distribution after each diffusion prediction. ； in , , Let (i,j) be the coordinates of the network point. Let t represent the nuclear contamination concentration value at network point (i,j), and t represent the stage of the diffusion prediction.

2. The method for predicting marine nuclear pollution spread based on multimodal data fusion according to claim 1, characterized in that: Preprocessing of the marine nuclear contamination dataset includes: Interpolate and fill in cloud-obscured areas in satellite remote sensing data; Data augmentation: Simulate island blocking effects in densely populated island areas; add Gaussian noise to open ocean areas; simulate sudden high-concentration areas and concentration gradient changes in nuclear pollution data; Min-Max normalization was used for marine environmental data; The marine nuclear pollution data were subjected to logarithmic transformation and then Min-Max normalization.

3. The method for predicting marine nuclear pollution spread based on multimodal data fusion according to claim 1, characterized in that: In graph-structured modeling, nodes include marine environment node i and nuclear pollution data node j. Node features are the parameter values ​​of the corresponding spatiotemporal point (t, h, w, z). Edges include connections between nodes of the same modality based on spatial proximity, and connections between nodes of different modalities reflecting the impact of ocean currents on pollution diffusion. The formula for dynamically updating edge weights is: ； in, This represents the dynamic edge weight between node i and node j at time t. Indicates the characteristics of the marine environment. Indicates characteristics of nuclear contamination. The correlation between ocean currents and concentration. It is based on the advection-diffusion equation. , These are learnable parameters; Graph convolution aggregation node features: ； in, It is the ReLU activation function. Let i represent the set of neighbors of node i. This represents the learnable weight matrix of the l-th layer. This represents the characteristics of neighbor node j at layer l.

4. A method for predicting marine nuclear pollution spread based on multimodal data fusion according to claim 1 or 3, characterized in that: ConvNeXt V2 is used to extract features from the input data and generate feature maps with shapes (T, H, W, Z, C). The formula for ConvNeXt V2 is: ； Where GELU is the activation function, GRN is the global response normalization, and LN is the layer normalization. It uses a 7x7 depthwise separable convolution, performing independent convolution on each channel to extract spatial features. For global response normalization, As a learnable scaling factor, the feature representation is adjusted channel by channel, and residual connections are used to ensure the stability of multimodal data and to handle the high dimensionality and noise characteristics of marine nuclear pollution data.

5. The method for predicting marine nuclear pollution spread based on multimodal data fusion according to claim 1, characterized in that: The diffusion evolution network includes a global diffusion module and two time-coded iTransformer modules; The three diffusion evolutions include: Based on the initial distribution input into the global diffusion module, the first temporal coding diffusion evolution is performed to obtain the first diffusion evolution result; Based on the results of the first diffusion evolution, the feature map and time step encoding are input into the first temporal coding iTransformer module to perform the second temporal coding diffusion evolution, and the second diffusion evolution result is obtained. Based on the results of the second diffusion evolution and the feature map input into the second temporal coding iTransformer module, a third temporal coding diffusion evolution is performed to obtain the final diffusion prediction result.

6. The method for predicting marine nuclear pollution spread based on multimodal data fusion according to claim 5, characterized in that: In the second timing-coded iTransformer module: The Douglas-Puk algorithm is used to simplify the concentration distribution and filter out key points in high-concentration areas. At key points, priority is given to matching the diffusion trend of high-concentration areas, and Euclidean distance is used to calculate the nearest neighbor matching between key points.

7. The method for predicting marine nuclear pollution spread based on multimodal data fusion according to claim 1, characterized in that: For the diffusion evolution network, training is performed using a training dataset, and a multi-task loss function is introduced; the deviation between the predicted concentration and the actual concentration is calculated. The multi-task loss function is: ;in: Concentration distribution loss: ； Centroid trajectory loss: ； , These are the weighting coefficients. , These represent predicted and actual nuclear contamination concentrations, respectively. , , and , , These represent the predicted and actual centroid positions, respectively. For time step, This refers to the size of the spatial network.

8. The method for predicting marine nuclear pollution spread based on multimodal data fusion according to claim 1, characterized in that: Identify key monitoring areas for nuclear contamination and set different data collection frequencies based on the expected rate of spread and monitoring needs; Ensure data consistency in each data collection; Use monitoring equipment to collect data from a predetermined monitoring area, ensuring that the location and parameters of the monitoring equipment are consistent, and mark and archive the collected data by date.

Images