Overwater emergency rescue path prediction and rescue collaborative optimization system and method

Abstract

The invention discloses an overwater emergency rescue path prediction and rescue collaborative optimization system and method, and belongs to the technical field of artificial intelligence and overwater emergency rescue, and the system comprises a multi-source data acquisition and processing module, a federal learning path prediction module, a multi-target collaborative optimization module and a rescue command and dispatch module. The method comprises the following steps that: each rescue subject locally trains a water area environment feature model; generating a global path prediction model through federal learning safety aggregation; carrying out multi-rescue-force collaborative optimization based on a prediction result; and dynamically generating an optimal rescue scheme and commanding execution. According to the method, on the premise of protecting data privacy of all rescue subjects, high-precision rescue path prediction and efficient cooperation of multiple rescue forces can be achieved, and the success rate and efficiency of water emergency rescue are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and water emergency rescue technology, specifically to a water emergency rescue path prediction and rescue collaborative optimization system and method. Background Technology

[0002] Maritime emergency rescue faces severe technical challenges and a complex natural environment; its technological development directly affects rescue efficiency and success rate. Traditional maritime rescue systems have revealed several key technical bottlenecks in long-term application:

[0003] Data silos and privacy dilemmas are the primary challenges facing modern maritime rescue. Various rescue entities, including maritime authorities, maritime rescue centers, port authorities, and fishing associations, have created significant data silos due to differences in management systems, data standards, and trade secrets. Each entity has accumulated valuable local rescue data through long-term practice, containing unique waterway characteristics and rescue experience. However, for privacy and security reasons, these entities are generally unwilling to share raw data, and this data fragmentation severely restricts the improvement of overall rescue effectiveness.

[0004] Insufficient accuracy in path prediction is a key factor affecting rescue success rates. Existing methods for predicting the drift paths of people or vessels in the water are mainly based on simplified physical models or empirical formulas, such as the traditional Leeway model. These methods cannot effectively integrate multi-source heterogeneous environmental data, especially in complex sea conditions, where traditional models struggle to accurately characterize the nonlinear drift behavior under the coupled effects of wind, waves, and currents. Practical applications show that in severe sea conditions, traditional prediction methods have significant errors, severely impacting the precise deployment of rescue forces.

[0005] Weak collaborative optimization capabilities constrain the effectiveness of multi-rescue force coordination. In complex rescue scenarios involving multiple forces, the coordination between these forces lacks scientific optimization methods, relying primarily on the commander's personal experience and intuitive judgment. In rescue scenarios with limited resources and tight deadlines, it is difficult to quickly generate optimized solutions that consider multiple constraints and objectives, leading to unreasonable allocation of rescue resources, poor action coordination, and an inability to form a cohesive rescue force.

[0006] Poor adaptability to dynamic environments is another significant drawback of traditional systems. The water rescue environment is highly dynamic and uncertain, with environmental factors such as wind direction and current speed changing in real time. Traditional static rescue plans lack effective online learning and adjustment mechanisms, making it difficult to adapt to rapid environmental changes, often resulting in significantly reduced effectiveness of the rescue plans during actual execution.

[0007] Federated learning, as an emerging privacy-preserving machine learning paradigm, theoretically offers a new approach to solving the aforementioned problems. However, directly applying a general federated learning framework to the specific scenario of maritime emergency rescue faces numerous challenges: the aquatic environment has complex spatiotemporal characteristics, the data distribution of various rescue entities exhibits significant heterogeneity, rescue missions have extremely high real-time requirements, and multiple optimization objectives often conflict. To address these challenges, a maritime emergency rescue path prediction and collaborative optimization system is proposed. Summary of the Invention

[0008] The technical problem to be solved by this invention is: how to solve the problems of data silos among various rescue entities, low accuracy of rescue path prediction, and difficulty in coordination among multiple rescue forces in traditional water emergency rescue, and to provide a water emergency rescue path prediction and rescue coordination optimization system.

[0009] The present invention solves the above-mentioned technical problems through the following technical solutions: the present invention includes a multi-source data acquisition and processing module, a federated learning path prediction module, a multi-objective collaborative optimization module, and a rescue command and dispatch module;

[0010] The multi-source data acquisition and processing module is deployed in each rescue entity to collect water environment data, rescue resource status data and historical rescue case data, and to perform data cleaning, standardization and spatiotemporal feature extraction processing.

[0011] The federated learning path prediction module is used to securely aggregate the local model parameters of each rescue entity and generate a global rescue path prediction model by adopting a dual protection mechanism based on homomorphic encryption and differential privacy.

[0012] The multi-objective collaborative optimization module is used to establish optimization objectives based on path prediction results and environmental constraints, and to solve for the Pareto optimal solution set using an improved multi-objective particle swarm optimization algorithm.

[0013] The rescue command and dispatch module is used to select the optimal rescue plan from the Pareto optimal solution set, coordinate the execution of various rescue forces, and monitor and adjust the rescue strategy in real time.

[0014] Furthermore, in the multi-source data acquisition and processing module, the aquatic environment data includes real-time marine meteorological data and aquatic topographic data. The real-time marine meteorological data includes wind speed, wind direction, wave height, and current speed, while the aquatic topographic data includes water depth, reef distribution, and channel information. The rescue resource status data includes the location and status of rescue vessels, the endurance of drones, and the status of personnel and equipment. The historical rescue case data includes successful and failed cases, rescue time statistics, and resource consumption records.

[0015] Furthermore, in the multi-source data acquisition and processing module, spatiotemporal feature extraction is the extraction of spatiotemporal features with predictive value from standardized data. Spatiotemporal features include time features and spatial features. Time features include short-term features, medium-term features and long-term features. Spatial features take into account the gridding of water areas. Each grid cell contains complete environmental parameters, including real-time marine meteorological data and water area topography data.

[0016] Furthermore, in the federated learning path prediction module, a secure aggregation algorithm based on homomorphic encryption is adopted, and an attention mechanism is introduced to dynamically adjust the aggregation weights of the local model parameters of each rescue entity. The aggregation weight calculation formula is as follows:

[0017] ;

[0018] in, Score the data quality. To score for regional representativeness, As a contribution to history, For learnable parameters, This represents the total number of clients participating in federated learning; during the federated learning process, the clients are the rescue entities.

[0019] Furthermore, in the federated learning path prediction module, differential privacy protection processing involves adding differential privacy noise to the model gradient using a Gaussian mechanism after the local model training is completed.

[0020] Furthermore, in the federated learning path prediction module, the local model is built based on the MST-GNN model, including a meteorological perception branch, a hydrodynamic modeling branch, a terrain coding branch, a resource status branch, and a fusion center;

[0021] The meteorological sensing branch is used to capture the dynamic impact of external meteorological conditions on the water environment. This branch receives time-series meteorological parameter data from each grid cell, first extracts the meteorological change pattern at each location through time-series convolution, and then performs graph convolution operation on a graph that reflects the spatial continuity of the atmospheric field, so that the meteorological information of adjacent areas can be smoothed and interacted, and finally generates a feature representation containing the regional meteorological scenario. The time-series meteorological parameter data includes wind speed, wind direction, and wave height.

[0022] The hydrodynamic modeling branch is used to simulate the advection and diffusion processes of substances within water bodies. Based on gridded hydrodynamic data and historical water quality data, this branch constructs a directed weighted graph according to the actual water flow direction. On the directed weighted graph, diffusion convolution is applied to explicitly simulate the transport of substances from upstream to downstream, thereby learning how water environment parameters are dynamically redistributed in space by water flow and obtaining corresponding feature representations. The hydrodynamic data includes flow velocity and flow direction.

[0023] The terrain coding branch is used to encode relatively static geographic background features. This branch processes the inherent properties of each grid cell and transforms them into feature vectors that characterize the terrain environment through feature embedding and a fully connected network. It can also be incorporated into the local spatial context through lightweight graph convolution to form a feature representation that reflects long-term, stable geographic constraints. The inherent properties include water depth, seabed type, and shoreline distance.

[0024] The resource status branch is used to encode the real-time status and availability of rescue resources. This branch receives real-time status parameters from each rescue unit and transforms discrete or continuous real-time status parameters into feature vectors of a unified dimension through temporal embedding and state coding networks. A graph attention mechanism is introduced to capture the coordination and constraint relationships between multiple resources, forming a feasible dynamic representation that reflects the current distribution and effectiveness of rescue resources, and obtaining the corresponding feature representation. Among them, the real-time status parameters include location, endurance, equipment integrity, and personnel fatigue.

[0025] The fusion center, serving as the decision hub of the model, adaptively integrates the heterogeneous information from the four branches. First, the feature representations of each branch are aligned and stitched together at the node level. Then, the contribution of different physical processes is dynamically weighed through gating fusion and attention mechanisms to obtain deeply fused features. Finally, the deeply fused features are fed into a spatiotemporal prediction head, which performs final multi-round spatiotemporal evolution inference on a unified graph, and outputs the prediction results of future changes in water environment parameters.

[0026] Furthermore, in the federated learning path prediction module, the global rescue path prediction model is an agent based on a deep reinforcement learning framework. The specific process by which the global rescue path prediction model generates path prediction results is as follows:

[0027] S11: Multi-source Spatiotemporal State Coding and Fusion

[0028] The heterogeneous client prediction fields are calibrated and fused into a unified dynamic probability field using a spatiotemporal attention network. At the same time, the task constraints and navigation graph structure are encoded into machine-readable feature representations to obtain the corresponding spatiotemporal graph, where the heterogeneous client prediction fields are output by each local model.

[0029] S12: Serialization Path Decision Generation

[0030] By performing a Markov decision process on the spatiotemporal graph encoded in step S11, each decision comprehensively evaluates the immediate navigation cost, long-term goal orientation, and prediction of future environmental risks, and iteratively generates an original spatiotemporal path.

[0031] S13: Multipath Optimization and Post-processing Output

[0032] By sampling, perturbing, or parallel inference on the original spatiotemporal path generated in step S12, multiple candidate spatiotemporal paths are generated. After multi-objective evaluation and Pareto front screening, the optimal path is obtained. The optimal path is then dynamically smoothed and corrected. Finally, the optimized rescue trajectory scheme with a precise timestamp is output, which is the path prediction result. The whole process is iteratively optimized under the federated learning framework, which enables the model to continuously integrate global knowledge and realize intelligent path planning in dynamic environments.

[0033] Specifically, the global rescue path prediction model receives and integrates four types of input information:

[0034] Multi-client environment prediction field: The gridded spatiotemporal prediction results of target parameters for a future period of time generated by each rescue entity based on its local MST-GNN model constitute a multi-source spatiotemporal data field describing the dynamic evolution of risks or targets;

[0035] Rescue mission constraints include specific mission descriptions such as starting location, target area or location, mission type, and performance parameters of the vehicle (maximum speed, turning radius, endurance) and mission deadline.

[0036] Global static / quasi-static navigation map: A unified spatial map covering all relevant waters, where nodes represent the center points of regularly or irregularly divided grid cells, and edges represent navigable connections between adjacent water cells. Each edge includes a basic navigation cost and can identify fixed obstacles (such as reefs, shoals, and bridges).

[0037] Wide-area real-time status information: Integrates regional-level real-time or short-term forecast data, such as large-scale wind fields, ocean current fields, and storm warning information, to correct the global environmental background field.

[0038] Furthermore, in the multi-objective collaborative optimization module, the optimization objectives are to minimize the total rescue time, maximize the rescue success rate, minimize the rescue resource consumption, and minimize the rescue risk, as detailed below:

[0039] Minimize total rescue time:

[0040] ;

[0041] in, The total number of units participating in the rescue operation. A unit participating in the rescue operation refers to an independent mobile vehicle or operational unit that is dispatched and deployed to perform search, rescue, or support tasks in a specific rescue mission. Let be the travel time of the i-th unit. Let be the search time for the i-th unit. Let i be the execution time of the rescue operation for the i-th unit. For path planning variables, For rescue strategy variables;

[0042] Maximizing the success rate of rescue:

[0043] ;

[0044] in, The number of targets awaiting rescue. Let be the probability of finding the j-th target in the i-th unit. Let i be the probability that the i-th rescue unit successfully rescues the j-th target.

[0045] Minimize the consumption of rescue resources:

[0046] ;

[0047] in, The term "rescue resources" refers to the total number of resource types. It encompasses all equipment, vehicles, personnel, and consumables that can be used for maritime emergency rescue missions. Resource types are categorized based on their core functions, power attributes, operating costs, and risk characteristics. Let be the fuel cost coefficient for resource type k. Let k be the time cost coefficient for the k-th type of resource. Let be the equipment loss coefficient for the k-th type of resource. The total sailing distance corresponding to the k-th type of resource. Let k be the total usage time of the k-th type of resource. Let be the consumption quantity of the k-th type of resource. For path planning variables, For time scheduling variables, Assign variables to resources;

[0048] Minimize rescue risks:

[0049] ;

[0050] in, The total number of resource types, For environmental risks, For equipment risks, For personnel risks, For path planning variables, Assign variables to resources. For rescue strategy variables.

[0051] Furthermore, in the multi-objective collaborative optimization module, the multi-objective particle swarm optimization algorithm introduces a dynamic weight adjustment mechanism and an elite retention strategy to ensure that the algorithm achieves an optimal balance between convergence speed and solution set diversity; wherein, the weight update formula is as follows:

[0052] ;

[0053] in, =0.9, =0.4, where t is the current iteration number. This represents the maximum number of iterations.

[0054] This invention also provides a method for predicting water emergency rescue routes and optimizing rescue coordination, applied to the above-mentioned system, comprising the following steps:

[0055] S1: Each rescue entity trains a local rescue path prediction model based on local historical data. The local rescue path prediction model is the ontology model.

[0056] S2: Securely aggregate model parameters from various rescue entities using a federated learning framework to generate a global rescue path prediction model;

[0057] S3: Obtain path prediction results through the global rescue path prediction model, and perform multi-objective collaborative optimization calculations based on the path prediction results and environmental constraints to obtain the Pareto optimal solution set;

[0058] S4: Generate the optimal rescue plan through the Pareto optimal solution set and coordinate the execution of various rescue forces;

[0059] S5: Monitor the rescue process in real time and dynamically adjust the rescue strategy based on feedback data.

[0060] Furthermore, in step S1, the loss function during training is as follows:

[0061] ;

[0062] in, Loss for trajectory prediction; Quantify the loss for uncertainty; Loss due to physical constraints; , This is a hyperparameter.

[0063] The present invention has the following advantages over the prior art:

[0064] 1. A breakthrough in prediction accuracy has been achieved, providing reliable technical support for precision rescue. Rapid response and efficient rescue have greatly improved the survival rate of people in distress. Resource utilization has been significantly improved; through multi-objective collaborative optimization, the utilization rate of rescue resources has been significantly increased under the same rescue effect.

[0065] 2. Significantly enhanced system robustness. The distributed architecture ensures that the system can maintain basic functions even in the event of single points of failure, communication interruptions, or other anomalies, and the regional autonomy capability is significantly improved. The system also has self-healing capabilities, automatically reconstructing service links when some functions fail.

[0066] 3. A perfect balance is achieved between privacy protection and knowledge sharing. While strictly meeting privacy protection requirements, knowledge sharing and collaborative learning across rescue entities are realized, laying the foundation for broader rescue cooperation. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the structure of the water emergency rescue path prediction and rescue collaborative optimization system based on federated learning in an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of the structure of the Multimodal Spatiotemporal Graph Neural Network (MST-GNN) in an embodiment of the present invention;

[0069] Figure 3 This is a flowchart of the water emergency rescue path prediction based on federated learning in an embodiment of the present invention;

[0070] Figure 4 This is a flowchart of the collaborative optimization process for water emergency rescue based on federated learning in an embodiment of the present invention. Detailed Implementation

[0071] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0072] Example 1

[0073] The system architecture design is the fundamental innovation of this invention. The system adopts a three-tiered collaborative distributed intelligent rescue architecture of "cloud-edge-device". The device side, as the basic unit of the system, consists of local intelligent terminals of each rescue entity, responsible for real-time data acquisition, feature engineering, and preliminary model training. These terminal devices are equipped with multimodal sensors and edge computing capabilities, enabling rapid data processing on-site in water areas. The edge side, as the regional coordination center, consists of federated learning nodes of the regional rescue center, responsible for model aggregation and preliminary optimization calculations within the region, forming a regional rescue knowledge base. The cloud side, as the global command center, is hosted by the central rescue command platform, responsible for global model aggregation, cross-regional collaborative optimization, and final command decision-making. This hierarchical architecture ensures both the effectiveness of global optimization and meets the real-time requirements of local response.

[0074] The core modules embody the technological depth of this invention. The system comprises four core modules: a multi-source data acquisition and processing module, a federated learning path prediction module, a multi-objective collaborative optimization module, and a rescue command and dispatch module. The multi-source data acquisition module employs advanced multimodal sensor fusion technology, integrating AIS receivers, meteorological sensors, hydrological monitoring equipment, and satellite communication modules to achieve comprehensive perception of aquatic environmental data and rescue resource status data. This module can collect in real-time various environmental parameters, including wind speed, wind direction, current speed, and wave height, as well as resource information such as the location and status of rescue vessels and the endurance of drones. The federated learning path prediction module designs a personalized federated learning algorithm based on an attention mechanism, achieving global knowledge sharing while protecting data privacy. This module includes four closely interconnected sub-modules: model initialization, local training, secure aggregation, and global update, forming a complete learning loop. The multi-objective collaborative optimization module develops an improved multi-objective particle swarm optimization algorithm, capable of simultaneously optimizing four key objectives: rescue time, success rate, resource consumption, and risk. This module includes three core sub-modules: problem modeling, algorithm solving, and scheme evaluation, ensuring the scientific rigor and effectiveness of the optimization process. The rescue command and dispatch module constructs an intelligent command and decision-making system, realizing the generation, distribution, execution, and monitoring of rescue plans. This module includes four functional units: plan generation, instruction distribution, status monitoring, and dynamic adjustment, forming a complete command closed loop.

[0075] The innovative methodology demonstrates the systematic thinking of this invention. This invention proposes a five-stage rescue optimization process: First, the data perception and feature engineering stage, where each rescue entity locally collects multi-source data and extracts spatiotemporal features to construct a standardized feature dataset; second, the personalized model training stage, where personalized path prediction models are trained based on local data, fully utilizing the characteristics of local data; third, the federated knowledge aggregation stage, where a high-precision global prediction model is generated through secure aggregation to achieve knowledge sharing; fourth, the multi-objective collaborative optimization stage, where rescue forces are collaboratively optimized based on prediction results to generate the optimal solution; and fifth, the dynamic execution and feedback stage, where the rescue plan is executed and continuously optimized based on feedback, forming a learning loop.

[0076] The core technology of this invention lies in its innovative core algorithms. The Multimodal Spatiotemporal Graph Neural Network (MST-GNN) forms the technological foundation for path prediction. This network innovatively designs four dedicated processing branches and one intelligent fusion center. The meteorological perception branch uses a temporal convolutional network (TCN) to extract multi-scale temporal features of meteorological data such as wind speed and direction, capturing the dynamic changes in meteorological conditions. The hydrodynamic modeling branch uses a graph attention network (GAT) to capture the spatial correlation of flow and wave fields, accurately describing the characteristics of water movement. The terrain coding branch learns the topological structure of water terrain data through a graph convolutional network (GCN) to understand the impact of the geographical environment on rescue efforts. The rescue resource branch utilizes embedding techniques to represent the spatiotemporal state of rescue resources and quantifies the distribution of rescue capabilities. The cross-modal fusion center designs a cross-modal attention mechanism to achieve adaptive weighted fusion of multi-source information, ensuring the rational utilization of information from each modality. The network's total loss function comprehensively considers trajectory prediction accuracy, uncertainty quantification accuracy, and conformity to physical laws, ensuring the reliability and practicality of the prediction results.

[0077] A privacy-preserving federated learning framework is crucial for achieving secure data sharing. This framework proposes a dual privacy protection mechanism: at the communication level, Paillier homomorphic encryption ensures secure transmission of model parameters and prevents man-in-the-middle attacks; at the algorithm level, strict (ε,δ)-differential privacy prevents privacy inference attacks, providing quantifiable privacy guarantees. Simultaneously, a contribution-based incentive mechanism is introduced, comprehensively considering the prediction accuracy, data quality, and privacy costs of each rescue entity, dynamically adjusting their aggregation weights to incentivize participation. The contribution-based incentive mechanism includes:

[0078]

[0079] in, For the first The incentive score for each rescue entity is used to dynamically adjust its aggregate weight.

[0080] The Improved Multi-Objective Particle Swarm Optimization (IMOPSO) algorithm is an optimization tool tailored to the specific needs of rescue scenarios. This algorithm makes three key improvements over the traditional IMOPSO algorithm: dynamic parameter adjustment allows the inertia weight and learning factor to adaptively adjust with the iteration process, balancing exploration and development capabilities; an elite-oriented mechanism introduces external archiving and elite selection strategies to maintain solution set diversity and avoid premature convergence; and a fast convergence design employs a reference point-based environment selection to accelerate algorithm convergence, meeting the real-time requirements of rescue operations. These improvements ensure that the algorithm can generate high-quality rescue plans within a limited time.

[0081] Example 2

[0082] like Figure 1The diagram shown is a schematic representation of a water emergency rescue path prediction and rescue collaborative optimization system based on federated learning provided in this embodiment, including:

[0083] The data acquisition and processing layer is the basic support layer of the system, responsible for the acquisition, processing and feature engineering of multi-source data.

[0084] Multi-source data acquisition module: Real-time acquisition of aquatic environmental data, rescue resource status data, and historical rescue case data. Deployment of a multimodal sensor network, including AIS receivers, meteorological sensors, hydrological monitoring equipment, and satellite communication modules. Utilizing IoT technology for automatic data acquisition and transmission, supporting multiple communication methods such as 5G and satellite communication.

[0085] Data cleaning and standardization module: Performs quality control and standardization processing on raw data. Establishes a data quality assessment index system, including dimensions such as completeness, accuracy, and timeliness. Employs a combination of rule-based and machine learning methods to identify and handle outliers and missing values. Data standardization uses Z-score normalization to ensure effective integration of data from different dimensions.

[0086] The spatiotemporal feature extraction module extracts predictive spatiotemporal features from the raw data. A multi-scale feature extraction algorithm is designed, including short-term features (0-6 hours), medium-term features (6-24 hours), and long-term features (24-72 hours). Spatial features are processed using a water area grid, with each grid cell containing complete environmental parameters, including real-time marine meteorological data and water area topography data. The feature dimension reaches 128 dimensions, fully capturing the spatiotemporal variation patterns of the environment.

[0087] Quality Assessment and Control Module: Monitors data quality and ensures the reliability of the processing. It establishes a real-time data quality monitoring dashboard and sets quality thresholds for automatic alerts. Multiple copy storage and backup mechanisms are employed to ensure data security. In actual operation, this module uses data quality scores as an important basis for federated learning aggregation weights.

[0088] The federated learning path prediction layer is the core technology layer of the system, enabling accurate path prediction while protecting privacy.

[0089] Model initialization module: Establishes the initial model framework for federated learning. A multimodal spatiotemporal graph neural network (MST-GNN) is designed, including meteorological sensing branches, hydrodynamic modeling branches, terrain coding branches, resource status branches, and a fusion center. Model parameters are pre-trained based on large-scale historical data, providing a solid foundation for subsequent personalized learning.

[0090] like Figure 2The diagram shown is a schematic of the multimodal spatiotemporal graph neural network (MST-GNN) structure provided in this embodiment, including a meteorological sensing branch, a hydrodynamic modeling branch, a terrain coding branch, a resource status branch, and a fusion center;

[0091] The meteorological perception branch is mainly responsible for capturing the dynamic impact of external meteorological conditions on the water environment. This branch receives time-series data of meteorological parameters (such as wind speed, wind direction, and wave height) from each grid cell. First, it extracts the meteorological change pattern at each location through time-series convolution. Then, it performs graph convolution operation on a graph that reflects the spatial continuity of the atmospheric field, so that the meteorological information of adjacent areas can be smoothed and interacted, and finally generates a feature representation containing the regional meteorological scenario.

[0092] The core of the hydrodynamic modeling branch is to simulate the advection and diffusion processes of substances within a water body. This branch is based on gridded hydrodynamic data (such as flow velocity and direction) and historical water quality data. It constructs a directed weighted graph according to the actual water flow direction. On this graph, operations such as diffusion convolution are applied to explicitly simulate the transport of substances from upstream to downstream, thereby learning how water environment parameters are dynamically redistributed in space by water flow and obtaining the corresponding feature representations.

[0093] The topographic coding branch focuses on encoding relatively static geographic background features. This branch processes the inherent attributes of each grid cell, such as water depth, substrate type, and shoreline distance, and transforms them into feature vectors that characterize the topographic environment through feature embedding and fully connected networks. These vectors can be incorporated into the local spatial context through lightweight graph convolutions to form feature representations that reflect long-term, stable geographic constraints.

[0094] The resource status branch encodes the real-time status and availability of rescue resources. This branch receives real-time status parameters from each rescue unit (such as rescue boats, drones, and personnel), including dynamic information such as location, endurance, equipment integrity, and personnel fatigue. Through temporal embedding and a state coding network, discrete or continuous state parameters are transformed into feature vectors of a unified dimension. A graph attention mechanism is introduced to capture the collaborative and constraining relationships between multiple resources, forming a feasible dynamic representation that reflects the current distribution and effectiveness of rescue resources, and obtaining the corresponding feature representation.

[0095] The fusion center, serving as the decision hub of the model, adaptively integrates the heterogeneous information from the four branches. First, the feature vectors of each branch are aligned and concatenated at the node level. Then, the contribution of different physical processes is dynamically weighed through gating fusion and attention mechanisms to obtain deeply fused features. Finally, the deeply fused features are fed into a spatiotemporal prediction head, which performs several rounds of spatiotemporal evolution inference on a unified graph to output accurate predictions of future changes in water environment parameters.

[0096] Local training module: Model training and optimization are performed locally by each rescue entity. A multi-task learning framework is adopted, simultaneously optimizing trajectory prediction loss, uncertainty quantification loss, and physical constraint loss. The training process supports incremental learning, enabling continuous absorption of new rescue experience.

[0097] Secure Aggregation Module: Aggregates knowledge from various entities while protecting privacy. It employs a weighted aggregation strategy based on an attention mechanism, with weights dynamically calculated based on data quality, prediction accuracy, and historical contribution. Homomorphic encryption and differential privacy technologies are used to ensure the secure transmission and aggregation of model parameters.

[0098] The global update module is responsible for updating and maintaining the global rescue path prediction model and its related performance metrics after security aggregation. This module supports a complete model version management mechanism, allowing for archiving, querying, and rapid rollback of historical versions to ensure system reliability. After each update iteration, the module performs a multi-dimensional performance evaluation of the newly generated global model (global rescue path prediction model). To adapt to potential network instability at rescue sites, the module employs an asynchronous update mechanism to ensure robust execution of update tasks even under network fluctuations. Furthermore, the module establishes a systematic model performance evaluation framework, primarily covering key indicators such as prediction accuracy, generalization ability, and convergence speed, providing a quantitative basis for model iteration and optimization.

[0099] Prediction Generation Module: Generates the final path prediction results. Based on a converged global model, it takes real-time environmental data as input and outputs drift path predictions for the next 6-48 hours, including the optimal estimated path and confidence interval. The prediction results are visualized in the form of spatial probability density.

[0100] The collaborative optimization decision layer is the system's intelligent decision layer, which generates the optimal rescue plan based on the prediction results.

[0101] The optimization problem modeling module formalizes the rescue scenario into a mathematical optimization problem. It establishes a four-objective optimization model, including minimizing rescue time, maximizing success rate, minimizing resource consumption, and minimizing rescue risk. A complete system of constraints is defined, including dynamic constraints, resource constraints, time window constraints, and safety constraints.

[0102] Multi-objective optimization solution module: This module solves optimization problems and generates Pareto optimal solution sets. It employs an improved multi-objective particle swarm optimization algorithm (IMOPSO), introducing dynamic parameter tuning, an elitist-oriented mechanism, and a fast convergence design. The algorithm supports parallel computation, obtaining high-quality solution sets within a finite time.

[0103] Solution Evaluation and Selection Module: Selects the final implementation plan from the Pareto optimal solution set. Based on multi-attribute utility theory, it establishes a solution evaluation index system that comprehensively considers factors such as time urgency, resource availability, and risk tolerance. It supports interactive decision-making, allowing command personnel to participate in the solution selection process.

[0104] Dynamic re-optimization module: Adjusts rescue plans dynamically based on actual conditions. Establishes a feedback loop mechanism, triggering re-optimization based on execution results and environmental changes. Supports both local adjustment and global re-optimization modes to ensure the adaptability and effectiveness of the plan.

[0105] The application service layer is the system's business application layer, providing complete rescue service functions.

[0106] Rescue Command and Dispatch Module: Executes rescue plans and coordinates various rescue forces. Establishes an intelligent command system supporting functions such as task distribution, progress tracking, and resource coordination. Provides multiple communication interfaces to ensure timely and accurate transmission of instructions.

[0107] Real-time monitoring and early warning module: Monitors the rescue process and provides timely warnings of abnormal situations. Integrates multi-source monitoring data to establish a situational awareness dashboard. Sets up a multi-level early warning mechanism to automatically issue warnings for situations such as progress delays, resource anomalies, and sudden environmental changes.

[0108] Decision Support Display Module: Provides intuitive display of decision support information. Develops a visualization system supporting various display formats such as 2D maps, 3D scenes, and data charts. Provides a mobile application for convenient use by on-site command personnel.

[0109] Effectiveness Evaluation and Feedback Module: This module evaluates the effectiveness of rescue efforts and provides feedback for system optimization. It establishes an effectiveness evaluation index system, including dimensions such as response time, rescue success rate, and resource utilization. The feedback mechanism continuously optimizes model parameters and algorithm strategies.

[0110] Example 3

[0111] like Figure 3 The diagram shown is a flowchart of a maritime emergency rescue path prediction method based on federated learning provided in this embodiment, including the following steps:

[0112] Process Initialization: The federated learning path prediction process begins with global model initialization. The central server pre-trains the MST-GNN model based on historical data, which fully considers the spatiotemporal characteristics of the aquatic environment. Model parameters are distributed to various rescue entities through a secure channel, initiating the distributed learning process.

[0113] Local training phase: Each rescue entity trains the model individually based on local data. The training process employs a multi-task learning framework, and the loss function is designed as follows:

[0114]

[0115] in, The trajectory prediction loss is used to ensure prediction accuracy. Uncertainty Quantification Loss is designed to calibrate the uncertainty of the model's predictions, ensuring that the confidence level of the model's output matches the true statistical characteristics of the prediction error, thereby improving the reliability of probabilistic predictions. The physical constraint loss is used to ensure that the prediction results conform to physical laws; hyperparameters , This was determined through cross-validation.

[0116] Privacy protection measures: After local training, differential privacy noise is added to the model gradient; a Gaussian mechanism is used, and the noise standard deviation is calculated according to the formula:

[0117]

[0118] in, For privacy budget, To ensure gradient sensitivity, gradient clipping is controlled within the range of 1.0-3.0 to meet stringent privacy protection requirements.

[0119] Secure Aggregated Updates: The server-side employs a weighted aggregation strategy, with weights calculated based on multi-factor evaluation.

[0120]

[0121] in, The data quality score is used to quantify the reliability and completeness of the local data of the c-th rescue entity. Its calculation is based on multi-dimensional evaluation indicators of the data provided by the client (the specific reference of the rescue entity in the context of federated learning technology) during the training period. Each indicator is normalized and then calculated through a small neural network (quality evaluation network). The regional representativeness score is used to measure the uniqueness and importance of the environmental characteristics of the waters under the jurisdiction of the c-th rescue entity in the overall mission, aiming to promote the diversity of model knowledge. Its acquisition is based on geographical and mission characteristics. Historical contribution score is used to incentivize continuous, high-quality participation and smooth out occasional fluctuations in a single training round, recording the contribution of the c-th client to the global model performance improvement in past aggregation rounds. As learnable parameters, they are acquired through a lightweight meta-learning feedback adjustment mechanism deployed on the server side. This mechanism dynamically optimizes the allocation ratio of data quality, regional representativeness, and historical contribution to the final aggregation weights. Each adjustment of the allocation ratio affects... Normalization is performed to ensure that the sum is 1, thus forming a continuously self-optimizing weight allocation strategy. The total number of clients participating in federated learning.

[0122] Convergence criterion: A dual convergence criterion is used: the rate of change of the loss function is less than... Or, the maximum number of training epochs can be reached, up to 50. Simultaneously, the prediction accuracy on the validation set is monitored to ensure the model's generalization ability.

[0123] Example 4

[0124] like Figure 4 The diagram shown illustrates a collaborative optimization process for maritime emergency rescue based on federated learning, comprising the following steps:

[0125] Problem modeling phase: A four-objective optimization problem is constructed based on the path prediction results. Decision variables include path planning, time scheduling, resource allocation, and rescue strategy. The objective function comprehensively considers:

[0126] Minimize total rescue time:

[0127]

[0128] in, The total number of units participating in the rescue operation. A unit participating in the rescue operation refers to an independent mobile vehicle or operational unit that is dispatched and deployed to perform search, rescue, or support tasks in a specific rescue mission. Let be the travel time of the i-th unit (rescue unit). Let be the search time for the i-th unit. Let i be the execution time of the rescue operation for the i-th unit. For path planning variables, For rescue strategy variables.

[0129] Maximizing the success rate of rescue:

[0130]

[0131] in, The number of targets awaiting rescue. Let be the probability of finding the j-th target in the i-th unit. Let i be the probability that the i-th rescue unit successfully rescues the j-th target.

[0132] Minimize the consumption of rescue resources:

[0133]

[0134] in, The term "rescue resources" refers to the total number of resource types. It encompasses all equipment, vehicles, personnel, and consumables that can be used for maritime emergency rescue missions. Resource types are categorized based on their core functions, power attributes, operating costs, and risk characteristics. Let be the fuel cost coefficient for resource type k. Let k be the time cost coefficient for the k-th type of resource. Let be the equipment loss coefficient for the k-th type of resource. The total sailing distance corresponding to the k-th type of resource. Let k be the total usage time of the k-th type of resource. Let be the consumption quantity of the k-th type of resource. For path planning variables, For time scheduling variables, Assign variables to resources.

[0135] Minimize rescue risks:

[0136]

[0137] in, The total number of resource types, For environmental risks, For equipment risks, For personnel risks, For path planning variables, Assign variables to resources. For rescue strategy variables.

[0138] The specific constraints are as follows:

[0139] Dynamic constraints:

[0140]

[0141]

[0142]

[0143] , : These represent the velocity and acceleration vectors of the i-th unit at time t, respectively, derived from the path planning variables;

[0144] , : The minimum and maximum feasible speeds of the i-th unit;

[0145] : Maximum acceleration amplitude limit for the i-th unit;

[0146] : The turning angle of the i-th unit at time t;

[0147] : The maximum permissible steering angle of the i-th unit (related to the minimum turning radius).

[0148] Resource constraints:

[0149]

[0150] : The set of indices of all rescue units belonging to the k-th resource category;

[0151] : A binary decision variable, representing whether the i-th unit is scheduled (1 for yes, 0 for no), which is a component of the resource allocation variable;

[0152] The maximum number of units of resource type k available in the current task;

[0153] Time window constraint:

[0154]

[0155] The time it takes for the first unit to arrive at the region where target j is located is a function of the time scheduling variable;

[0156] The latest deadline for the rescue operation targeting target j is determined by the urgency of the mission and the state of the target (such as the survival limit of the person who fell into the water).

[0157] Safety constraints:

[0158]

[0159] : The position coordinates of the i-th unit at time t, defined by the path planning variables.

[0160] : The geometric representation of the o-th obstacle or danger zone (such as a reef, a no-navigation zone, or a zone with severe sea conditions).

[0161] dist(·): Calculates the shortest Euclidean distance from a point to a region.

[0162] : The preset minimum safe distance threshold.

[0163] Optimize the solution process: The improved IMOPSO algorithm is adopted, and the main improvements include dynamic inertia weight, adaptive learning factor, elite selection mechanism (maintaining solution set diversity based on crowding distance), and reference point guidance (accelerating algorithm convergence).

[0164] Dynamic inertia weight adjustment mechanism: A nonlinear adaptive decay strategy was designed. The update of the inertia weight depends not only on the number of iterations, but also on the convergence state index of the current particle swarm and the time urgency of the external rescue environment (the time pressure coefficient calculated from the mission deadline).

[0165] Adaptive learning factor mechanism: Personalized adjustment based on the density of the Pareto front region where each particle is located. For particles located in sparse front regions, the social learning factor is increased, causing them to move more towards globally known excellent regions (which may represent efficient rescue modes), guiding resources to explore these regions; for particles located in dense front regions, the individual learning factor is increased, encouraging them to leverage their individual experience to find better solutions that may be overlooked near these dense regions, thus refining the solution.

[0166] Elite selection and solution set diversity preservation mechanism: After each generation update, the non-dominated solutions in the current particle swarm are merged with an external archive. When selecting archive members to enter the next generation, a hybrid selection strategy based on crowding distance ranking and rescue target preference weights is not used alone.

[0167] Reference point-guided convergence acceleration mechanism: A set of dynamically generated reference points is introduced. Based on the target probability distribution provided by the path prediction results and the initial feasible solutions of resources and environment, the reference points delineate the most promising "potential optimal solution region" in the target space.

[0168] Solution Generation and Execution: After selecting the optimal solution from the Pareto front, a detailed rescue execution plan is generated. The plan includes a resource allocation table, route planning map, time schedule, and emergency response plan. Execution is distributed through a command system, and a real-time monitoring mechanism is established.

[0169] Dynamic adjustment mechanism: Dynamic adjustments are triggered based on execution feedback and environmental changes. Re-optimization conditions include: significant changes in environmental parameters, rescue progress lagging beyond a threshold, resource status updates, and receipt of new intelligence information. The adjustment process is completed within 3 minutes to ensure the timeliness of the plan.

[0170] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A system for predicting waterborne emergency rescue routes and optimizing rescue coordination, characterized in that, It includes a multi-source data acquisition and processing module, a federated learning path prediction module, a multi-objective collaborative optimization module, and a rescue command and dispatch module; The multi-source data acquisition and processing module is deployed in each rescue entity to collect water environment data, rescue resource status data and historical rescue case data, and to perform data cleaning, standardization and spatiotemporal feature extraction processing. The federated learning path prediction module is used to securely aggregate the local model parameters of each rescue entity and generate a global rescue path prediction model by adopting a dual protection mechanism based on homomorphic encryption and differential privacy. The multi-objective collaborative optimization module is used to establish optimization objectives based on path prediction results and environmental constraints, and to solve for the Pareto optimal solution set using an improved multi-objective particle swarm optimization algorithm. The rescue command and dispatch module is used to select the optimal rescue plan from the Pareto optimal solution set, coordinate the execution of various rescue forces, and monitor and adjust the rescue strategy in real time.

2. The water emergency rescue route prediction and rescue coordination optimization system according to claim 1, characterized in that, In the multi-source data acquisition and processing module, the aquatic environment data includes real-time marine meteorological data and aquatic topography data. The real-time marine meteorological data includes wind speed, wind direction, wave height, and current speed. The aquatic topography data includes water depth, reef distribution, and channel information. The rescue resource status data includes the location and status of rescue vessels, the endurance of drones, and the status of personnel and equipment. The historical rescue case data includes successful and failed cases, rescue time statistics, and resource consumption records.

3. The water emergency rescue route prediction and rescue coordination optimization system according to claim 1, characterized in that, In the multi-source data acquisition and processing module, spatiotemporal feature extraction involves extracting predictive spatiotemporal features from standardized data. These features include temporal and spatial characteristics. Temporal features include short-term, medium-term, and long-term features, while spatial features are considered... The water area is processed into a grid, with each grid cell containing complete environmental parameters, including real-time marine meteorological data and water topography data.

4. The water emergency rescue route prediction and rescue coordination optimization system according to claim 1, characterized in that, In the federated learning path prediction module, a secure aggregation algorithm based on homomorphic encryption is adopted, and an attention mechanism is introduced to dynamically adjust the aggregation weights of the local model parameters of each rescue entity. The aggregation weight calculation formula is as follows: ； in, Score the data quality. To score for regional representativeness, As a contribution to history, For learnable parameters, This represents the total number of clients participating in federated learning; during the federated learning process, the clients are the rescue entities.

5. The water emergency rescue path prediction and rescue coordination optimization system according to claim 4, characterized in that, In the federated learning path prediction module, differential privacy protection processing involves adding differential privacy noise to the model gradient using a Gaussian mechanism after the local model training is completed.

6. The water emergency rescue path prediction and rescue coordination optimization system according to claim 5, characterized in that, In the federated learning path prediction module, the local model is built based on the MST-GNN model, including a meteorological perception branch, a hydrodynamic modeling branch, a terrain coding branch, a resource status branch, and a fusion center; The meteorological sensing branch is used to capture the dynamic impact of external meteorological conditions on the water environment. This branch receives time-series meteorological parameter data from each grid cell, first extracts the meteorological change pattern at each location through time-series convolution, and then performs graph convolution operation on a graph that reflects the spatial continuity of the atmospheric field, so that the meteorological information of adjacent areas can be smoothed and interacted, and finally generates a feature representation containing the regional meteorological scenario. The time-series meteorological parameter data includes wind speed, wind direction, and wave height. The hydrodynamic modeling branch is used to simulate the advection and diffusion processes of substances within water bodies. Based on gridded hydrodynamic data and historical water quality data, this branch constructs a directed weighted graph according to the actual water flow direction. On the directed weighted graph, diffusion convolution is applied to explicitly simulate the transport of substances from upstream to downstream, thereby learning how water environment parameters are dynamically redistributed in space by water flow and obtaining corresponding feature representations. The hydrodynamic data includes flow velocity and flow direction. The terrain coding branch is used to encode relatively static geographic background features. This branch processes the inherent properties of each grid cell and transforms them into feature vectors that characterize the terrain environment through feature embedding and a fully connected network. It can also be incorporated into the local spatial context through lightweight graph convolution to form a feature representation that reflects long-term, stable geographic constraints. The inherent properties include water depth, seabed type, and shoreline distance. The resource status branch is used to encode the real-time status and availability of rescue resources. This branch receives real-time status parameters from each rescue unit and transforms discrete or continuous real-time status parameters into feature vectors of a unified dimension through temporal embedding and state coding networks. A graph attention mechanism is introduced to capture the coordination and constraint relationships between multiple resources, forming a feasible dynamic representation that reflects the current distribution and effectiveness of rescue resources, and obtaining the corresponding feature representation. Among them, the real-time status parameters include location, endurance, equipment integrity, and personnel fatigue. The fusion center, serving as the decision hub of the model, adaptively integrates the heterogeneous information from the four branches. First, the feature representations of each branch are aligned and stitched together at the node level. Then, the contribution of different physical processes is dynamically weighed through gating fusion and attention mechanisms to obtain deeply fused features. Finally, the deeply fused features are fed into a spatiotemporal prediction head, which performs final multi-round spatiotemporal evolution inference on a unified graph, and outputs the prediction results of future changes in water environment parameters.

7. The water emergency rescue path prediction and rescue coordination optimization system according to claim 6, characterized in that, In the federated learning path prediction module, the global rescue path prediction model is an agent based on a deep reinforcement learning framework. The specific process by which the global rescue path prediction model generates path prediction results is as follows: S11: Multi-source Spatiotemporal State Coding and Fusion The heterogeneous client prediction fields are calibrated and fused into a unified dynamic probability field using a spatiotemporal attention network. At the same time, the task constraints and navigation graph structure are encoded into machine-readable feature representations to obtain the corresponding spatiotemporal graph, where the heterogeneous client prediction fields are output by each local model. S12: Serialization Path Decision Generation By performing a Markov decision process on the spatiotemporal graph encoded in step S11, each decision comprehensively evaluates the immediate navigation cost, long-term goal orientation, and prediction of future environmental risks, and iteratively generates an original spatiotemporal path. S13: Multipath Optimization and Post-processing Output By sampling, perturbing, or parallel inference on the original spatiotemporal path generated in step S12, multiple candidate spatiotemporal paths are generated. After multi-objective evaluation and Pareto front screening, the optimal path is obtained. The optimal path is then dynamically smoothed and corrected. Finally, the optimized rescue trajectory scheme with a precise timestamp is output, which is the path prediction result. The whole process is iteratively optimized under the federated learning framework, which enables the model to continuously integrate global knowledge and realize intelligent path planning in dynamic environments.

8. The water emergency rescue route prediction and rescue coordination optimization system according to claim 7, characterized in that, In the multi-objective collaborative optimization module, the optimization objectives are to minimize the total rescue time, maximize the rescue success rate, minimize the rescue resource consumption, and minimize the rescue risk, as detailed below: Minimize total rescue time: ； in, The total number of units participating in the rescue operation. A unit participating in the rescue operation refers to an independent mobile vehicle or operational unit that is dispatched and deployed to perform search, rescue, or support tasks in a specific rescue mission. Let be the travel time of the i-th unit. Let be the search time for the i-th unit. Let i be the execution time of the rescue operation for the i-th unit. For path planning variables, For rescue strategy variables; Maximizing the success rate of rescue: ； in, The number of targets awaiting rescue. Let be the probability of finding the j-th target in the i-th unit. Let i be the probability that the i-th rescue unit successfully rescues the j-th target. Minimize the consumption of rescue resources: ； in, The term "rescue resources" refers to the total number of resource types. It encompasses all equipment, vehicles, personnel, and consumables that can be used for maritime emergency rescue missions. Resource types are categorized based on their core functions, power attributes, operating costs, and risk characteristics. Let be the fuel cost coefficient for resource type k. Let k be the time cost coefficient for the k-th type of resource. Let be the equipment loss coefficient for the k-th type of resource. The total sailing distance corresponding to the k-th type of resource. Let k be the total usage time of the k-th type of resource. Let be the consumption quantity of the k-th type of resource. For path planning variables, For time scheduling variables, Assign variables to resources; Minimize rescue risks: ； in, The total number of resource types, For environmental risks, For equipment risks, For personnel risks, For path planning variables, Assign variables to resources. For rescue strategy variables.

9. A waterborne emergency rescue route prediction and rescue coordination optimization system according to claim 8, characterized in that, In the multi-objective collaborative optimization module, the multi-objective particle swarm optimization algorithm introduces a dynamic weight adjustment mechanism and an elite retention strategy to ensure that the algorithm achieves an optimal balance between convergence speed and solution set diversity; the weight update formula is as follows: ； in, =0.9, =0.4, where t is the current iteration number. This represents the maximum number of iterations.

10. A method for predicting waterborne emergency rescue routes and optimizing rescue coordination, characterized in that, Applied to the system as described in any one of claims 1 to 9, Includes the following steps: S1: Each rescue entity trains a local rescue path prediction model based on local historical data. The local rescue path prediction model is the ontology model. S2: Securely aggregate model parameters from various rescue entities using a federated learning framework to generate a global rescue path prediction model; S3: Obtain path prediction results through the global rescue path prediction model, and perform multi-objective collaborative optimization calculations based on the path prediction results and environmental constraints to obtain the Pareto optimal solution set; S4: Generate the optimal rescue plan through the Pareto optimal solution set and coordinate the execution of various rescue forces; S5: Monitor the rescue process in real time and dynamically adjust the rescue strategy based on feedback data.

Images