Intelligent Fusion of Multi-Source Heterogeneous Marine Data and Marine Disaster Prediction Method and Platform

Abstract

The present application relates to the technical field of data processing, and discloses a method and platform for intelligent fusion of multi-source heterogeneous marine data and marine disaster prediction. The method includes: hierarchically deploying and constructing a network topology for marine environment monitoring devices, communication relay devices, and data processing devices to obtain an integrated sea-air-shore monitoring network; collecting marine environment data with spatio-temporal identifiers; performing marine element modeling and dynamic feature extraction to obtain a marine environment digital model; performing feature fusion and correlation analysis on multi-source heterogeneous marine data to obtain spatio-temporal feature data; using transfer learning methods to train and integrate multiple models for typhoon path prediction models, storm surge prediction models, and red tide prediction models to obtain a marine disaster prediction model set, thereby improving the prediction accuracy and timeliness, optimizing the allocation and scheduling of computing resources, and meeting the real-time requirements for processing massive data.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and in particular, to a method and platform for intelligent fusion of multi-source heterogeneous marine data and marine disaster prediction. Background Art

[0002] Traditional marine disaster prediction methods mainly rely on single data sources and simple numerical models, making it difficult to accurately grasp the complex dynamic change characteristics of the marine environment, resulting in insufficient prediction accuracy and timeliness.

[0003] Currently, the data generated by marine monitoring systems has characteristics such as multi-source heterogeneity, uneven spatio-temporal distribution, and severe noise interference. How to effectively process and fuse this data has become a key challenge in marine disaster prediction. At the same time, there are complex interaction relationships between different types of marine disasters, and a single prediction model is difficult to accurately depict the multi-scale dynamic characteristics during the disaster evolution process. In addition, the real-time processing of massive multi-source heterogeneous data places higher requirements on computing resources, and existing data processing methods lack effective resource scheduling and optimization mechanisms, making it difficult to meet the dual requirements of real-time and accuracy for marine disaster early warning. Summary of the Invention

[0004] This application provides a method and platform for intelligent fusion of multi-source heterogeneous marine data and marine disaster prediction, thereby improving prediction accuracy and timeliness, optimizing the allocation and scheduling of computing resources, and meeting the real-time requirements for massive data processing.

[0005] In the first aspect of this application, a method for intelligent fusion of multi-source heterogeneous marine data and marine disaster prediction is provided. The method for intelligent fusion of multi-source heterogeneous marine data and marine disaster prediction includes:

[0006] Hierarchically deploy and construct a network topology for marine environment monitoring devices, communication relay devices, and data processing devices to obtain an integrated sea-air-shore monitoring network;

[0007] According to the integrated sea-air-shore monitoring network, use a periodic deterministic sampling algorithm to collect and standardize water temperature, salinity, sea level height, wind speed, and air pressure data to obtain marine environment data with spatio-temporal identifiers;

[0008] Based on the marine environment data with spatio-temporal identifiers, perform marine element modeling and dynamic feature extraction to obtain a marine environment digital model;

[0009] Based on the marine environment digital model, use a selective caching mechanism and a deep learning algorithm to perform feature fusion and correlation analysis on multi-source heterogeneous marine data to obtain spatio-temporal feature data;

[0010] According to the spatio-temporal feature data, a transfer learning method is used to train and integrate multiple models for the typhoon path prediction model, storm surge prediction model, and red tide prediction model, to obtain a set of ocean disaster prediction models.

[0011] The second aspect of this application provides a multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction platform, which includes:

[0012] A construction module, used to hierarchically deploy and construct a network topology for ocean environment monitoring devices, communication relay devices, and data processing devices, to obtain an air-sea-shore integrated monitoring network;

[0013] An acquisition module, used to collect and standardize water temperature, salinity, sea level height, wind speed, and air pressure data according to the air-sea-shore integrated monitoring network, using a periodic deterministic sampling algorithm, to obtain ocean environment data with spatio-temporal identifiers;

[0014] A feature extraction module, used to perform ocean element modeling and dynamic feature extraction based on the ocean environment data with spatio-temporal identifiers, to obtain an ocean environment digital model;

[0015] An association analysis module, used to perform feature fusion and association analysis on multi-source heterogeneous ocean data based on the ocean environment digital model, using a selective caching mechanism and a deep learning algorithm, to obtain spatio-temporal feature data;

[0016] A model integration module, used to train and integrate multiple models for the typhoon path prediction model, storm surge prediction model, and red tide prediction model according to the spatio-temporal feature data, using a transfer learning method, to obtain a set of ocean disaster prediction models.

[0017] The third aspect of this application provides an electronic device, including: a memory and at least one processor, where instructions are stored in the memory; the at least one processor calls the instructions in the memory, so that the electronic device executes the above multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method.

[0018] The fourth aspect of this application provides a computer-readable storage medium, where instructions are stored in the computer-readable storage medium, and when it runs on a computer, it causes the computer to execute the above multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method.

[0019] Compared with the prior art, the present application has the following beneficial effects: By constructing an integrated sea-air-shore monitoring network, the unified collection and transmission of multi-source heterogeneous ocean data are realized, solving the problems of scattered data acquisition and inconsistent formats. The periodic deterministic sampling algorithm and data standardization processing are adopted to effectively reduce the noise interference in the data acquisition process, ensuring data quality and timeliness; Based on digital twin technology, a digital model of the marine environment is established, accurately depicting the dynamic correlation relationships among marine environmental elements and enhancing the cognitive ability of the laws of marine environmental changes; The combination of the selective caching mechanism and deep learning algorithm optimizes the usage efficiency of computing resources while improving the accuracy of feature fusion; Through the transfer learning method and multi-model integration technology, the knowledge of pre-trained models is fully utilized, enhancing the generalization performance of different types of marine disaster prediction models; A complete model training and optimization process is established, realizing the full-process intelligent processing from data acquisition to disaster prediction, significantly improving the prediction accuracy and timeliness; The distributed computing architecture and edge computing technology are adopted to optimize the allocation and scheduling of computing resources, meeting the real-time requirements for processing massive data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0021] The structures, proportions, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those familiar with this technology to understand and read, and are not used to limit the limiting conditions under which the present invention can be implemented. Therefore, they do not have technical essential meanings. Any modification of the structure, change of the proportional relationship or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope that can be covered by the technical content disclosed in the present invention.

[0022] Figure 1 is a schematic flowchart of a method for intelligent fusion of multi-source heterogeneous ocean data and marine disaster prediction provided by an embodiment of the present invention;

[0023] Figure 2 is a schematic block diagram of the structure of a platform for intelligent fusion of multi-source heterogeneous ocean data and marine disaster prediction provided by an embodiment of the present invention;

[0024] Figure 3 is a schematic block diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0026] The flowcharts shown in the accompanying drawings are only illustrative examples, and do not necessarily include all the content and operations / steps, nor do they necessarily need to be executed in the described order. For example, some operations / steps can also be decomposed, combined, or partially merged, so the actual execution order may change according to the actual situation.

[0027] It should also be understood that the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. As used in the specification of this application and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0028] It should be further understood that the term "and / or" used in the specification of this application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Please refer to Figure 1 , an embodiment of the multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method in the embodiments of this application includes:

[0029] Step 100: Hierarchically deploy and construct a network topology for ocean environment monitoring devices, communication relay devices, and data processing devices to obtain an air-sea-shore integrated monitoring network;

[0030] It can be understood that the execution subject of this application can be a multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction platform, or a terminal or a server. Specifically, it is not limited here. In the embodiments of this application, the server is taken as an example of the execution subject for illustration.

[0031] Specifically, the monitoring area is divided into grids at three levels: ocean, air, and shore-based, to form monitoring grid units with clear spatial attributes. By analyzing the spatial distribution characteristics of the monitoring grid units, the monitoring density requirements under different environmental conditions are evaluated, and the layout plan of the monitoring equipment is optimized, so that the equipment can cover the largest area under resource constraints and meet the data collection requirements. After the equipment is deployed, the functional requirements of each equipment are refined and allocated, and its operations are divided into a data collection module, a data preprocessing module, and a communication module to achieve the efficient operation of the equipment functions. Through the analysis of the specific uses of the equipment, the data collection module is used to collect environmental parameters such as water temperature, salinity, and air pressure in real time, the data preprocessing module cleans and formats the raw data, and the communication module is used for data uploading and coordination of communication between devices to ensure the stability and efficiency of the entire monitoring system. According to the functional allocation information of the equipment, it is input into the network topology generator, and the preliminary planning of the network is completed through a multi-level topology construction algorithm to generate a hierarchical network topology structure diagram covering the ocean, air, and shore-based. Based on the network topology structure diagram, the communication relay equipment is deployed in different regions and the communication protocol is configured to ensure smooth data transmission between different levels. At the ocean level, the communication relay equipment includes the wireless transmission device of the sea surface buoy and the communication station of the unmanned ship; at the air level, drones and satellites are used as efficient data relay tools; while at the shore-based level, it mainly relies on fixed communication stations or coastal stations for data aggregation. To improve the transmission efficiency, a hierarchical communication protocol is adopted. Short-distance, high-frequency transmission technology is used between the underlying communication relay devices, and a satellite communication protocol with high stability and broadband is used for long-distance transmission to achieve data synchronization and efficient sharing across the network. At the same time, to improve the immediacy of data processing and distributed processing capabilities, lightweight prediction models are deployed on the edge computing servers in the data transmission network to form local data processing units. These local units mainly perform preliminary processing and analysis on the data collected by the monitoring equipment, reduce the bandwidth requirements for data transmission, and provide rapid prediction results to support real-time decision-making. For tasks that require higher computing power and data integration capabilities, the data fusion model of the cloud computing center is decomposed into tasks and allocated computing resources according to the principle of distributed computing, and works in coordination with the local data processing units through network connections to form a central computing and processing unit. The responsibility of the central unit is to globally integrate the scattered data from each monitoring area and complete more complex analysis tasks, including the intelligent fusion of multi-source data and the training and update of disaster prediction models. The data transmission network, local data processing units, and central computing and processing units are integrated through a unified system integration and communication protocol specification to form a highly collaborative and scalable integrated sea-air-shore monitoring network.This network can support multi-level monitoring and processing functions, coordinate and complete data collection, transmission, processing, and analysis tasks at different levels, and provide a full-process service from raw data to prediction results.

[0032] Step 200: According to the sea-air-shore integrated monitoring network, use the periodic deterministic sampling algorithm to collect and standardize water temperature, salinity, sea level height, wind speed, and air pressure data to obtain ocean environmental data with spatio-temporal identifiers.

[0033] Specifically, in the sea-air-shore integrated monitoring network, the working time slots of marine environmental monitoring equipment are periodically divided. By evaluating the functions, data importance, and resource constraints of different equipment, the sampling tasks are prioritized to generate a sampling scheduling sequence. This sequence can optimize the data acquisition process of monitoring equipment, enabling high-priority data to be preferentially collected during critical periods, while reasonably arranging low-priority tasks to ensure the full utilization of system resources and the balance of data coverage. According to the generated sampling scheduling sequence, error calculation and threshold judgment are performed on the original data collected by the monitoring equipment to evaluate the credibility of the data. By introducing error calculation methods, such as statistical analysis based on historical mean deviation, equipment accuracy range, and data distribution law, the quality of each piece of data is accurately evaluated, and a data credibility index is assigned to it. For records with data credibility lower than the preset threshold, they are input into the anomaly processing module for marking and elimination to avoid interference from abnormal data to subsequent analysis and ensure the quality reliability of the effective data set. The anomaly processing module can adopt the combination of rule detection and machine learning methods. For example, by setting reasonable boundary conditions to filter out outliers and using clustering or classification algorithms to further identify potential anomalies. Based on the time distribution characteristics of the effective data set, the data sampling period is dynamically adjusted and synchronized to enable the data acquisition task to better adapt to the dynamic changes of the marine environment. For example, when monitoring emergencies such as typhoons or storm surges, the sampling period is shortened to improve the timeliness of data, while in a stable state, the sampling interval is appropriately extended to save resources. The dynamic adjustment mechanism is realized through an adaptive control algorithm, which can not only flexibly respond to environmental changes but also reduce the energy consumption of the equipment. After completing the time sequence synchronization, the calibrated time sequence data is cleaned, redundant or repeated records are eliminated, and the data is converted into a standard format through normalization and numerical quantization processing. The normalization process maps data with different dimensions to a unified range, thus eliminating the influence of dimension differences on model input, while numerical quantization helps to improve the storage efficiency and calculation convenience of data. According to the spatial distribution of the monitoring equipment, a geographical location code is assigned to each data record, and a timestamp is marked with the time information collected by the equipment to form a basic data set with spatio-temporal correlation. Based on the spatio-temporal correlated data, the data change rate and transmission delay between adjacent sampling points are calculated to quantitatively describe the dynamic and transmission characteristics of the data. The calculation of the data change rate is achieved through incremental change measurement on a continuous time series, reflecting the dynamic change trend of environmental elements, while the transmission delay is measured by analyzing the transmission time of data from the equipment to the processing center, reflecting the real-time nature of data acquisition and processing. The spatio-temporal correlated data and its additional timeliness parameters are encapsulated into data packets and encoded with a transmission protocol for efficient transmission in the network. The encapsulation process includes packing the data records according to certain priorities and network protocol requirements, and adding necessary check information and routing information to improve the reliability and integrity of transmission.Through the above steps, the marine environmental data with spatio-temporal identifiers are finally obtained.

[0034] Step 300: Based on the marine environmental data with spatio-temporal identifiers, conduct marine element modeling and dynamic feature extraction to obtain a digital marine environmental model;

[0035] It should be noted that the marine environmental data with spatio-temporal identifiers are classified according to different physical elements, and the water temperature field, salinity field, sea level height field, wind speed field and air pressure field are respectively constructed. Based on this, digital encoding is carried out for each type of physical element data to generate a unified data structure. In this process, combined with the physical field theory and data distribution law, standardized formats and coding rules are used to transform complex physical environment information into a structured marine element data structure. The physical quantities in the above data structure are mapped in time series and spatial distribution. By constructing a spatio-temporal distribution model, the variation laws of marine environmental elements in different time and space ranges are revealed. The time series mapping adopts the dynamic time warping algorithm or Fourier transform technology to extract long-term trends and periodic characteristics, while the spatial distribution mapping constructs a high-resolution spatial distribution map through interpolation methods and zonal modeling to achieve an accurate description of the spatio-temporal dynamics of the marine environment. Through this process, the spatio-temporal distribution characteristics of marine environmental elements are described. Based on the spatio-temporal distribution characteristics, the coupling analysis and correlation degree calculation of the interaction relationships between marine environmental elements are carried out. The coupling analysis is based on the hydrodynamic theory and statistical models. By studying the linkage relationships between elements such as water temperature and salinity, wind speed and air pressure, a correlation matrix reflecting the physical mechanism is constructed. The calculation of the correlation degree uses methods such as mutual information, Granger causality or correlation coefficient to quantify the coupling strength between different elements and generate a marine element correlation matrix that can reflect the internal interactions of complex systems. The above correlation matrix is input into the dynamic evolution model to analyze the variation laws of the marine environmental state. The dynamic evolution model simulates the evolution process of the marine environment with time and space changes by using differential equation systems, state space models or data-driven machine learning algorithms, and generates an environmental evolution sequence. After obtaining the environmental evolution sequence, stability analysis and mutation feature detection are carried out to identify abnormal events in the marine environment. For example, through methods such as principal component analysis and singular value decomposition, the existing system mutations are detected, and combined with the physical field model, the spatio-temporal distribution and influence range of abnormal events are located to generate complete marine environmental abnormal event characteristics. Based on these abnormal event characteristics, a mapping function between the physical ocean environment and the digital space is established to realize the mapping from the physical environment to the digital model. This process is the digital twin mapping. The digital twin mapping associates physical elements with digital features in the virtual space by constructing multi-dimensional mapping relationships to ensure that the virtual environment can accurately reflect the actual physical state. The digital twin mapping relationship is input into the feature extractor to extract the key features that are most sensitive to the environmental state and dynamic changes, and generate a highly condensed marine environmental feature vector. Model integration and parameter optimization are carried out for the marine environmental feature vector and the digital twin mapping relationship. Model integration adopts ensemble learning methods (such as Boosting or Bagging) to synergistically optimize multiple modeling algorithms, while parameter optimization adjusts model parameters through techniques such as particle swarm optimization or Bayesian optimization to improve the prediction accuracy and generalization ability.In this process, by combining the digital twin model with the spatio-temporal distribution characteristics, data from different sources and with different structures are comprehensively integrated to generate a unified digital model of the ocean environment.

[0036] Step 400: Based on the digital model of the ocean environment, a selective caching mechanism and a deep learning algorithm are used to perform feature fusion and correlation analysis on multi-source heterogeneous ocean data to obtain spatio-temporal feature data;

[0037] Specifically, in the digital model of the marine environment, according to the characteristics of different marine environmental elements, such as water temperature, salinity, sea level height, wind speed, and air pressure, etc., based on their data types (such as continuous, discrete, or categorical) and different patterns of spatio-temporal distribution, these characteristics are systematically classified and stratified to construct the hierarchical structure of feature data. On this basis, by analyzing the usage frequency, correlation importance, and update period of historical data characteristics, a selective caching mechanism is established to manage data caching in a hierarchical and prioritized manner, generating a hierarchical caching database. Through the caching mechanism, the data access latency is effectively reduced, and at the same time, the processing efficiency of high-priority data is improved. The feature data in the hierarchical caching database is input into a deep neural network for processing to extract potential feature representations and complete non-linear mapping. The deep neural network can capture the high-order relationships and complex patterns hidden in the feature data, especially non-linear characteristics, through a multi-layer abstract learning process, thereby generating deep feature representations. These feature representations include spatial features (such as geographical location and regional distribution), temporal features (such as periodic changes and trends), and physical features (such as hydrodynamics characteristics and atmospheric circulation characteristics), which together constitute a multi-dimensional description of multi-source data. After the deep feature representations are generated, multi-modal feature fusion and complementary analysis are performed on the spatial, temporal, and physical features among them. This process uses multi-modal learning methods, through techniques such as joint distribution modeling, collaborative representation, and feature alignment, to give full play to the complementary role between features of each modality, obtaining fused feature vectors. These vectors represent the internal connections and synergistic effects between different features, enabling subsequent analysis to be more comprehensive and accurate. The fused feature vectors are input into an attention calculation unit to perform feature weight assignment and significance calculation. The attention mechanism dynamically evaluates the importance of different features, assigns appropriate weights to each feature, and generates a weighted feature set. The weighted feature set is organized according to the time dimension and the space dimension, and an indexing mechanism is established to form a structured spatio-temporal feature organization system. Similarity measurement and correlation mining are performed on the feature data in the spatio-temporal feature organization system. The correlation between different features is identified through similarity measurement, while correlation mining can reveal the potential complex relationships between features. This process generates a feature correlation map, describing the interaction network of multi-source data in the spatio-temporal dimension. The feature correlation map is input into a deep spatio-temporal fusion network for cross-modal feature learning and spatio-temporal dependence modeling. The deep spatio-temporal fusion network combines the advantages of convolutional neural network (CNN) and recurrent neural network (RNN), can capture the local patterns of spatial features, and model the dependencies in the time series, thereby generating a high-dimensional fused feature tensor. Dimensionality reduction processing and feature selection are performed on the fused feature tensor. Dimensionality reduction processing reduces feature redundancy through techniques such as principal component analysis, singular value decomposition, or autoencoders, and retains the most representative features.Meanwhile, feature selection is based on importance scores or contribution evaluations to select the most relevant features and generate the final spatio-temporal feature data.

[0038] Step 500: According to the spatio-temporal feature data, use transfer learning methods to train and integrate multiple models for the typhoon track prediction model, storm surge prediction model, and red tide prediction model to obtain a set of ocean disaster prediction models.

[0039] Specifically, according to the content of spatiotemporal feature data, it is divided into typhoon path prediction data set, storm surge prediction data set and red tide prediction data set, corresponding to different disaster types and prediction needs. In this process, the spatiotemporal characteristics and physical characteristics of the data are combined to ensure that each data set can fully cover the key elements of the target phenomenon. Based on the typhoon path prediction data set, a typhoon path prediction model with an encoder-decoder structure as the core is constructed. The encoder network of the model consists of a 5-layer convolution structure. Each convolution layer uses a ReLU activation function to enhance the nonlinear expression ability, and is equipped with a batch normalization layer to stabilize the training process and accelerate convergence. These convolution layers extract the motion characteristics of the typhoon path, and gradually capture the spatial distribution law and dynamic change trend of the typhoon through layer-by-layer convolution operations. The corresponding decoder network also consists of a 5-layer deconvolution structure, which can generate the prediction results of the typhoon path. In order to avoid feature loss and enhance information transfer, jump connections are added between the encoder and decoder to ensure that high-resolution features can be directly transferred to the decoding stage to generate the initial prediction results of the typhoon path. For the storm surge prediction dataset, a storm surge prediction model is constructed. The model consists of two parts: a temporal feature extraction network and a spatial feature fusion network. The temporal feature extraction network adopts a bidirectional long short-term memory network (Bi-LSTM) structure, which can extract the previous and next dependencies from the time series of tidal level changes and capture the dynamic change pattern and long-term trend of the tidal level. At the same time, the spatial feature fusion network adopts a graph convolutional network (GCN) structure. By performing correlation analysis on multi-point tidal level data, a spatial distribution map of tidal level characteristics in the region is constructed to explore the linkage effect between different regions. Through the synergy of these two sub-networks, the model can generate initial prediction results for storm surges. Based on the red tide prediction dataset, a red tide prediction model is constructed. The model adopts a structure that combines a multi-scale feature extraction module and a spatiotemporal attention module to deal with the complex mechanism and variability of red tide occurrence. In the multi-scale feature extraction module, three parallel residual networks are used, each of which focuses on feature extraction at different scales to ensure that the model can simultaneously capture local and global information of red tide occurrence. The spatiotemporal attention module focuses on analyzing key areas and time periods, and uses the attention mechanism to dynamically adjust the degree of attention to different data to generate initial prediction results for red tides. After all initial prediction models are built, they are input into the transfer learning network together with the pre-trained marine disaster prediction model library for knowledge transfer and model optimization. The transfer learning network contains a feature mapping layer and a task adaptation layer. The feature mapping layer transfers the knowledge of the pre-trained model to the data distribution of the current task, while the task adaptation layer further adjusts the model structure to adapt to the current task requirements. Through transfer learning, the model can make full use of the existing marine disaster prediction experience, optimize the initial prediction results of typhoon paths, storm surges, and red tides, and generate a more accurate set of optimized prediction results.Perform uncertainty analysis and confidence calculation on the optimized prediction result set. By analyzing the sources of uncertainty in model predictions, such as insufficient data or model bias, calculate the confidence interval for each prediction result to quantify the reliability of the prediction. Input the prediction results containing the confidence interval into the ensemble learning module for multi-model fusion and weight optimization. The ensemble learning module synthesizes the prediction results of multiple models through methods such as weighted voting and stacking generalization, and uses weight optimization techniques to dynamically adjust the contribution of each model to generate a more robust ensemble prediction result. On this basis, perform model integration and parameter calibration on the ensemble prediction result and the optimized prediction result set to improve the prediction accuracy and applicability, and finally obtain an ocean disaster prediction model set that can comprehensively cover the prediction requirements for three types of disasters: typhoon path, storm surge, and red tide.

[0040] In this embodiment, the weight parameters in the pre-trained ocean disaster prediction model library are input into the feature mapping layer of the transfer learning network. The feature mapping layer constructs a shared feature space through general feature analysis. This shared space aims to capture the common features of three disaster prediction tasks: typhoon path, storm surge, and red tide, ensuring that the existing knowledge of the pre-trained model can be maximally utilized during the subsequent model optimization process. Based on the shared feature space, a feature vector mapping is performed on the initial typhoon path prediction result. The feature mapping layer adopts a three-layer fully connected network structure for this task. Each layer of the fully connected network processes non-linear features through the LeakyReLU activation function, which can effectively avoid the problem of gradient disappearance and maintain sensitivity to the feature distribution. Through this mapping method, the mapped features of the typhoon path are generated, which are characterized by retaining the core laws of the path dynamics and being closely aligned with the shared feature space, providing a unified feature representation for subsequent cross-task transfer. A feature vector mapping is performed on the initial storm surge prediction result. The feature mapping layer adopts a three-layer convolutional network enhanced by an attention mechanism for the spatio-temporal correlation characteristics of the storm surge. Each layer of the convolutional network dynamically adjusts the attention degree to different feature channels through the attention mechanism, and combines batch normalization and the ReLU activation function to achieve the stability and efficiency of the training process. This structure can not only extract the spatio-temporal distribution features in the storm surge water level data, but also strengthen the key region features through the attention mechanism, making the storm surge mapped features have higher expressiveness and discrimination ability. For the initial red tide prediction result, the feature mapping layer adopts a three-layer graph neural network (GNN) structure. Each layer of the GNN aggregates the red tide-related features through the GraphSAGE aggregation function to generate the red tide mapped features. The GraphSAGE function can utilize the neighborhood information during the node feature aggregation process and is suitable for dealing with phenomena such as red tide with complex spatial distributions and mutual influence characteristics. After completing the feature mapping, the typhoon path mapped features, storm surge mapped features, and red tide mapped features are input into the task adaptation layer for cross-task knowledge transfer. The task adaptation layer designs a structure of a generator and a discriminator to compete through an adversarial learning mechanism. The generator generates a unified transfer feature representation, while the discriminator guides the generator to optimize the transfer process by distinguishing the feature distributions of different tasks. This mechanism effectively reduces the distribution difference between the source task and the target task, enables the transfer feature representation to be shared among multiple tasks, and retains the key features of each task. Task relevance analysis and domain adaptation training are performed on the transfer feature representation. The adversarial discriminator structure in the task adaptation layer discriminates the distribution of the transfer features, identifies the features specifically related to each task, and further optimizes the distribution of these features to better adapt to the data characteristics of the target task. Through this process, task-specific features are obtained. These features are the exclusive high-dimensional representations of each task and simultaneously possess the shared knowledge required during the transfer learning process. Based on the task-specific features, fine-grained tuning and gradient update are performed on the parameters in the pre-trained ocean disaster prediction model library.In this process, the transfer learning network optimizes the weight parameters of the pre-trained model through backpropagation to adapt to the current data distribution and prediction requirements. During the tuning process, an adaptive gradient descent algorithm (such as Adam) is used to accelerate parameter updates, and a learning rate decay strategy is utilized to ensure the stability of the optimization process. Through parameter updates and model reconstruction, an optimized prediction result set is generated. This result set combines the rich knowledge of the pre-trained model and the specific data characteristics of the target task, significantly improving the prediction accuracy and robustness of typhoon tracks, storm surges, and red tides.

[0041] In the embodiments of this application, by constructing an integrated sea-air-shore monitoring network, the unified collection and transmission of multi-source heterogeneous marine data are realized, solving the problems of scattered data acquisition and inconsistent formats. The periodic deterministic sampling algorithm and data standardization processing are adopted to effectively reduce the noise interference during data acquisition and ensure data quality and timeliness; based on the digital twin technology, a digital model of the marine environment is established to accurately depict the dynamic correlation relationships among marine environmental elements and enhance the cognitive ability of the laws of marine environmental changes; the combination of the selective caching mechanism and the deep learning algorithm optimizes the usage efficiency of computing resources and improves the accuracy of feature fusion at the same time; through the transfer learning method and multi-model integration technology, the knowledge of the pre-trained model is fully utilized to enhance the generalization performance of different types of marine disaster prediction models; a complete model training and optimization process is established to realize the full-process intelligent processing from data acquisition to disaster prediction, significantly improving the prediction accuracy and timeliness; the distributed computing architecture and edge computing technology are adopted to optimize the allocation and scheduling of computing resources to meet the real-time requirements of massive data processing.

[0042] In a specific embodiment, the process of executing step 100 may specifically include the following steps:

[0043] The monitoring area is divided into grids at three levels: marine, air, and shore-based to obtain monitoring grid units, and the density of marine environmental monitoring devices is optimized and arranged according to the spatial distribution of the monitoring grid units to obtain a monitoring device deployment plan;

[0044] Function allocation of the data acquisition module, data preprocessing module, and communication module is performed on each monitoring device in the monitoring device deployment plan to obtain device function configuration information;

[0045] The device function configuration information is input into a network topology generator to construct multi-level network connections to obtain a network topology structure diagram, and communication relay devices are deployed in different regions and communication protocols are configured according to the network topology structure diagram to obtain a data transmission network;

[0046] Deploy a lightweight prediction model on the edge computing server of the data processing device in the data transmission network to obtain a local data processing unit, and based on the local data processing unit, perform distributed deployment and computing resource allocation on the data fusion model in the cloud computing center to obtain a central computing processing unit;

[0047] Integrate the data transmission network, the local data processing unit, and the central computing processing unit systemically and unify the protocols to obtain an integrated sea-air-shore monitoring network.

[0048] Specifically, considering the complexity of the monitoring area, a three-dimensional grid division method is adopted to divide the area into three levels: ocean, air, and shore-based. Assume that the ocean range of the monitoring area is a two-dimensional rectangular area , whose boundary is described by and . At the same time, the air monitoring layer is further divided by the vertical stratification height range . Then the three-dimensional grid division is expressed as:

[0049] Number of grid cells ;

[0050] where is the spatial resolution of the two-dimensional plane grid, and is the resolution in the vertical height direction. After completing the grid division, according to the spatial distribution and regional characteristics of the monitoring grid cells, optimize the density layout of the marine environment monitoring equipment. The density optimization layout is achieved through the following objective function:

[0051] ;

[0052] where is the total number of grid cells, is the importance weight of the th cell (e.g., the weight of the grid near the typhoon path is large), and is the distance between the th cell and the nearest monitoring device. During the optimization process, adjust the layout position of the equipment to ensure that more equipment is arranged in high-weight areas and minimize the coverage redundancy between devices. After completing the equipment layout, assign three major functions to each device: a data acquisition module, a data preprocessing module, and a communication module. The data acquisition module is responsible for collecting basic environmental parameters such as water temperature, salinity, and air pressure; the data preprocessing module formats, removes outliers, and compresses the collected data; the communication module uploads the data to the relay device through a wireless communication protocol. Assume that the data volume of device is , the data acquisition rate is , and the processing delay is . Then the transmission delay of the device is expressed as:

[0053] ;

[0054] After the device function allocation is completed, the device function configuration information is input into the network topology generator, and a multi-level network topology structure diagram is constructed through hierarchical design. Assume that the monitoring network consists of communication relay devices and monitoring devices. Through minimizing the total communication cost to achieve regional deployment:

[0055] ;

[0056] Among them, represents the transmission cost (such as power consumption and delay) between the relay device and the monitoring device . is a connection indication variable. If the device is connected to the relay device , then , otherwise it is 0. By solving this optimization problem, the regional deployment of communication relay devices is completed, and communication protocols are configured in combination with actual network requirements. For example, LoRa protocol is used for ocean buoys, while satellite communication is adopted for air layer devices. In the data transmission network, to improve data processing efficiency, a lightweight prediction model is deployed on the edge computing server to form a local data processing unit. Assume that the computing power of the edge server is , and the local model computing requirement is , then the following constraints need to be satisfied:

[0057] ;

[0058] The edge server can predict short-term path changes in real time and transmit the results to the cloud computing center simultaneously. The cloud computing center deploys a data fusion model to process global prediction tasks through distributed computing resource allocation. Assume that the number of cloud computing nodes is , the computing power of each node is , and the total task volume is . The distributed resource allocation strategy needs to satisfy:

[0059] ;

[0060] By coordinating the task allocation between the cloud computing center and the edge computing server, a central computing processing unit is formed. The data transmission network, local data processing unit, and central computing processing unit are system-integrated through a unified communication protocol and interface standard to form a sea-air-shore integrated monitoring network. This network can achieve efficient linkage from data collection to processing, and support real-time monitoring and disaster warning of the marine environment.

[0061] Among them, the device function configuration information is input into a network topology generator to construct a multi-level network connection, obtaining a network topology structure diagram. Then, according to the network topology structure diagram, communication relay devices are deployed in regions and communication protocols are configured to obtain a data transmission network, including: performing state space modeling on the data acquisition module, data preprocessing module, and communication module in the device function configuration information to obtain an intelligent agent environment state set; performing quantitative calculations on communication resource consumption, data transmission delay, and network load according to the environment state set to obtain an action reward function; inputting the environment state set and the action reward function into a deep Q network for policy iteration training to obtain a network topology optimization strategy; constructing hierarchical connections for the communication links between devices according to the network topology optimization strategy to obtain an initial network topology structure; performing topology optimization and update on the initial network topology structure using a multi-agent collaborative learning algorithm to obtain a network topology structure diagram; based on the network topology structure diagram, using a density clustering algorithm to plan the regional layout of communication relay devices to obtain the deployment locations of relay devices; according to the deployment locations of relay devices, adaptively configuring communication protocol parameters using a dynamic programming algorithm to obtain a protocol configuration plan; inputting the protocol configuration plan into a network simulator for performance evaluation and parameter fine-tuning to obtain a data transmission network.

[0062] In a specific embodiment, the process of executing step 200 may specifically include the following steps:

[0063] Periodically divide and prioritize the data acquisition time slots of ocean environment monitoring devices in the sea-air-shore integrated monitoring network to obtain a sampling scheduling sequence;

[0064] Calculate the error and threshold judgment for the original data collected by the monitoring devices according to the sampling scheduling sequence to obtain a data credibility index, and input the data with a data credibility index less than a preset threshold into an exception handling module for marking and elimination to obtain an effective data set;

[0065] Dynamically adjust and synchronize control the data sampling period based on the time distribution characteristics of the effective data set to obtain time series calibration data, and perform data cleaning, normalization, and numerical quantization processing on the time series calibration data to obtain data in standard format;

[0066] Perform geographical location coding and timestamp marking on the data in standard format according to the spatial distribution of the monitoring devices to obtain spatio-temporal associated data, and calculate the data change rate and transmission delay between adjacent sampling points based on the spatio-temporal associated data to obtain data timeliness parameters;

[0067] Perform data packet encapsulation and transmission protocol encoding on the spatio-temporal associated data and the data timeliness parameters to obtain ocean environment data with spatio-temporal identifiers.

[0068] Specifically, the data acquisition time slots of marine environment monitoring devices are periodically divided and prioritized to ensure that high-priority monitoring tasks can obtain resource allocation first. Assume the set of monitoring devices is , the importance weight of its acquisition task is , the acquisition duration is , the total acquisition period is , then the acquisition time slot of each device meets the following optimization objectives:

[0069] ;

[0070] Among them, is the weight set according to the urgency of the task or the importance of the data. For example, in the typhoon path prediction task, the monitoring devices near the typhoon center area will be given a higher weight to ensure the acquisition and processing of their data first. According to the optimized sampling scheduling sequence, error calculation and threshold judgment are performed on the original data collected by the monitoring devices to evaluate the credibility of the data. Assume the original data value is , the standard value of the data is , and its error calculation formula is:

[0071] ;

[0072] Among them, represents the relative error of the th data point. When exceeds the preset threshold , this data point is considered untrustworthy and is marked and excluded. Through this process, abnormal data is effectively excluded, and a valid data set with higher credibility is obtained. Based on the time distribution characteristics of the valid data set, dynamic adjustment and synchronization control are performed on the data sampling period. Assume the sampling period is , and the time distribution characteristics of the data are described by its autocorrelation function :

[0073] ;

[0074] Among them, is the time lag, is the data value at time . By analyzing the change trend, the main periodic characteristics of the data are determined, and is dynamically adjusted accordingly.。The adjusted data ensures the alignment of data timestamps across different devices through synchronization control, generating time-series calibration data. Based on the time-series calibration data, data cleaning, normalization, and numerical quantization processes are performed to generate data in a standard format. Data cleaning ensures the integrity and consistency of the data by removing missing values and outliers; normalization maps the data to a unified range [0, 1], and its formula is:

[0075] ;

[0076] where, and are the minimum and maximum values of the data respectively; numerical quantization reduces the storage and computational burden by discretizing continuous data into numerical values with a fixed precision. The data in the standard format is geographically encoded and timestamped according to the spatial distribution of the monitoring devices to generate spatio-temporal correlation data. Assuming the monitoring device is located at the longitude and latitude coordinates , and its acquisition time is , then each piece of data is marked as a triple . Based on these spatio-temporal correlation data, the data change rate and transmission delay between adjacent sampling points are calculated. The change rate is expressed as:

[0077] ;

[0078] while the transmission delay is expressed as:

[0079] ;

[0080] where, and are the timestamps of data reception and transmission respectively. These parameters reflect the timeliness and dynamic change characteristics of the data. The spatio-temporal correlation data and data timeliness parameters are packet encapsulated and transmission protocol encoded to generate ocean environment data with spatio-temporal identifiers. Packet encapsulation groups the data through standard network protocols (such as TCP / IP or MQTT) and attaches necessary check information and header information; transmission protocol encoding selects appropriate communication methods (such as LoRa or satellite communication) according to different devices and network levels to ensure the reliability and efficiency of data transmission.

[0081] In a specific embodiment, the process of executing step 300 may specifically include the following steps:

[0082] Classify and digitally encode the ocean environment data with spatio-temporal identifiers according to the water temperature field, salinity field, sea level height field, wind speed field, and air pressure field to obtain an ocean element data structure;

[0083] Map the physical quantities in the ocean element data structure to time series and spatial distributions to obtain the spatio-temporal distribution characteristics of ocean environmental elements;

[0084] Based on the spatio-temporal distribution characteristics, conduct coupling analysis and correlation degree calculation on the interaction relationships between ocean environmental elements to obtain an ocean element correlation matrix;

[0085] Input the ocean element correlation matrix into a dynamic evolution model to analyze the changing rules of the ocean environmental state, obtain an environmental evolution sequence, and conduct stability analysis and mutation feature detection on the environmental evolution sequence to obtain the characteristics of ocean environmental abnormal events;

[0086] Based on the characteristics of ocean environmental abnormal events, establish a mapping function between the physical ocean environment and the digital space to obtain a digital twin mapping relationship, and input the digital twin mapping relationship into a feature extractor to extract key features of the ocean environment to obtain an ocean environment feature vector;

[0087] Conduct model integration and parameter optimization on the ocean environment feature vector and the digital twin mapping relationship to obtain an ocean environment digital model.

[0088] Specifically, classify the ocean environmental data with spatio-temporal identifiers according to physical elements such as water temperature field, salinity field, sea level height field, wind speed field, and air pressure field. By extracting the physical attributes of the data and grouping them according to their scientific meanings, for example, the water temperature field describes the sea water temperature distribution at a specific time and spatial coordinates , and the salinity field describes the salinity distribution at the same space and time, etc. Digitally encode each type of data, convert it into a unified ocean element data structure, and represent it by the following formula:

[0089] ;

[0090] where respectively represent the discrete numerical values of water temperature, salinity, sea level height, wind speed, and air pressure, and are the discretized dimensions of time and space. After completing the digital encoding, map the physical quantities in the ocean element data structure to time series and spatial distributions to reveal the spatio-temporal distribution characteristics of ocean environmental elements. The time series mapping analyzes the continuous values in the time dimension to capture long-term trends, periodic fluctuations, and mutation points, and its mathematical representation is:

[0091] ;

[0092] where is the autocorrelation function corresponding to the time lag , represents the th value of the time series. The spatial distribution map is constructed by interpolation techniques and distribution fitting methods to generate a continuous spatial distribution map, such as a spatial prediction model based on Gaussian process regression:

[0093] ;

[0094] where, is the physical quantity value of the prediction point, is the Gaussian kernel function, is the weight. Based on the spatio-temporal distribution characteristics, the interaction relationships between ocean environmental elements are analyzed through coupling analysis and correlation calculation. Taking water temperature and salinity as an example, the coupling analysis quantifies the causal relationship between them through Granger causality test or mutual information method. The formula for mutual information is:

[0095] ;

[0096] where, represents the joint probability distribution, and are the marginal probability distributions respectively. The ocean element correlation matrix generated by the correlation calculation can describe the relationship strength between elements, where represents the correlation degree between the th and the th elements. The correlation matrix is input into the dynamic evolution model to analyze the variation law of the ocean environmental state. The dynamic evolution model describes the spatio-temporal evolution of the system through a system of differential equations. For example, the diffusion of the water temperature field is described using partial differential equations:

[0097] ;

[0098] where, is the diffusion coefficient, is the external heat source. Through numerical simulation, an environmental evolution sequence can be generated, and the stability of the evolution sequence is analyzed through eigenvalue analysis and singular spectrum analysis to extract possible mutation characteristics. Based on the mutation characteristics of the environmental evolution, a mapping function between the physical ocean environment and the digital space is established to generate the digital twin mapping relationship. This mapping is fitted through a multivariate regression model or a deep learning network. For example, a nonlinear mapping is established using a multi-layer perceptron (MLP) model:

[0099] ;

[0100] where, is the weight matrix, is the input feature matrix, is the bias, It is a non-linear activation function. The generated digital twin mapping relationship is input into the feature extractor to further extract key features and generate a high-dimensional ocean environment feature vector. Model integration and parameter optimization are performed on the ocean environment feature vector and the digital twin mapping relationship. The integration method is, for example, random forest or gradient boosting decision tree, which improves the prediction performance through multi-model fusion. Parameter optimization uses Bayesian optimization or particle swarm optimization algorithm. Finally, an ocean environment digital model with optimal performance is obtained, which comprehensively reflects the spatio-temporal dynamic characteristics of the ocean environment.

[0101] In a specific embodiment, the process of executing step 400 may specifically include the following steps:

[0102] Classify and construct the ocean environment element features in the ocean environment digital model according to data types and spatio-temporal distributions to obtain a feature data hierarchical structure, and cache and prioritize the historical data features based on the feature data hierarchical structure to obtain a hierarchical cache database;

[0103] Input the feature data in the hierarchical cache database into a deep neural network for latent feature extraction and non-linear mapping to obtain a deep feature representation;

[0104] Perform multi-modal feature fusion and complementary analysis on the spatial features, temporal features, and physical features in the deep feature representation to obtain a fused feature vector;

[0105] Input the fused feature vector into an attention calculation unit for feature weight assignment and significance calculation to obtain a weighted feature set, and establish a feature organizational structure and indexing mechanism for the weighted feature set according to the time dimension and space dimension to obtain a spatio-temporal feature organization system;

[0106] Perform similarity measurement and correlation mining on the feature data in the spatio-temporal feature organization system to obtain a feature correlation map;

[0107] Input the feature correlation map into a deep spatio-temporal fusion network for cross-modal feature learning and spatio-temporal dependence modeling to obtain a fused feature tensor, and perform dimensionality reduction processing and feature selection on the fused feature tensor to obtain spatio-temporal feature data.

[0108] Specifically, classify the ocean environment element features in the ocean environment digital model according to data types and spatio-temporal distributions to construct a hierarchical structure of feature data. This process groups different types of data based on the physical attributes and statistical characteristics of the data. For example, water temperature and salinity are classified as continuous spatio-temporal features, sea level height is classified as event-driven features, and wind speed and air pressure are classified as dynamic meteorological features. Based on these classifications, priorities are assigned to the feature data according to their importance in spatio-temporal distribution, and a hierarchical cache database is constructed by optimizing the following objective function:

[0109] ;

[0110] Among them, is the total number of feature data, is the importance weight of the th feature, is the data update frequency, is the data volume of the feature. Through this optimization function, the data in the cache is hierarchically managed, enabling higher-priority data to receive a higher cache resource allocation. The feature data in the hierarchical cache database is input into a deep neural network for latent feature extraction and non-linear mapping. Assume the input data is , where is the input feature dimension. The deep neural network realizes feature abstraction through multiple fully connected layers and activation functions. The output of each layer is expressed as:

[0111] ;

[0112] Among them, is the output of the th layer, is the weight matrix, is the bias vector, is the non-linear activation function (such as ReLU). The finally output deep feature representation contains the latent features of the data and has higher expressive power. Multimodal feature fusion and complementary analysis are performed on the spatial features, temporal features, and physical features in the deep feature representation. Assume the spatial feature is , the temporal feature is , and the physical feature is . Their fusion is achieved by weighted summation:

[0113] ;

[0114] Among them, is the weight coefficient, reflecting the importance of different modal features. Through complementary analysis, the synergy of these features is enhanced to obtain the fused feature vector . The fused feature vector is input into an attention calculation unit for feature weight assignment and saliency calculation. The attention mechanism dynamically assigns weights by calculating the importance score of each feature. Its calculation formula is:

[0115] ;

[0116] Among them, is the scoring function, and are learnable parameters respectively, is the feature vectors. Through the attention mechanism, a weighted feature set is generated, and a feature organizational structure and an indexing mechanism are established based on the time dimension and the space dimension to form a spatio-temporal feature organization system. Based on the spatio-temporal feature organization system, similarity measurement and correlation mining are performed on the feature data therein. The similarity measurement is achieved through the cosine similarity formula:

[0117] ;

[0118] The correlation mining is achieved through the construction of a graph structure, where the adjacency matrix of the graph element represents the feature and the correlation strength between. Finally, a feature correlation map is generated to describe the complex relationships between features. The feature correlation map is input into a deep spatio-temporal fusion network for cross-modal feature learning and spatio-temporal dependence modeling. Assuming the input map is where is the node set, is the adjacency matrix, and the deep spatio-temporal fusion network captures spatial features through graph convolutional layers and captures temporal dependencies through recurrent neural networks. The graph convolution operation formula is:

[0119] ;

[0120] wherein, is the node feature of the layer, is the weight matrix, is the activation function. The temporal modeling is implemented through LSTM, and its update formula is:

[0121] ;

[0122] Through the above operations, a fused feature tensor is generated. Dimensionality reduction processing and feature selection are performed on the fused feature tensor, such as through principal component analysis or recursive feature elimination methods, to extract the most important features, and finally high-quality spatio-temporal feature data is obtained.

[0123] In a specific embodiment, the process of executing step 500 may specifically include the following steps:

[0124] The spatio-temporal feature data is divided into a typhoon track prediction data set, a storm surge prediction data set, and a red tide prediction data set;

[0125] Build a typhoon path prediction model based on the typhoon path prediction dataset. The typhoon path prediction model includes an encoder network and a decoder network. The encoder network uses a 5-layer convolutional structure to extract typhoon movement features. Each layer of the convolutional structure uses a ReLU activation function and a batch normalization layer. The decoder network uses a 5-layer transposed convolutional structure to generate the path prediction result, and a skip connection is set between the encoder network and the decoder network to obtain the initial typhoon path prediction result;

[0126] Build a storm surge prediction model for the storm surge prediction dataset. The storm surge prediction model includes a temporal feature extraction network and a spatial feature fusion network. The temporal feature extraction network uses a bidirectional long short-term memory network structure to model the tidal level change. The spatial feature fusion network uses a graph convolutional network structure to perform correlation analysis on multi-point tidal level data to obtain the initial storm surge prediction result;

[0127] Build a red tide prediction model according to the red tide prediction dataset. The red tide prediction model includes a multi-scale feature extraction module and a spatio-temporal attention module. The multi-scale feature extraction module uses 3 parallel residual networks to process different scale features. The spatio-temporal attention module focuses on analyzing key regions and time periods to obtain the initial red tide prediction result;

[0128] Input the pre-trained marine disaster prediction model library into the transfer learning network for knowledge transfer. The transfer learning network includes a feature mapping layer and a task adaptation layer to optimize the typhoon path initial prediction result, the storm surge initial prediction result and the red tide initial prediction result to obtain the optimized prediction result set;

[0129] Perform uncertainty analysis and confidence calculation on the optimized prediction result set to obtain the model prediction confidence interval;

[0130] Input the model prediction confidence interval into the ensemble learning module for multi-model fusion and weight optimization to obtain the ensemble prediction result, and perform model integration and parameter calibration on the ensemble prediction result and the optimized prediction result set to obtain the marine disaster prediction model set.

[0131] Specifically, the spatio-temporal feature data is divided into three sub-datasets, which are respectively used for typhoon path prediction, storm surge prediction and red tide prediction. Among them, the typhoon path prediction dataset includes data on the change of the typhoon center position (latitude and longitude) over time. The storm surge prediction dataset covers multi-point tidal level observations and meteorological conditions, while the red tide prediction dataset contains spatial distribution data of water temperature, salinity and nutrient concentration. This division method ensures the sufficient expression ability of each dataset for the corresponding prediction task. Based on the typhoon path prediction dataset, a typhoon path prediction model is built. The model adopts an encoder-decoder structure, where the encoder network contains a 5-layer convolutional structure and the decoder network contains a 5-layer transposed convolutional structure. Assuming the input data is , the convolution process is expressed as:

[0132] ;

[0133] Among them, is the output feature map of the th layer, and are the convolution kernel and bias parameters respectively. * represents the convolution operation, BN is batch normalization, and ReLU is the activation function. After the typhoon motion features are extracted by the encoder, the decoder generates the path prediction result through transposed convolution:

[0134] ;

[0135] Among them, is the output feature map of the decoder, and are the weights and biases of the transposed convolution. To avoid information loss, skip connections are set between the encoder and the decoder, enabling high-resolution features to be directly transmitted to the decoder layer to generate the initial prediction result of the typhoon path. For the storm surge prediction dataset, a storm surge prediction model is constructed. The model includes a time series feature extraction network and a spatial feature fusion network. The time series feature extraction network uses a bidirectional long short-term memory network (Bi-LSTM) to model the time-dependent relationship of tidal level changes, and its state update formula is:

[0136] ;

[0137] Among them, are the forget gate, input gate, and output gate respectively, is the cell state, is the hidden state, is the Sigmoid activation function, represents element-wise multiplication. The spatial feature fusion network uses a graph convolutional network (GCN) to capture the spatial correlation of tidal level data through the following formula:

[0138] ;

[0139] Among them, is the node feature matrix of the th layer, is the adjacency matrix, is the weight matrix, is the activation function. Combining the time series and spatial features, the initial prediction result of the storm surge is generated. For the red tide prediction dataset, a red tide prediction model is constructed. The model includes a multi-scale feature extraction module and a spatio-temporal attention module. The multi-scale feature extraction module uses 3 parallel residual networks, and the output of each residual network is:

[0140] ;

[0141] Among them, is the input feature, represents the non-linear transformation through multiple convolutional layers, is the weight set. The spatio-temporal attention module dynamically assigns weights through the following formula:

[0142] ;

[0143] Among them, is the attention weight, is the attention score, is the input feature, is the learnable parameter. Combining the multi-scale and attention mechanisms, the initial prediction result of the red tide is generated. The above initial prediction result is input into the transfer learning network for knowledge transfer. The transfer learning network includes a feature mapping layer and a task adaptation layer. The feature mapping layer maps the general features of the pre-trained model to the current task through the following formula:

[0144] ;

[0145] Among them, and are the mapping layer parameters, is the input feature. The task adaptation layer uses the adversarial learning mechanism to optimize the task-specific features, and realizes knowledge transfer through the interactive training of the generator and the discriminator. Uncertainty analysis and confidence calculation are performed on the optimized prediction result set. The uncertainty of the prediction result is estimated by Monte Carlo sampling, and the confidence interval calculation formula is:

[0146] ;

[0147] Among them, is the prediction mean, is the standard deviation, is the confidence coefficient, is the number of samples. Through uncertainty analysis, the confidence interval of the prediction result is generated. The confidence interval is input into the ensemble learning module, and multi-model fusion is achieved through weight optimization. The prediction result of the ensemble learning is:

[0148] ;

[0149] Among them, is the prediction result of the th model, is the weight, is the number of models. Through model integration and parameter calibration, the marine disaster prediction model set is finally generated.

[0150] In a specific embodiment, the process of inputting the pre-trained marine disaster prediction model library into the transfer learning network for knowledge transfer, where the transfer learning network includes a feature mapping layer and a task adaptation layer, and optimizing the initial typhoon path prediction result, initial storm surge prediction result, and initial red tide prediction result to obtain an optimized prediction result set may specifically include the following steps:

[0151] Input the weight parameters in the pre-trained marine disaster prediction model library into the feature mapping layer for general feature analysis to obtain a shared feature space;

[0152] Based on the shared feature space, perform feature vector mapping on the initial typhoon path prediction result. The feature mapping layer adopts a 3-layer fully connected network structure, and each layer of the fully connected network structure uses the LeakyReLU activation function to obtain the typhoon path mapping feature;

[0153] Perform feature vector mapping on the initial storm surge prediction result. The feature mapping layer adopts a 3-layer convolutional network enhanced by an attention mechanism, and each layer of the convolutional network enhanced by the attention mechanism uses batch normalization and the ReLU activation function to obtain the storm surge mapping feature;

[0154] According to the initial red tide prediction result, perform feature vector mapping. The feature mapping layer adopts a 3-layer graph neural network structure, and each layer of the graph neural network structure uses the GraphSAGE aggregation function to obtain the red tide mapping feature;

[0155] Input the typhoon path mapping feature, storm surge mapping feature, and red tide mapping feature into the task adaptation layer for cross-task knowledge transfer. The task adaptation layer adopts an adversarial learning mechanism to obtain a transfer feature representation;

[0156] Perform task relevance analysis and domain adaptation training on the transfer feature representation. The task adaptation layer uses an adversarial discriminator structure to obtain task-specific features;

[0157] Based on the task-specific features, perform fine-grained tuning and gradient update on the parameters in the pre-trained marine disaster prediction model library to obtain model optimization parameters, and based on the model optimization parameters, perform parameter update and model reconstruction to obtain an optimized prediction result set.

[0158] Specifically, input the weight parameters in the pre-trained marine disaster prediction model library into the feature mapping layer. The main role of the feature mapping layer is to analyze the shared general features in the model and construct a high-dimensional shared feature space. This process is achieved through the following formula:

[0159] ;

[0160] Wherein, is the shared feature representation, is the input feature of the pre-trained model, and are the weight matrix and the bias vector respectively, represents a non-linear activation function (e.g., ReLU). After extracting general features through the feature mapping layer, the constructed shared feature space can capture the basic knowledge across tasks in the pre-trained model. Based on the shared feature space, a feature vector mapping is performed on the initial typhoon path prediction result. The feature mapping layer designs a three-layer fully connected network for the typhoon path task, and the output of each layer is expressed as:

[0161] ;

[0162] where, is the output feature vector of the -th layer, and are the weight matrix and the bias vector respectively, and LeakyReLU is a ReLU activation function with a leakage parameter, which is used to avoid the problem of gradient disappearance. Through three-layer non-linear mapping, the typhoon path mapping features are finally generated, retaining the high-order representation of the path dynamic features. For the initial storm surge prediction result, the feature mapping layer adopts a three-layer convolutional network enhanced by the attention mechanism, and the output of each layer is expressed as:

[0163] ;

[0164] where * represents the convolution operation, BN is batch normalization, is the attention weight, and the calculation formula is:

[0165] ;

[0166] where, and are the attention weight and the bias parameter respectively, is the score vector, is the -th feature. By combining the convolution operation and the attention mechanism, the storm surge mapping features are generated, which can accurately capture the spatial correlation in the tide level distribution. For the initial red tide prediction result, the feature mapping layer adopts a three-layer graph neural network structure (GNN), and each layer uses the GraphSAGE aggregation function to update the node features. The formula is:

[0167] ;

[0168] where, represents the feature of node at the -th layer, is the neighbor set of node , and AGGREGATE is the feature aggregation function (such as taking the mean or maximum value), and are the weights and biases. Through three-layer iterative updates of the graph structure, the red tide mapping features are generated, which can effectively capture the complex spatial relationships in the red tide occurrence areas. The typhoon path mapping features, storm surge mapping features, and red tide mapping features are input into the task adaptation layer for cross-task knowledge transfer. The task adaptation layer achieves feature alignment through an adversarial learning mechanism and generates transfer feature representations. Adversarial learning consists of two parts: a generator and a discriminator. The goal of the generator is to generate representations consistent with the feature distribution of the target task, and its optimization objective is:

[0169] ;

[0170] where is the discriminator, is the generator, and are the data distribution and noise distribution respectively. Through adversarial optimization, the transfer feature representations can retain both general knowledge and task-specific characteristics. Task relevance analysis and domain adaptation training are performed on the transfer feature representations. Task relevance is measured by calculating the Kullback-Leibler divergence (KL divergence) of the feature distributions:

[0171] ;

[0172] where and are the feature distributions of the source task and the target task respectively. Domain adaptation training is further optimized through an adversarial discriminator structure to make the source task features more consistent with the target task feature distribution, thereby generating task-specific features. Based on the task-specific features, fine-grained tuning and gradient updates are performed on the parameters in the pre-trained model library. Parameter tuning updates the weights through backpropagation, and the formula is:

[0173] ;

[0174] where is the model parameter, is the learning rate, is the loss function. The tuned model parameters are used for model reconstruction, and finally an optimized prediction result set is generated.

[0175] The multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method in the embodiments of the present application is described above. Next, the multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction platform 10 in the embodiments of the present application will be described. Please refer to Figure 2 , an embodiment of the multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction platform 10 in the embodiments of the present application includes:

[0176] Building module 11 is used to hierarchically deploy and construct a network topology for marine environment monitoring devices, communication relay devices, and data processing devices to obtain an integrated sea-air-shore monitoring network;

[0177] Acquisition module 12 is used to collect and standardize water temperature, salinity, sea level height, wind speed, and air pressure data according to the integrated sea-air-shore monitoring network by using a periodic deterministic sampling algorithm to obtain marine environment data with spatio-temporal identifiers;

[0178] Feature extraction module 13 is used to perform marine element modeling and dynamic feature extraction based on the marine environment data with spatio-temporal identifiers to obtain a marine environment digital model;

[0179] Correlation analysis module 14 is used to perform feature fusion and correlation analysis on multi-source heterogeneous marine data based on the marine environment digital model by using a selective caching mechanism and a deep learning algorithm to obtain spatio-temporal feature data;

[0180] Model integration module 15 is used to train and integrate multiple models for typhoon path prediction models, storm surge prediction models, and red tide prediction models according to the spatio-temporal feature data by using a transfer learning method to obtain a set of marine disaster prediction models.

[0181] Through the collaborative cooperation of the above-mentioned various components, by constructing an integrated sea-air-shore monitoring network, the unified collection and transmission of multi-source heterogeneous marine data are realized, the problems of scattered data acquisition and inconsistent formats are solved, and the periodic deterministic sampling algorithm and data standardization processing are used to effectively reduce the noise interference in the data acquisition process and ensure the data quality and timeliness; a marine environment digital model is established based on the digital twin technology to accurately depict the dynamic correlation relationship between marine environment elements and enhance the cognitive ability of the laws of marine environment changes; the combination of the selective caching mechanism and the deep learning algorithm optimizes the use efficiency of computing resources and improves the accuracy of feature fusion at the same time; through the transfer learning method and multi-model integration technology, the knowledge of pre-trained models is fully utilized to enhance the generalization performance of different types of marine disaster prediction models; a complete model training and optimization process is established to realize the full-process intelligent processing from data acquisition to disaster prediction, significantly improving the prediction accuracy and timeliness; a distributed computing architecture and edge computing technology are adopted to optimize the allocation and scheduling of computing resources and meet the real-time requirements of massive data processing.

[0182] Please refer to Figure 3 , Figure 3 which is a schematic block diagram of the structure of the electronic device 300 provided by the embodiment of the present application. The electronic device 300 includes a processor 301 and a memory 302. The processor 301 and the memory 302 are connected through a platform bus 303. Among them, the memory 302 may include a non-volatile storage medium and an internal memory.

[0183] The non-volatile storage medium can store a computer program. The computer program includes program instructions which, when executed by the processor 301, can cause the processor 301 to execute any of the above-mentioned multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction methods.

[0184] The processor 301 is used to provide computing and control capabilities to support the operation of the entire electronic device 300.

[0185] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, it can cause the processor 301 to execute any of the above-mentioned multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction methods.

[0186] Those skilled in the art can understand that Figure 3 the structure shown in [[ ]] is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the electronic device 300 involved in the solution of this application. The specific electronic device 300 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0187] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0188] It should be noted that those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described electronic device 300 can refer to the corresponding process of the foregoing multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method, and will not be elaborated herein.

[0189] This application embodiment also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program which, when executed by one or more processors, causes the one or more processors to implement the multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method provided by this application embodiment.

[0190] Among them, the computer-readable storage medium may be an internal storage unit of the electronic device 300 in the foregoing embodiments, such as the hard disk or memory of the electronic device 300. The computer-readable storage medium may also be an external storage device of the electronic device 300, such as a plug-in hard disk equipped with the electronic device 300, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc.

[0191] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, platforms, and units described above may refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0192] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing an electronic device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0193] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A method for intelligent fusion of multi-source heterogeneous ocean data and prediction of marine disasters, characterized in that: The method comprises: Hierarchical deployment and network topology construction of marine environment monitoring equipment, communication relay equipment and data processing equipment to obtain an integrated sea, air and shore monitoring network; According to the integrated sea-air-shore monitoring network, a periodic deterministic sampling algorithm is used to collect and standardize water temperature, salinity, sea level, wind speed and air pressure data to obtain marine environmental data with time and space identification; Based on the marine environment data with time and space labels, marine element modeling and dynamic feature extraction are performed to obtain a marine environment digital model; specifically, the marine environment data with time and space labels are classified and digitally encoded according to water temperature field, salinity field, sea level height field, wind speed field and air pressure field to obtain a marine element data structure; time series and spatial distribution mapping are performed on the physical quantities in the marine element data structure to obtain the spatiotemporal distribution characteristics of the marine environment elements; coupling analysis and correlation calculation are performed on the interaction relationship between the marine environment elements based on the spatiotemporal distribution characteristics to obtain a marine element association matrix; The marine element association matrix is ​​input into a dynamic evolution model to analyze the law of marine environment state changes, and an environmental evolution sequence is obtained. The environmental evolution sequence is subjected to stability analysis and mutation feature detection to obtain the characteristics of marine environmental abnormal events. A mapping function between the physical marine environment and the digital space is established based on the characteristics of the marine environmental abnormal events to obtain a digital twin mapping relationship. The digital twin mapping relationship is input into a feature extractor to extract key features of the marine environment to obtain a marine environment feature vector. Model integration and parameter optimization are performed on the marine environment feature vector and the digital twin mapping relationship to obtain a marine environment digital model. Based on the digital model of the ocean environment, a selective caching mechanism and a deep learning algorithm are used to perform feature fusion and correlation analysis on multi-source heterogeneous ocean data to obtain spatiotemporal feature data; specifically, the following steps are performed: the features of the ocean environment elements in the digital model of the ocean environment are classified and constructed according to the data type and spatiotemporal distribution to obtain a feature data hierarchy, and historical data features are cached and prioritized based on the feature data hierarchy to obtain a hierarchical cache database; the feature data in the hierarchical cache database is input into a deep neural network for potential feature extraction and nonlinear mapping to obtain a deep feature representation; the spatial features, temporal features and physical features in the deep feature representation are analyzed. Perform multimodal feature fusion and complementarity analysis to obtain a fused feature vector; input the fused feature vector into the attention calculation unit to perform feature weight allocation and significance calculation to obtain a weighted feature set, and establish a feature organization structure and index mechanism for the weighted feature set according to the time dimension and space dimension to obtain a spatiotemporal feature organization system; perform similarity measurement and correlation mining on the feature data in the spatiotemporal feature organization system to obtain a feature association map; input the feature association map into a deep spatiotemporal fusion network to perform cross-modal feature learning and spatiotemporal dependency modeling to obtain a fused feature tensor, and perform dimensionality reduction processing and feature selection on the fused feature tensor to obtain spatiotemporal feature data; According to the spatiotemporal characteristic data, the typhoon path prediction model, the storm surge prediction model and the red tide prediction model are trained and integrated by the transfer learning method to obtain a marine disaster prediction model set.

2. The method for intelligent fusion of multi-source heterogeneous ocean data and prediction of marine disasters according to claim 1 is characterized in that: The marine environment monitoring equipment, communication relay equipment and data processing equipment are hierarchically deployed and the network topology is constructed to obtain an integrated sea, air and shore monitoring network, including: Divide the monitoring area into three grids: ocean, air and shore-based, to obtain monitoring grid units, and optimize the density of marine environment monitoring equipment according to the spatial distribution of the monitoring grid units to obtain a monitoring equipment deployment plan; Allocate functions of a data acquisition module, a data preprocessing module and a communication module to each monitoring device in the monitoring device deployment scheme to obtain device function configuration information; Input the device function configuration information into a network topology generator to construct a multi-level network connection to obtain a network topology structure diagram, and deploy the communication relay device in different areas and configure the communication protocol according to the network topology structure diagram to obtain a data transmission network; Performing lightweight prediction model deployment on the edge computing server of the data processing device in the data transmission network to obtain a local data processing unit, and performing distributed deployment and computing resource allocation on the data fusion model of the cloud computing center based on the local data processing unit to obtain a central computing processing unit; The data transmission network, the local data processing unit and the central computing processing unit are system-integrated and protocol-unified to obtain a sea-air-shore integrated monitoring network.

3. The method for intelligent fusion of multi-source heterogeneous ocean data and prediction of marine disasters according to claim 2 is characterized in that: According to the integrated sea, air and shore monitoring network, a periodic deterministic sampling algorithm is used to collect and standardize water temperature, salinity, sea level, wind speed and air pressure data to obtain marine environmental data with time and space identification, including: Periodically dividing and prioritizing the data collection time slots of the marine environment monitoring equipment in the integrated sea-air-shore monitoring network to obtain a sampling scheduling sequence; According to the sampling scheduling sequence, error calculation and threshold judgment are performed on the raw data collected by the monitoring device to obtain a data credibility index, and the data whose data credibility index is less than a preset threshold is input into the abnormal processing module for marking and elimination to obtain a valid data set; Dynamically adjusting and synchronously controlling the data sampling period based on the time distribution characteristics of the effective data set to obtain time series calibration data, and performing data cleaning, normalization and numerical quantization processing on the time series calibration data to obtain standard format data; The standard format data is geographically encoded and time-stamped according to the spatial distribution of the monitoring equipment to obtain spatiotemporal correlation data, and the data change rate and transmission delay between adjacent sampling points are calculated based on the spatiotemporal correlation data to obtain data timeliness parameters; The spatiotemporal correlation data and the data timeliness parameters are packaged into data packets and encoded with a transmission protocol to obtain ocean environment data with spatiotemporal identification.

4. The method for intelligent fusion of multi-source heterogeneous ocean data and prediction of marine disasters according to claim 1 is characterized in that: According to the spatiotemporal feature data, a transfer learning method is used to train and integrate the typhoon path prediction model, the storm surge prediction model and the red tide prediction model to obtain a marine disaster prediction model set, including: The spatiotemporal characteristic data are divided into a typhoon path prediction data set, a storm surge prediction data set and a red tide prediction data set; A typhoon path prediction model is constructed based on the typhoon path prediction dataset, wherein the typhoon path prediction model includes an encoder network and a decoder network, wherein the encoder network uses a 5-layer convolution structure to extract typhoon motion features, and each convolution structure uses a ReLU activation function and a batch normalization layer, and the decoder network uses a 5-layer deconvolution structure to generate a path prediction result, and a jump connection is set between the encoder network and the decoder network to obtain an initial typhoon path prediction result; A storm surge prediction model is established for the storm surge prediction data set. The storm surge prediction model includes a temporal feature extraction network and a spatial feature fusion network. The temporal feature extraction network uses a bidirectional long short-term memory network structure to model tide level changes. The spatial feature fusion network uses a graph convolutional network structure to perform correlation analysis on multi-point tide level data to obtain an initial storm surge prediction result. A red tide prediction model is constructed according to the red tide prediction data set, wherein the red tide prediction model includes a multi-scale feature extraction module and a spatiotemporal attention module, wherein the multi-scale feature extraction module uses three parallel residual networks to process features of different scales, and the spatiotemporal attention module performs a focused analysis on key areas and time periods to obtain an initial red tide prediction result; Inputting the pre-trained marine disaster prediction model library into the transfer learning network for knowledge transfer, the transfer learning network includes a feature mapping layer and a task adaptation layer, and optimizing the model of the initial prediction result of the typhoon path, the initial prediction result of the storm surge, and the initial prediction result of the red tide to obtain an optimized prediction result set; Performing uncertainty analysis and confidence calculation on the optimized prediction result set to obtain a model prediction confidence interval; The model prediction confidence interval is input into the integrated learning module for multi-model fusion and weight optimization to obtain an integrated prediction result, and the integrated prediction result and the optimized prediction result set are subjected to model integration and parameter calibration to obtain a marine disaster prediction model set.

5. The method for intelligent fusion of multi-source heterogeneous ocean data and prediction of ocean disasters according to claim 4 is characterized in that: The pre-trained marine disaster prediction model library is input into the transfer learning network for knowledge transfer. The transfer learning network includes a feature mapping layer and a task adaptation layer. The model is optimized for the initial prediction result of the typhoon path, the initial prediction result of the storm surge, and the initial prediction result of the red tide to obtain an optimized prediction result set, including: The weight parameters in the pre-trained marine disaster prediction model library are input into the feature mapping layer for common feature analysis to obtain a shared feature space; Based on the shared feature space, feature vector mapping is performed on the initial prediction result of the typhoon path, the feature mapping layer adopts a 3-layer fully connected network structure, and each layer of the fully connected network structure uses a LeakyReLU activation function to obtain typhoon path mapping features; Performing feature vector mapping on the initial storm surge prediction result, the feature mapping layer adopts a 3-layer attention mechanism enhanced convolutional network, each layer of the attention mechanism enhanced convolutional network uses batch normalization and ReLU activation function to obtain storm surge mapping features; Performing feature vector mapping according to the initial red tide prediction result, the feature mapping layer adopts a 3-layer graph neural network structure, and each layer of the graph neural network structure uses a GraphSAGE aggregation function to obtain red tide mapping features; Inputting the typhoon path mapping feature, the storm surge mapping feature and the red tide mapping feature into a task adaptation layer for cross-task knowledge transfer, wherein the task adaptation layer adopts an adversarial learning mechanism to obtain a transfer feature representation; Performing task relevance analysis and domain adaptation training on the migration feature representation, wherein the task adaptation layer uses an adversarial discriminator structure to obtain task-specific features; Based on the task-specific features, fine-grained tuning and gradient updating are performed on the parameters in the pre-trained marine disaster prediction model library to obtain model optimization parameters, and parameter updating and model reconstruction are performed based on the model optimization parameters to obtain an optimized prediction result set.

6. A multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction platform, characterized in that: Used to execute the multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method as described in any one of claims 1 to 5, the multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction platform comprises: The construction module is used to hierarchically deploy marine environment monitoring equipment, communication relay equipment and data processing equipment and construct network topology to obtain an integrated sea, air and shore monitoring network; A collection module is used to collect and standardize water temperature, salinity, sea level, wind speed and air pressure data according to the integrated sea, air and shore monitoring network using a periodic deterministic sampling algorithm to obtain marine environmental data with time and space identification; A feature extraction module, used to perform ocean element modeling and dynamic feature extraction based on the ocean environment data with time and space identification to obtain a digital model of the ocean environment; An association analysis module is used to perform feature fusion and association analysis on multi-source heterogeneous ocean data based on the ocean environment digital model using a selective caching mechanism and a deep learning algorithm to obtain spatiotemporal feature data; The model integration module is used to train and integrate the typhoon path prediction model, the storm surge prediction model and the red tide prediction model according to the spatiotemporal feature data using a transfer learning method to obtain a set of marine disaster prediction models.

7. An electronic device, characterized in that: The electronic device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory so that the electronic device executes the multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method as described in any one of claims 1-5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the multi-source heterogeneous ocean data intelligent fusion and ocean disaster prediction method as described in any one of claims 1 to 5 is implemented.

Images