Federal multi-agent distributed semantic fusion method and system for marine early warning

Abstract

The present application relates to marine environment monitoring, distributed multi-agent and federal privacy computing technical field, disclose a kind of federal multi-agent distributed semantic fusion method and system for marine early warning, method includes: S1, edge data agent collects the original marine data of local edge data acquisition terminal, executes abnormal value rejection, format standardization lightweight pretreatment, extracts metadata based on the marine feature ontology stored in shared knowledge base, and generates multidimensional semantic vector by semantic mapping function, upload metadata, semantic vector to shared knowledge base;S2, task coordination agent receives the early warning task of marine disaster, calls marine feature ontology to generate task semantic vector by semantic analysis on task demand etc.;The method and system construct distributed semantic fusion mechanism, using federal learning and multi-agent system deep integration architecture, can be dynamically adjusted according to the emergency degree of early warning Calculation accuracy and resources.

Description

Technical Field

[0001] This invention relates to the fields of marine environmental monitoring, distributed multi-agent and federated privacy computing technologies, specifically to a federated multi-agent distributed semantic fusion method and system for marine early warning. Background Technology

[0002] Marine data mining is a core supporting technology for marine environmental monitoring, involving multi-source heterogeneous data (such as buoy sensor data, satellite remote sensing data, ship observation data, and underwater robot data). These data are characterized by strong spatiotemporal dynamics, multi-source heterogeneity, massive volume (such as TB-level buoy data and GB-level satellite data per day), and high noise. Traditional marine data mining methods are difficult to effectively handle these characteristics, resulting in low mining efficiency and insufficient accuracy.

[0003] Defects and shortcomings of existing technology: 1. The Challenge of Distributed Fusion of Heterogeneous Marine Data. Marine data comes from diverse sources and exhibits significant heterogeneity in spatiotemporal resolution, data format, coordinate system, quality standards, and update frequency. Traditional centralized ETL (Extract, Transform, Load) processes require the physical aggregation of all data to a central server, resulting in enormous network bandwidth pressure, high transmission latency, and difficulty in adapting to the open environment of dynamic addition / exit of observation platforms. The fusion process is rigid and lacks fault tolerance.

[0004] 2. Computational Bottlenecks in Real-Time Sensing and Early Warning of Complex Marine Phenomena. For sudden marine disasters such as red tides and tsunamis, low-latency correlation analysis of multi-source data is required to generate early warnings. Existing models based on centralized cloud computing or single-machine algorithms are prone to creating computational hotspots when data volumes surge, and response times cannot meet the demands for minute-level or even second-level early warnings. Simultaneously, to protect the data sovereignty and privacy of various data sources, raw data is difficult to directly upload to the center for joint analysis, creating a dilemma of both "data silos" and "computing silos."

[0005] In summary, existing technologies lack a complete technical solution that takes into account distributed collaborative computing methods, automatic semantic fusion, edge computing, differential privacy protection, and adaptive scheduling of early warning resources, and therefore cannot meet the business needs of rapid early warning of marine disasters across the entire domain. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a federated multi-agent distributed semantic fusion method and system for marine early warning. This method and system constructs a distributed semantic fusion mechanism and adopts a deep integration architecture of federated learning (FL) and multi-agent system (MAS), which can dynamically adjust the computational accuracy and resources according to the urgency of the early warning.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A federated multi-agent distributed semantic fusion method for ocean early warning includes: S1, an edge data agent collects raw ocean data from local edge data acquisition terminals, performs outlier removal and format standardization lightweight preprocessing, extracts metadata based on the ocean feature ontology stored in a shared knowledge base, and generates multi-dimensional semantic vectors through a semantic mapping function, uploading the metadata and semantic vectors to the shared knowledge base; S2, a task coordination agent receives ocean disaster early warning tasks, calls the ocean feature ontology to perform semantic parsing on the task requirements to generate task semantic vectors, and breaks down the early warning task into several parallel fusion sub-tasks; S3, the task coordination agent performs semantic similarity calculation between the task semantic vectors and the metadata semantic vectors in the shared knowledge base, filters candidate edge data agents with semantic similarity greater than a preset threshold, publishes sub-tasks to all available analysis execution agents based on an improved contract network protocol, selects the optimal set of analysis execution agents through a comprehensive scoring formula, solves for the edge node with the best latency, and each analysis execution agent migrates to the edge node with the best latency; S4, the analysis execution agent obtains preprocessed data from the edge data agents, performs time alignment, and performs other preprocessing operations. S5: Spatial alignment generates distributed fusion intermediate results; the task coordination agent distributes the previous round's global model to all participating analysis and execution agents. Each analysis and execution agent trains a local model based on preprocessed data from its local edge data agent, calculates the model increment, adds Gaussian noise to the increment to generate an encrypted increment, and uploads it to the task coordination agent. The task coordination agent completes secure aggregation to generate the updated global model and simultaneously distributes it to all analysis and execution agents; S6: each analysis and execution agent uses the updated global model to perform sub-task disaster inference, generates encrypted local prediction results, and uploads them to the task coordination agent. The task coordination agent calculates the sub-task voting results using a weighted voting formula and outputs a standardized warning level based on the marine disaster rule base; S7: the task coordination agent calculates the urgency of the current warning task according to the task urgency calculation formula and dynamically adjusts the number of participating analysis and execution agents and the model iteration round in the next round of federated iterations based on the urgency value; S8: the task coordination agent pushes out the marine disaster warning level and the fused sea area data visualization results, and writes the metadata, fusion error, and warning accuracy of this task into the shared knowledge base.

[0008] In this invention, preferably, in step S4, the calculation formula for fusing intermediate results is as follows: , in, To integrate intermediate results, For spatiotemporal alignment operators, For multi-source preprocessed data splicing operators, This is the local preprocessed dataset for the nth site.

[0009] In this invention, preferably, in step S5, the aggregation method of the global model is secure weighted average aggregation, and the aggregation process only processes the encrypted model increment and does not acquire the original observation data.

[0010] In this invention, preferably, the marine feature ontology is defined as a quintuple, which includes a set of observation parameters, a set of parameter relationships, a set of marine observation instances, a set of domain axiom constraints, and a semantic vector mapping function; the standard structure of the metadata includes a unique data identifier, observation station location, observation unit, observation start time, observation end time, internal calling interface, and external preprocessing data calling interface; the unique data identifier is used to distinguish each piece of observation data, the observation station location is used to record the acquisition coordinates, and the external preprocessing data calling interface is a data retrieval channel opened to the outside world by the edge data agent.

[0011] In this invention, preferably, in step S3, the formula for calculating semantic similarity is: , in, For semantic similarity, For task semantic vectors, Metadata semantic vector, Let be the angle between the two vectors in the semantic space.

[0012] In this invention, preferably, in step S3, the objective function for solving the optimal edge node with the best time delay is: , in, For the optimal target edge node, Solving operators for the minimum value of the independent variable, For unidirectional network delay function, For the first An edge data acquisition terminal, s represents candidate edge nodes, and s represents the total number of edge data agents participating in this round of fusion task.

[0013] In this invention, preferably, in step S4, time alignment employs linear interpolation, linearly completing the target time data based on two sets of discrete observation values. The formula is as follows: , in, For the interpolation target time, The interpolation is performed using standardized oceanographic observation values ​​corresponding to the target time. This refers to the sampling time of the first set of original observation data. This refers to the sampling time of the later set of raw observation data. ; Spatial alignment uses inverse distance weighted interpolation, selecting known observation points around the target point to calculate hydrological values ​​using a weighted average. The formula is as follows: , in, To fuse the planar coordinates of the grid points to the target, for Standardized marine hydrological interpolation results, For the first Spatial weighting coefficients corresponding to each known observation point For the first Real ocean observation points Preprocessed hydrological observation data at the location, The number of nearby valid observation stations participating in the interpolation calculation. for To the Known observation points The planar Euclidean straight-line distance between them This is the distance power-law decay parameter.

[0014] In this invention, preferably, in step S6, the weighted voting formula is: , in, The overall fit score for the j-th candidate object. The total number of federal agents participating in the evaluation. For the first The trusted accuracy weights of each federated agent For the first The current operating load coefficient of each federal agent For the first Each agent gives a single-item fit score for the j-th candidate object.

[0015] In this invention, preferably, in step S7, the formula for calculating the urgency of the task is: , in, Depending on the urgency of the task, For the target sea area, In order to monitor the sea area, This represents the highest risk level currently predicted. The time elapsed since the warning was issued. Assign weights to the sea area coverage dimension. As the weight of the disaster risk level dimension, To handle the remaining time dimension weights.

[0016] A federated multi-agent distributed semantic fusion system for ocean early warning, used to execute any of the methods described above, includes a hierarchically deployed multi-type agents and a shared knowledge base; the multi-type agents include: edge data agents, edge coordination agents, analysis and execution agents, task coordination agents, and central management agents; the shared knowledge base stores ocean feature ontology, standardized metadata and multi-dimensional semantic vectors uploaded by edge data agents, providing unified domain knowledge and vector retrieval support for metadata extraction and semantic similarity calculation, and simultaneously receives task data, fusion errors, and early warning accuracy written by task coordination agents, completing closed-loop storage of task data throughout the entire process; the edge data agents are deployed on buoys, ocean radars, and underwater... The submersible and satellite ground station are used to collect raw ocean data from local observation terminals, perform outlier removal, format standardization, and lightweight preprocessing, extract metadata based on the five-tuple ocean feature ontology stored in the shared knowledge base, generate multi-dimensional semantic vectors through a semantic mapping function, and upload the metadata and semantic vectors to the shared knowledge base to provide local preprocessed ocean data to the analysis and execution agent. The edge coordination agent is deployed on the edge server of the nearshore ocean observation station to manage all edge data agents within its jurisdiction, assist in performing local observation data consistency pre-verification, preprocessed data caching and forwarding, and cooperate with the analysis and execution agent to retrieve preprocessed data from edge data agents within its jurisdiction. The central management agent is deployed in the central computer room of the shore station for... The system integrates a comprehensive knowledge base for overall management and the marine feature ontology, maintains the baseline version of the global federated model, coordinates the collaborative scheduling of multiple batches of early warning tasks across sea areas, and synchronously updates the resource ledgers of all observation nodes and edge servers, providing top-level management support for the global resource scheduling of the task coordination agent. The task coordination agent receives marine disaster early warning tasks and performs semantic parsing and task decomposition; it calculates the similarity between the task semantic vector and the metadata semantic vector, selects candidate edge data agents, and, based on the improved contract network protocol, completes the bidding for analysis and execution agents, the selection of the optimal agent, and the solution for the optimal edge migration node; it distributes the global model, receives encrypted model increments and securely aggregates them to generate the updated global model; and it receives encrypted local prediction results. The system calculates comprehensive prediction results using a weighted voting formula and outputs standardized early warning levels based on a disaster rule base. It quantifies task urgency using a task urgency calculation formula, dynamically adjusting the number of participating agents and model iteration rounds in the federated iteration. It pushes out early warning levels and sea area visualization results, and writes task operation indicators into a shared knowledge base. The analysis execution agent is a dynamically transferable computational agent, used to retrieve preprocessed data from the edge data agent after migrating to the edge node with optimal latency, and performs linear interpolation time alignment and inverse distance weighted spatial alignment operations to generate distributed fusion intermediate results. Based on the local fusion intermediate results, a local model is trained, the model increment is calculated, and Gaussian noise is superimposed to generate an encrypted increment that is then sent back to the task coordination agent.After loading and updating, the global model completes the sub-task of disaster inference and outputs encrypted local prediction results, which are then uploaded to the task coordinating agent.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This method and system construct a unified domain semantic space based on a standardized five-tuple marine feature ontology. It maps local metadata from various observation terminals and early warning task requirements into semantic vectors with unified dimensions. A cosine similarity algorithm automatically filters and matches observation data sources, eliminating the need for manual configuration of cross-device and cross-institutional data mapping rules. This addresses the data silo problem caused by the heterogeneity of multi-source marine observation data from a semantic level. The entire process achieves local retention of raw preprocessed data, uploading only lightweight metadata, semantic vectors, and model increments encrypted with Gaussian noise. It avoids interacting with confidential raw hydrological and geographic data throughout the process, and leverages a differential privacy mechanism to balance the needs of multi-institutional joint modeling with the data sovereignty security of each entity. This invention improves the contract network by optimally allocating and analyzing execution agents, and determines the optimal edge deployment nodes to achieve nearest-neighbor computation with the goal of minimizing the total network latency. Simultaneously, it considers factors such as the scope of disaster impact and risks. The system quantifies task urgency using three dimensions: level, remaining processing time, and automatically expands the computational agent and compresses model iteration rounds for high-urgency early warning tasks. This addresses the shortcomings of traditional centralized cloud computing, such as transmission congestion and fixed computing power allocation, enabling minute-level graded early warnings for marine disasters. Linear interpolation time-series alignment and inverse distance weighted spatial grid normalization eliminate spatiotemporal misalignment biases in multi-source observation data. Furthermore, it combines historical accuracy of the agent with real-time load-weighted voting to output early warning results, weakening the decision-making weight of low-reliability computing units and effectively reducing the probability of missed or false early warnings. Upon task completion, the system automatically writes task data, fusion errors, and early warning accuracy back to the shared knowledge base to complete a closed-loop knowledge iteration. Combined with a layered agent architecture, it achieves comprehensive management of regional data caching, all-domain resources, and ontology versions. The system is highly scalable, allowing for the flexible addition of various marine observation terminals to meet the continuous expansion needs of large-scale, all-domain monitoring operations. Attached Figure Description

[0018] Figure 1 This is a flowchart of the federated multi-agent distributed semantic fusion method for ocean early warning according to the present invention.

[0019] Figure 2 This is a logic diagram of the federated multi-agent distributed semantic fusion method for ocean early warning according to the present invention.

[0020] Figure 3 This is a multi-agent architecture diagram of the federated multi-agent distributed semantic fusion system for ocean early warning, as described in this invention.

[0021] Figure 4 This is a data flow diagram of the federated multi-agent distributed semantic fusion system for ocean early warning, as described in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] Please also see Figures 1 to 4 The basic idea of ​​this invention is to construct an architecture of "metadata registration - automatic semantic discovery - edge-based fusion," achieving distributed fusion where the original data does not leave the local network, only exchanging metadata and intermediate results, while ensuring semantic consistency and fusion accuracy across source data. Based on this architecture, to address the need for multi-source data collaboration and high real-time requirements in early warning, a federated learning and multi-agent system deep integration architecture is adopted to achieve dynamic adjustment of "data remains stationary while the model moves," while supporting dynamic adjustment of computational accuracy and resources based on the urgency of the early warning, balancing speed and accuracy requirements. To this end, a preferred embodiment of this invention provides a federated multi-agent distributed semantic fusion method for marine early warning, such as... Figure 1 As shown, it includes: S1, the edge data agent collects raw ocean data from the local edge data acquisition terminal, performs outlier removal and format standardization in a lightweight preprocessing manner, extracts metadata based on the ocean feature ontology stored in the shared knowledge base, and generates multi-dimensional semantic vectors through a semantic mapping function, and uploads the metadata and semantic vectors to the shared knowledge base.

[0025] The Edge Data Agent (EDA) is deployed on local edge data acquisition terminals such as buoys, underwater vehicles, marine radars, and satellite ground stations. It is responsible for lightweight preprocessing of local data, extraction of observation physical constraints (such as buoy EDA automatically identifying the location of the thermocline in the local profile, and radar EDA extracting the motion continuity constraints of the surface flow field), and metadata registration. The power consumption of a single node does not exceed 3W, and it continuously collects raw marine hydrological, ecological, and dynamic observation data output by sensors. First, a lightweight cleaning operation is performed on the raw data to remove abnormal noise data that exceeds the objective range of marine physical values. The coordinate reference, unit of measurement, and time format of all observation data are unified to complete the standardized preprocessing. Based on the pre-built five-tuple marine feature ontology in the shared knowledge base, standardized metadata containing observation points, monitoring indicators, collection time periods, and equipment types is extracted from the cleaned local data. Then, the semantic mapping function built into the ontology is called to convert the text-based metadata into dense multidimensional semantic vectors of a unified dimension. Only the metadata and semantic vectors with minimal storage space are encrypted and uploaded to the shared knowledge base. The original observation and preprocessed data are permanently stored on the local edge terminal and are not transmitted externally, thus avoiding the risk of leakage of original classified marine data from the source.

[0026] The marine feature ontology falls under the category of Domain Ontology. It is a standardized conceptual description of the marine observation domain, constructed based on ontological theory. It is a formalized, explicit, and unambiguous specification of shared and reusable knowledge in the marine environmental monitoring field, comprehensively characterizing core concepts, conceptual attributes, logical relationships between concepts, domain axioms, and semantic mapping mechanisms. Specifically, the marine feature ontology is defined as a quintuple, which includes a set of observation parameters, a set of parameter relationships, a set of marine observation instances, a set of domain axiom constraints, and a semantic vector mapping function. The standard structure of the metadata includes a unique data identifier, observation station location, observation unit, observation start time, observation end time, internal calling interface, and external preprocessing data calling interface. The unique data identifier is used to distinguish each piece of observation data, the observation station location is used to record the acquisition coordinates, and the external preprocessing data calling interface is a data retrieval channel opened to the edge data agent. The marine data feature ontology is defined as a quintuple, with the following structure: in, Represents the five-tuple oceanic feature ontology. It is a parameter set. It is a set of relations. It is a set of individual instances. It is an axiomatic set. It is a semantic vector mapping function that can map any metadata into a multidimensional semantic vector according to query requirements.

[0027] Define the metadata standard structure: Among them, M is a standardized observation metadata structured encapsulation defined by the five-tuple ocean feature ontology. It is a unified semantic carrier for edge data agents to extract from preprocessed ocean data and upload to a shared knowledge base. It fully records the full-dimensional attributes of a set of ocean observation indicators and is used to generate subsequent semantic vectors and cross-site semantic matching. It is a unique identifier for the data. It is the location of the data source. It is a dimensional unit. It is the observation start timestamp. It is the timestamp of the end of the observation. It is a local data reading and interaction interface. It is the metadata synchronization interface, which is the address of the preprocessed data service interface provided by each data agent (DA).

[0028] S2, the task coordination agent receives the early warning task of marine disaster, calls the marine feature ontology to perform semantic parsing of the task requirements to generate task semantic vectors, and breaks down the early warning task into several parallel fusion sub-tasks.

[0029] The Task Coordinator (TC) receives early warning requests for marine disasters such as red tides and tsunamis from external business platforms. It calls upon the same set of five-tuple marine feature ontology in the shared knowledge base to perform standardized semantic parsing on the business requirements such as target monitoring sea areas, hydrological indicators of concern, and disaster types in the task text. It outputs a task semantic vector with the same dimension as the edge metadata vector, thus achieving semantic space unification between business requirements and observation data. Then, according to the monitoring sea area range and observation equipment grouping, the complete early warning task is decomposed into multiple sets of parallelizable fusion subtasks. By splitting the task, the computational pressure of each batch is distributed, supporting distributed parallel federated computing.

[0030] S3, the task coordination agent performs semantic similarity calculation between the task semantic vector and the metadata semantic vector in the shared knowledge base, filters the candidate edge data agent set with semantic similarity greater than a preset threshold, publishes sub-tasks to all available analysis execution agents based on the improved contract network protocol, filters the optimal analysis execution agent set through a comprehensive scoring formula, solves the optimal edge node in terms of latency, and each analysis execution agent migrates to the optimal edge node in terms of latency.

[0031] The task coordination agent retrieves the metadata semantic vectors corresponding to all edge data agents within the shared knowledge base. It uses the cosine similarity formula to calculate the semantic similarity between the task semantic vector and each metadata vector, retaining only edge data agents with a matching degree higher than a preset threshold as the candidate data source set. It automatically filters observation stations irrelevant to the current early warning task, reducing invalid data retrieval. Based on the improved Contract Net Protocol, sub-task bidding information is published to all available analysis execution agents within the system. The improved Contract Net Protocol (CNP) is a customized optimization for the distributed fusion scenario of multi-source heterogeneous marine data, based on the traditional CNP. It expands the comprehensive evaluation dimensions of bids, introduces semantic matching degree and agent historical inference accuracy, constructs a multi-weighted bidding scoring model, and adds a scheduling step after awarding the bid to migrate the analysis execution agent to the edge node with the optimal latency. This overcomes the shortcomings of the traditional CNP, such as a single evaluation dimension, lack of marine semantic constraints, and failure to optimize network transmission latency, achieving high-precision, low-latency distributed task allocation for marine disaster early warning. By improving the comprehensive scoring formula of the Contract Network protocol, the computing power, historical inference accuracy, and real-time load of each analysis execution agent are quantitatively scored, and the group of analysis execution agents with the best comprehensive performance is selected to undertake this sub-task. The optimal deployment node (the edge node with the best latency) is solved with the goal of minimizing the total network transmission latency from the candidate edge server to all candidate edge data agents. The selected analysis execution agents are then dynamically migrated across edge servers to the edge node with the best latency, completing subsequent data retrieval and computation nearby, thus reducing network interaction latency.

[0032] Specifically, the formula for calculating semantic similarity is: , in, For semantic similarity, For task semantic vectors, Metadata semantic vector, Let be the angle between two vectors in the semantic space. For example, when When the threshold is set to 0.75 (usually a preset threshold), the two are semantically matched and included in the candidate dataset.

[0033] Specifically, the computing power, historical inference accuracy, and real-time load of each Analysis Agent (AA) are quantitatively scored, and the group of Analysis Agents with the best overall performance is selected to undertake this sub-task. The comprehensive scoring formula adopts the improved Contract Net Protol (CNP) scoring formula, which adds semantic matching degree to the traditional CNP, as follows: , in, For the first The overall score of the bid execution agent in the analysis. For the first The fusion computational cost of analyzing and performing predictions by intelligent agents. For the first Each analytical execution agent is composed of The converted historical precision score, ( To analyze the mean square error of the agent's history fusion task, the smaller the error, the higher the score. For the first The analysis agent estimates the completion time. For other analytical agents to perform the analysis on the first The analysis performs historical scoring on the agent. , For the reference dataset error, For the first The analysis execution agent performs the analysis on the first The verification error of the fusion results of the analysis execution agent. ~ As weight, satisfying The task coordinating agent publishes fusion tasks to the analysis and execution agent through the improved protocol CNP, and the analysis and execution agent selects the optimal set according to the scoring formula.

[0034] Specifically, the objective function for finding the optimal edge node with the best time delay is: , in, For the optimal target edge node, Solving operators for the minimum value of the independent variable, For unidirectional network delay function, For the first An edge data acquisition terminal, s represents candidate edge nodes, and s represents the total number of edge data agents participating in this round of fusion task.

[0035] S4 analyzes and executes the intelligent agent to obtain preprocessed data from the edge data intelligent agent, performs time alignment and spatial alignment respectively, and generates distributed fusion intermediate results.

[0036] The Analysis Agent (AA), having completed its migration and deployment, initiates data retrieval requests to the corresponding candidate edge data agents to obtain standardized preprocessed ocean data stored locally on each observation terminal. Two types of correction operations are performed: linear interpolation is used to align the temporal dimensions, filling in missing values ​​caused by varying sampling intervals between different devices; power-law interpolation is used... Inverse distance weighted interpolation is used to align spatial dimensions, converting discrete point observations into a continuous marine hydrological field under a unified grid coordinate system. Multi-source observation data, unified by both temporal and spatial benchmarks, are integrated to generate distributed fusion intermediate results stored locally at the edge for local model training, without transmitting the complete fusion dataset to the center throughout the process.

[0037] Specifically, time alignment uses linear interpolation to linearly complete the target time data based on two sets of discrete observations. The formula is as follows: , in, For the interpolation target time, The interpolation is performed using standardized oceanographic observation values ​​corresponding to the target time. This refers to the sampling time of the first set of original observation data. This refers to the sampling time of the later set of raw observation data. .

[0038] Spatial alignment uses inverse distance weighted interpolation, selecting known observation points around the target point to calculate hydrological values ​​using a weighted average. The formula is as follows: , in, To fuse the planar coordinates of the grid points to the target, for Standardized marine hydrological interpolation results, The spatial weighting coefficients are those corresponding to the i-th known observation point. For the i-th real ocean observation point Preprocessed hydrological observation data at the location, The number of nearby valid observation stations participating in the interpolation calculation. for To the i-th known observation point The planar Euclidean straight-line distance between them This is the distance power-law attenuation parameter (usually taken as 2, which minimizes spatial interpolation error).

[0039] Specifically, the calculation formula for integrating intermediate results is as follows: , in, To integrate intermediate results, For spatiotemporal alignment operators, For multi-source preprocessed data splicing operators, This is the local preprocessed dataset for the nth site.

[0040] S5, the task coordination agent distributes the previous round's global model to all participating analysis and execution agents. Each analysis and execution agent trains a local model based on the preprocessed data from the local edge data agent, calculates the model increment, adds Gaussian noise to the increment to generate an encrypted increment, and uploads it to the task coordination agent. The task coordination agent completes secure aggregation to generate the updated global model and distributes it synchronously to all analysis and execution agents.

[0041] The task coordination agent distributes the global marine disaster early warning model completed in the previous iteration to all analysis and execution agents undertaking the task. Each analysis and execution agent independently trains a local model using only the fusion intermediate results stored locally at the edge, and obtains the model increment by comparing the difference in model parameters before and after local training. Gaussian noise is added to the increment to achieve differential privacy encryption, and the encrypted increment without original data features is then sent back to the task coordination agent. The global model is aggregated using a secure weighted average aggregation method. The aggregation process only processes the encrypted model increment and does not acquire the original observation data. The task coordination agent performs secure weighted aggregation on all encrypted increments to generate an updated global early warning model, and synchronously distributes the new version of the global model to all analysis and execution agents participating in this iteration, completing a round of privacy-secure federated model update. Throughout the process, no original observation data from any organization flows out of the local edge node.

[0042] S6, each analysis execution agent uses the updated global model to perform sub-task disaster reasoning, generates encrypted local prediction results and uploads them to the task coordination agent. The task coordination agent calculates the sub-task voting results using a weighted voting formula and outputs a standardized early warning level in conjunction with the marine disaster rule base.

[0043] Each analytical agent loads the latest global model, independently performs marine disaster probabilistic inference for its own jurisdictional sub-task, generates encrypted local prediction results, and uploads them to the task coordination agent. The task coordination agent uses a weighted voting formula, combining the historical inference accuracy of each analytical agent with the current real-time computing load to calculate the weight of each local prediction result, and merges them to obtain a unified comprehensive risk prediction value for the entire domain. It then retrieves the pre-stored marine disaster classification rule base and, based on the comprehensive prediction value, outputs a standardized four-level marine disaster early warning system: no risk, low risk, medium risk, and high risk.

[0044] In weighted decision-making, when the task coordinating agent aggregates the subtask inference results from all the analysis and execution agents, it performs a weighted vote based on the historical accuracy of the analysis and execution agents and the current computational load to obtain the final result of the subtask. Specifically, the weighted voting formula is as follows: , in, The overall fit score for the j-th candidate object. The total number of federal agents participating in the evaluation. For the first The trusted accuracy weights of each federated agent For the first The current operating load coefficient of each federal agent For the first Each agent gives a single-item fit score for the j-th candidate object.

[0045] S7, the task coordination agent calculates the urgency of the current warning task according to the task urgency calculation formula, and dynamically adjusts the number of agents participating in the analysis and execution of the next round of federated iteration and the number of model iteration rounds according to the magnitude of the urgency value.

[0046] The task coordination agent invokes the task urgency calculation formula, quantifying the overall urgency of the current early warning task based on three dimensions: the proportion of the target sea area to the total monitored sea area, the preliminary predicted disaster risk level, and the remaining disaster response time. The configuration for the next round of federated iterations is dynamically adjusted based on the urgency value: higher urgency involves more analytical and execution agents participating in the calculation, reduces model iteration rounds, and prioritizes compressing the early warning output latency. For routine, low-urgency monitoring tasks, the number of participating agents is reduced, iteration rounds are increased, and prediction accuracy is improved, enabling dynamic allocation of computing resources on demand and balancing response speed and early warning accuracy.

[0047] Specifically, the formula for calculating the urgency of a task is as follows: , in, Depending on the urgency of the task, For the target sea area, In order to monitor the sea area, This represents the highest risk level currently predicted. The time elapsed since the warning was issued. Assign weights to the sea area coverage dimension. As the weight of the disaster risk level dimension, To handle the remaining time dimension weights, the larger the weight, the more urgent the task.

[0048] S8, the task coordination agent pushes out the marine disaster early warning level and the visualization results of the fused sea area data, and writes the metadata, fusion error and early warning accuracy of this task into the shared knowledge base.

[0049] The task coordination agent pushes standardized disaster warning levels and visualized charts of fused marine hydrological data to the marine early warning service platform. Simultaneously, it writes the metadata corresponding to this task, multi-source data fusion calculation errors, and the accuracy of this warning inference into a shared knowledge base for persistent storage. This continuously expands the marine feature ontology semantic matching samples and analyzes the historical performance ledger of the executing agent, forming a closed-loop optimization mechanism of "task execution - indicator write-back - knowledge iteration," providing continuously updated reference data for subsequent semantic matching and agent optimization scheduling.

[0050] The basic logic of a federated multi-agent distributed semantic fusion method for ocean early warning is as follows: Figure 2 As shown, the process executes sequentially: task access (task coordination agent), semantic parsing, metadata query, and analysis execution agent migration; then, the task urgency E is determined; when the task urgency is low, all agents collaborate and fuse, performing fusion computation and federated learning, and after model training, the model is output as an early warning; when the task urgency is high, some key agents collaborate and fuse, with the edge analysis execution agent calling the nearest edge data agent to perform parallel linear interpolation and inverse distance spatial weighted operations, generating a fusion result F before outputting an early warning; finally, a unified early warning result is output. The task coordination agent is responsible for global task scheduling, the analysis execution agent undertakes data fusion and model computation, and the edge data agent is deployed on the marine observation terminal, responsible for local marine data acquisition and preprocessing.

[0051] This invention also provides a federated multi-agent distributed semantic fusion system for ocean early warning, which can be used to execute the aforementioned federated multi-agent distributed semantic fusion method for ocean early warning, including hierarchically deployed multi-type agents and a shared knowledge base.

[0052] like Figure 3 As shown, the various types of intelligent agents include: edge data agents, edge coordination agents, analysis and execution agents, task coordination agents, and central management agents (CMA).

[0053] The shared knowledge base is used to store standardized metadata and multi-dimensional semantic vectors uploaded by marine feature ontology and edge data agents. It provides unified domain knowledge and vector retrieval support for metadata extraction and semantic similarity calculation. At the same time, it receives task data, fusion error and early warning accuracy written by task coordination agents, and completes closed-loop storage of task data throughout the entire process.

[0054] Edge data agents are deployed on buoys, marine radars, underwater vehicles, and satellite ground stations to collect raw marine data from local observation terminals. They perform lightweight preprocessing such as outlier removal and format standardization. Based on the five-tuple marine feature ontology stored in the shared knowledge base, metadata is extracted and multidimensional semantic vectors are generated through a semantic mapping function. The metadata and semantic vectors are then uploaded to the shared knowledge base to provide locally preprocessed marine data to the analysis and execution agents.

[0055] The edge coordination agent is deployed on the edge server of the nearshore marine observation station to manage all edge data agents within the jurisdiction, assist in completing the local observation data consistency pre-verification, pre-processed data caching and forwarding, and cooperate with the analysis and execution agent to retrieve the pre-processed data of the edge data agents within the jurisdiction.

[0056] The central management agent is deployed in the onshore station's central computer room. It is used for the overall control and sharing of the knowledge base and the iteration of the marine feature ontology version, maintaining the baseline version of the global federated model, coordinating the collaborative scheduling of multiple batches of early warning tasks across sea areas, and synchronously updating the resource ledgers of all observation nodes and edge servers. It provides top-level control support for the global resource scheduling of the task coordination agent. The central management agent is the top-level overall control intelligent unit of the entire marine early warning federated multi-agent system. It is uniformly deployed in the onshore station's central computer room and is the global control hub that overrides the task coordination agent (TC), edge coordination agent, data agent (DA), and analysis agent (AA). It does not directly participate in local operations such as single batch sub-task scheduling, local data calculation, and model aggregation. It focuses on the overall management of all resources, knowledge base, model baseline, and cross-sea area tasks.

[0057] The task coordination agent receives marine disaster early warning tasks and performs semantic parsing and task decomposition; it calculates the similarity between the task semantic vector and the metadata semantic vector, selects candidate edge data agents, and, based on the improved contract network protocol, completes the bidding for analysis and execution agents, selection of the optimal agent, and solution for the optimal edge migration node; it distributes the global model, receives encrypted model increments, and securely aggregates them to generate an updated global model; it receives encrypted local prediction results, calculates the comprehensive prediction results using a weighted voting formula, and outputs standardized early warning levels in conjunction with the disaster rule base; it quantifies the urgency of the task using a task urgency calculation formula, dynamically adjusts the number of agents participating in the federated iteration, and the number of model iteration rounds; it pushes out early warning levels and marine area visualization results, and writes task operation indicators into a shared knowledge base.

[0058] The analysis execution agent is a dynamically transferable computational agent used to retrieve edge data from the edge data agent after migrating to the edge node with the optimal latency. It performs linear interpolation time alignment and inverse distance weighted spatial alignment operations to generate distributed fusion intermediate results. Based on the local fusion intermediate results, a local model is trained, the model increment is calculated and Gaussian noise is superimposed to generate encrypted increments, which are then sent back to the task coordination agent. After loading and updating, the global model completes the sub-task disaster inference and outputs encrypted local prediction results, which are then uploaded to the task coordination agent.

[0059] Data flow of a federated multi-agent distributed semantic fusion system for ocean early warning, such as Figure 4 As shown, the data flow is as follows: First, the data source outputs multiple types of raw ocean observation data, including buoy data, radar data, satellite data, and underwater observation data. The raw data is input into the edge data agent preprocessing unit, where outlier removal, format standardization, time and coordinate unification, feature extraction, and metadata generation are performed sequentially to complete local lightweight processing of the observation data. After preprocessing, the data enters the metadata and semantic processing stage, generating standardized metadata and obtaining semantic vectors based on ocean feature ontology mapping. The metadata and semantic vectors are simultaneously pushed to a shared knowledge base for storage, supporting task semantic retrieval and matching. The analysis and execution agent retrieves the preprocessed observation data according to scheduling instructions and performs data acquisition sequentially. The process involves spatiotemporal alignment, multi-source fusion, and local model training. Spatiotemporal alignment achieves a unified spatiotemporal benchmark for multi-source observation data through linear interpolation and inverse distance weighted interpolation. Each analysis agent pushes encrypted model update information to the federated learning unit, which performs gradient aggregation, iteratively generates a global model, and distributes the updated global model to each analysis agent. Each analysis agent outputs local prediction results and pushes them to the early warning output unit. The early warning output unit performs weighted fusion of multiple sets of local prediction results, completes the early warning determination based on marine disaster rules, and releases early warning information. The metadata, model parameters, and early warning records generated in this round of tasks are synchronously written back to the shared knowledge base, achieving a closed-loop process.

[0060] This invention presents a federated multi-agent distributed semantic fusion method and system for marine early warning. It constructs a distributed multi-agent system composed of edge data agents, edge coordination agents, analysis and execution agents, and a central management agent. Employing a "metadata registration + semantic discovery + edge fusion" approach, it achieves unified organization and collaborative processing of heterogeneous data from multiple sources, including buoys, radar, satellites, and underwater observation equipment. Raw data does not need to be centrally uploaded; cross-source fusion is completed by exchanging only metadata and intermediate results, effectively reducing network transmission pressure, improving fusion efficiency, and solving the problem of data silos in the ocean. The system combines federated learning with multi-agent collaborative computation, where each node uses local data to complete model training and only uploads model parameter updates, achieving a dynamic collaborative mechanism. Simultaneously, combined with a task scheduling strategy, the number of agents participating in the computation and the number of model iterations are automatically adjusted according to the urgency of the early warning task, improving disaster early warning response speed and prediction accuracy while ensuring data security.

[0061] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A federated multi-agent distributed semantic fusion method for ocean early warning, characterized in that, include: S1, the edge data agent collects raw ocean data from the local edge data acquisition terminal, performs outlier removal and format standardization lightweight preprocessing, extracts metadata based on the ocean feature ontology stored in the shared knowledge base, and generates multi-dimensional semantic vectors through the semantic mapping function, and uploads the metadata and semantic vectors to the shared knowledge base; S2, the task coordination agent receives the early warning task of marine disaster, calls the marine feature ontology to perform semantic parsing of the task requirements to generate task semantic vectors, and breaks down the early warning task into several parallel fusion sub-tasks; S3, the task coordination agent performs semantic similarity calculation between the task semantic vector and the metadata semantic vector in the shared knowledge base, filters the candidate edge data agent set with semantic similarity greater than a preset threshold, publishes sub-tasks to all available analysis execution agents based on the improved contract network protocol, filters the optimal analysis execution agent set through a comprehensive scoring formula, solves the optimal edge node in terms of latency, and each analysis execution agent migrates to the optimal edge node in terms of latency; S4, the analysis execution agent obtains preprocessed data from the edge data agent, performs time alignment and spatial alignment respectively, and generates distributed fusion intermediate results; S5, the task coordination agent distributes the previous round of global model to all participating analysis and execution agents. Each analysis and execution agent trains a local model based on the preprocessed data of the local edge data agent, calculates the model increment, adds Gaussian noise to the increment to generate an encrypted increment, and uploads it to the task coordination agent. The task coordination agent completes secure aggregation to generate the updated global model and distributes it to all analysis and execution agents simultaneously. S6, each analysis execution agent uses the updated global model to perform sub-task disaster reasoning, generates encrypted local prediction results and uploads them to the task coordination agent. The task coordination agent calculates the sub-task voting results through a weighted voting formula and outputs a standardized early warning level in combination with the marine disaster rule base. S7, the task coordination agent calculates the urgency of the current early warning task according to the task urgency calculation formula, and dynamically adjusts the number of agents participating in the analysis and execution of the next round of federated iteration and the model iteration round according to the magnitude of the urgency value; S8, the task coordination agent pushes out the marine disaster early warning level and the visualization results of the fused sea area data, and writes the metadata, fusion error and early warning accuracy of this task into the shared knowledge base.

2. The federated multi-agent distributed semantic fusion method for ocean early warning as described in claim 1, characterized in that, In step S4, the calculation formula for the fusion intermediate results is as follows: ， in, To integrate intermediate results, For spatiotemporal alignment operators, For multi-source preprocessed data splicing operators, This is the local preprocessed dataset for the nth site.

3. The federated multi-agent distributed semantic fusion method for ocean early warning as described in claim 1, characterized in that, In S5, the global model is aggregated using a secure weighted average aggregation method. The aggregation process only processes the encrypted model increment and does not acquire the original observation data.

4. The federated multi-agent distributed semantic fusion method for ocean early warning according to claim 1, characterized in that, The marine feature ontology is defined as a quintuple, which includes a set of observation parameters, a set of parameter relationships, a set of marine observation instances, a set of domain axiom constraints, and a semantic vector mapping function. The standard structure of the metadata includes a unique data identifier, observation station location, observation unit, observation start time, observation end time, internal calling interface, and external preprocessing data calling interface. The unique data identifier is used to distinguish each piece of observation data, the observation station location is used to record the acquisition coordinates, and the external preprocessing data calling interface is a data retrieval channel opened to the outside world by the edge data agent.

5. The federated multi-agent distributed semantic fusion method for ocean early warning according to claim 1, characterized in that, In S3, the formula for calculating semantic similarity is: ， in, For semantic similarity, For task semantic vectors, Metadata semantic vector, Let be the angle between the two vectors in the semantic space.

6. The federated multi-agent distributed semantic fusion method for ocean early warning according to claim 1, characterized in that, In S3, the objective function for finding the edge node with the optimal time delay is: ， in, For the optimal target edge node, Solving operators for the minimum value of the independent variable, For unidirectional network delay function, For the first An edge data acquisition terminal, s represents candidate edge nodes, and s represents the total number of edge data agents participating in this round of fusion task.

7. The federated multi-agent distributed semantic fusion method for ocean early warning according to claim 1, characterized in that, In S4, time alignment uses linear interpolation to linearly complete the target time data based on two sets of discrete observation values. The formula is as follows: ， in, For the interpolation target time, The interpolation is performed using standardized oceanographic observation values ​​corresponding to the target time. This refers to the sampling time of the first set of original observation data. This refers to the sampling time of the later set of raw observation data. ; Spatial alignment uses inverse distance weighted interpolation, selecting known observation points around the target point to calculate hydrological values ​​using a weighted average. The formula is as follows: ， in, To fuse the planar coordinates of the grid points to the target, for Standardized marine hydrological interpolation results, The spatial weighting coefficients are those corresponding to the i-th known observation point. For the i-th real ocean observation point Preprocessed hydrological observation data at the location, The number of nearby valid observation stations participating in the interpolation calculation. for To the i-th known observation point The planar Euclidean straight-line distance between them This is the distance power-law decay parameter.

8. The federated multi-agent distributed semantic fusion method for ocean early warning according to claim 1, characterized in that, In S6, the weighted voting formula is: ， in, The overall fit score for the j-th candidate object. The total number of federal agents participating in the evaluation. For the first The trusted accuracy weights of each federated agent For the first The current operating load coefficient of each federal agent For the first Each agent gives a single-item fit score for the j-th candidate object.

9. The federated multi-agent distributed semantic fusion method for ocean early warning according to claim 1, characterized in that, In S7, the formula for calculating the urgency of the task is: ， in, Depending on the urgency of the task, For the target sea area, In order to monitor the sea area, This represents the highest risk level currently predicted. The time elapsed since the warning was issued. Assign weights to the sea area coverage dimension. As the weight of the disaster risk level dimension, To handle the remaining time dimension weights.

10. A federated multi-agent distributed semantic fusion system for ocean early warning, characterized in that: The method for performing any one of claims 1-9 includes a hierarchically deployed multi-type intelligent agents and a shared knowledge base; The various types of intelligent agents include: edge data intelligent agents, edge coordination intelligent agents, analysis and execution intelligent agents, task coordination intelligent agents, and central management intelligent agents; The shared knowledge base is used to store standardized metadata and multi-dimensional semantic vectors uploaded by marine feature ontology and edge data intelligent agents, providing unified domain knowledge and vector retrieval support for metadata extraction and semantic similarity calculation. At the same time, it receives task data, fusion error and early warning accuracy written by task coordination intelligent agents, and completes closed-loop storage of task data throughout the entire process. The edge data agent is deployed on buoys, marine radars, underwater vehicles, and satellite ground stations to collect raw marine data from local observation terminals, perform outlier removal and format standardization in lightweight preprocessing, extract metadata based on the five-tuple marine feature ontology stored in the shared knowledge base, generate multi-dimensional semantic vectors through a semantic mapping function, upload the metadata and semantic vectors to the shared knowledge base, and provide local preprocessed marine data to the analysis execution agent. The edge coordination agent is deployed on the edge server of the nearshore marine observation station. It is used to manage all edge data agents in the jurisdiction, assist in completing local observation data consistency pre-verification, pre-processed data caching and forwarding, and cooperate with the analysis and execution agent to retrieve the pre-processed data of the edge data agents in the jurisdiction. The central management agent is deployed in the central computer room of the shore station. It is used for the full-domain control and sharing of the knowledge base and the version iteration of the marine feature ontology, maintaining the baseline version of the global federated model, coordinating the collaborative scheduling of multiple batches of early warning tasks across sea areas, and synchronously updating the resource ledgers of the full-domain observation nodes and edge servers, providing top-level control support for the global resource scheduling of the task coordination agent. The task coordination agent is used to receive marine disaster early warning tasks and complete semantic parsing and task decomposition; to calculate the similarity between the task semantic vector and the metadata semantic vector, and to screen candidate edge data agents; to complete the bidding for analysis and execution agents, the selection of the optimal agent, and the solution of the optimal edge migration node based on the improved contract network protocol; to distribute the global model, receive encrypted model increments, and securely aggregate them to generate an updated global model; to receive encrypted local prediction results, calculate the comprehensive prediction results through a weighted voting formula, and output a standardized early warning level in conjunction with the disaster rule base; to quantify the urgency of the task through a task urgency calculation formula, and dynamically adjust the number of agents participating in the federated iteration and the model iteration rounds; to push out early warning levels and marine area visualization results, and to write task operation indicators into the shared knowledge base; The analysis execution agent is a dynamically transferable computational agent. After migrating to the edge node with the optimal latency, it retrieves the edge data agent's preprocessed data and performs linear interpolation time alignment and inverse distance weighted spatial alignment operations to generate distributed fusion intermediate results. Based on the local fusion intermediate results, it trains a local model, calculates the model increment, and adds Gaussian noise to generate an encrypted increment, which is then sent back to the task coordination agent. After loading and updating, the global model completes the sub-task disaster inference and outputs encrypted local prediction results, which are then uploaded to the task coordination agent.

Images