Unmanned ship cluster formation cooperative control method applying large language model

Abstract

The invention relates to the technical field of unmanned ship control, and provides an unmanned ship cluster formation cooperative control method applying a large language model. According to the method, multi-source information such as unmanned ship operation data and marine environment data is collected, deep customization training is carried out based on a large language model, and by combining transfer learning and reinforcement learning technologies, the method has autonomous learning and intelligent decision-making capabilities. During task execution, the model receives unmanned ship states, task requirements and marine environment information in real time, intelligently judges task types, plans formation forms by using an optimization algorithm and reasonably allocates tasks. For example, in a seabed surveying and mapping task, an optimal formation scheme and an optimal surveying and mapping route are determined according to surveying and mapping area characteristics and unmanned ship performance, task progress and environment change are continuously monitored, and a cooperation strategy is rapidly adjusted. According to the method, the cooperative control capability of the unmanned ship cluster formation in the complex marine environment is effectively improved, and compared with a traditional method, the method has higher autonomous learning, intelligent decision making and dynamic cooperative capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned boat control, and in particular to a collaborative control method for an unmanned boat cluster formation using a large language model. Background Art

[0002] With the continuous development of ocean exploration and utilization, unmanned aerial vehicles (UAVs), as a vital tool for ocean exploration and operations, are finding increasingly widespread application. From marine resource exploration to marine environmental monitoring, from maritime rescue operations to military reconnaissance missions, UAVs, with their flexibility, efficiency, and adaptability to harsh environments, play an indispensable role. In practical applications, complex tasks such as monitoring large ocean areas and mapping complex seafloor topography often fall short on a single UAV. Therefore, the coordinated operation of multiple UAVs in swarm formations has become a necessity.

[0003] However, current cooperative control technology for swarms of unmanned aerial vehicles (UAVs) faces numerous bottlenecks. Traditional cooperative control methods primarily rely on pre-set rules and algorithms. In relatively simple, stable scenarios, such as regular ocean water quality monitoring in calm waters, swarms of UAVs can complete data collection tasks in an orderly manner according to established formation patterns and routes. However, these methods expose significant limitations when faced with complex and volatile ocean environments. In severe sea conditions such as strong currents and high winds and waves, the flow of seawater and the impact of wind and waves can cause the actual position and navigation status of the UAVs to deviate from the preset trajectory. Using traditional control methods, the UAVs may be unable to adjust in a timely manner, thereby impacting the coordinated operation of the entire formation and even leading to mission failure. For example, in seabed mapping missions, when faced with complex seabed topography and adverse sea conditions, UAV formations that rely on pre-set rules struggle to flexibly adjust their mapping routes and formations, making it impossible to obtain accurate and complete seabed topography data.

[0004] Existing collaborative control of unmanned boat swarm formations also has shortcomings in intelligent decision-making. Faced with complex and changing ocean environments and dynamically changing mission requirements, unmanned boats lack the ability to learn autonomously and make flexible decisions. When emergencies arise, such as the discovery of new marine biomes or special geological phenomena during marine scientific expeditions, traditional control methods are unable to quickly analyze the situation and rationally allocate unmanned boat resources, making it difficult to achieve optimal resource allocation and efficient execution of tasks. In addition, traditional methods lack effective task priority allocation and resource coordination mechanisms when dealing with multiple tasks in parallel, resulting in low efficiency of unmanned boat formations in complex tasks.

[0005] With the continuous increase in marine development activities, the requirements for the coordinated control of unmanned aerial vehicle swarms are becoming increasingly stringent. Existing control technologies cannot meet the growing needs of marine operations. There is an urgent need for innovative technologies to enhance the coordinated control capabilities of unmanned aerial vehicle swarms in complex marine environments and achieve more efficient and intelligent operations. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to propose a collaborative control method for unmanned boat cluster formation using a large language model to solve the problem of low efficiency of unmanned boat formations in complex tasks in the prior art.

[0007] Based on the above objectives, the present invention provides a method for cooperative control of an unmanned watercraft swarm formation using a large language model, comprising the following steps:

[0008] S1. Collect multi-source information, including historical data on UAV operations and marine environmental data, perform multi-source information fusion, and use transfer learning technology to integrate relevant domain knowledge into the large language model. Combined with reinforcement learning, rewards are given based on task completion accuracy, efficiency, and formation stability indicators in a virtual marine environment, and the large language model is deeply customized and trained.

[0009] S2. The large language model receives the status information, mission requirements, and ocean environment data of the UAV swarm in real time. It analyzes and determines the mission type, uses an optimization algorithm to plan the UAV formation, and constructs a linear programming model to assign specific tasks, taking into account the UAV performance, location, and mission requirements.

[0010] S3. The large language model continuously monitors the progress of the mission, the status of the unmanned boats, and changes in the ocean environment. When the actual situation deviates from expectations, it calculates the position deviation and speed deviation, re-evaluates the current operating situation, adjusts the unmanned boat formation structure and navigation route, and optimizes task allocation.

[0011] Preferably, when fusing multi-source information, the attention mechanism is used to assign different weights to different types of information to achieve a deep understanding and analysis of the status of the unmanned boat, mission requirements, and marine environment data.

[0012] Preferably, the analysis and judgment task type adopts a classification model based on the combination of convolutional neural network and recurrent neural network, and the model parameters are optimized through the cross entropy loss function.

[0013] Preferably, the method further comprises:

[0014] The reward function considers multiple key indicators of task execution, including:

[0015] Rewards are given based on the accuracy of task completion, with smaller deviations resulting in higher rewards;

[0016] In terms of mission efficiency, if the unmanned boat can complete the mission ahead of schedule, additional rewards will be given according to the proportion of the time ahead of schedule;

[0017] The formation stability is measured and rewarded by calculating the relative position deviation and speed consistency between the unmanned boats.

[0018] Preferably, the multiple key indicators for task execution also include:

[0019] Record the total energy consumption of the unmanned boat during the mission and the theoretical minimum energy consumption to complete the same mission, and give rewards based on the actual energy consumption savings.

[0020] Preferably, constructing a linear programming model to assign specific tasks includes:

[0021] A linear programming model is constructed with the shortest task completion time as the objective function and the capability limitation and task priority of the unmanned boat as the constraints to determine the specific tasks of each unmanned boat.

[0022] Preferably, step S3 further includes:

[0023] When encountering severe weather such as sudden storms, the unmanned boat's route is adjusted by calculating the distance from the storm center and the new direction of movement formula based on the storm's intensity, range, direction of movement, and the location and performance of the unmanned boat. At the same time, the direction of movement is corrected by considering the influence of ocean currents and wind.

[0024] Preferably, the method further comprises:

[0025] When a new target is discovered, the distance from each unmanned boat to the new target is calculated based on the new target location and the distribution of unmanned boats, the task is assigned to the unmanned boat closest to the target, and other unmanned boats are coordinated to adjust the task priority.

[0026] Beneficial effects of the present invention:

[0027] 1. Intelligent Decision-Making: Traditional methods rely on pre-set rules, making it difficult to quickly analyze and rationally allocate resources in the face of complex and ever-changing ocean environments and mission requirements. This patented method utilizes a large language model to analyze the status of unmanned boats, mission requirements, and ocean environment data in real time, accurately determine mission types, and intelligently plan formations and task allocations to achieve optimal resource allocation.

[0028] 2. Autonomous Learning: Traditional methods lack autonomous learning capabilities and are unable to adjust strategies based on new situations. This patent uses transfer learning and reinforcement learning to enable a large language model to deeply learn the operating data of unmanned boats and related domain knowledge, continuously optimizing control strategies in complex scenarios and improving decision-making capabilities in complex situations.

[0029] 3. Dynamic Collaboration: Traditional methods struggle to quickly adjust collaborative strategies when actual conditions deviate from expectations. This patented large language model continuously monitors mission progress and environmental changes. If deviations occur, such as inclement weather or mission changes, strategies can be quickly reassessed and adjusted to ensure smooth mission execution.

[0030] 4. Efficient Execution: Traditional methods lack effective task priority allocation and resource coordination mechanisms when running multiple tasks in parallel, resulting in low efficiency. This patent significantly improves the execution efficiency of unmanned boat swarms in complex tasks through intelligent decision-making, reasonable task priority allocation, and optimized resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0032] Figure 1 The figure is a flow chart of a collaborative control method for an unmanned boat swarm formation using a large language model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0034] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0035] Example 1:

[0036] like Figure 1 As shown, the embodiment of this specification provides a method for cooperative control of an unmanned boat swarm formation using a large language model, the main steps of which are as follows:

[0037] Deep customization and training of large language models: Collect massive amounts of historical data on the operation of unmanned boats, covering the motion trajectories, state parameters, and corresponding collaborative control strategies and execution results of unmanned boats in different ocean environments and mission types. Use these data to deeply pre-train the large language model, introduce transfer learning technology in the process, and integrate knowledge from related fields such as ocean dynamics and navigation into the model to accelerate its learning of unmanned boat cluster formation control tasks. At the same time, combined with reinforcement learning, by constructing a virtual ocean environment, the model is constantly interacting with the environment in simulated complex scenarios, and rewards are given according to the set reward mechanism, such as accuracy, efficiency, formation stability and other indicators of task completion, to encourage the model to optimize the control strategy and improve decision-making capabilities in various complex situations. Specifically including:

[0038] 1) Multi-source data collection and preprocessing:

[0039] Data Collection: Leveraging the high-precision sensors onboard the unmanned boats, a comprehensive data collection system is constructed. For example, an inertial measurement unit (IMU) is used to obtain the unmanned boat's acceleration and angular velocity information to accurately track its motion. High-resolution sonar equipment is used to collect seabed terrain echo data, providing a foundation for subsequent seabed mapping missions. Meteorological sensors are used to monitor marine meteorological parameters such as wind speed, wind direction, air pressure, and sea temperature in real time. At the same time, historical environmental data from different sea areas during past missions is collected through satellite communication links, including seasonal ocean current distribution and tidal patterns. Furthermore, control commands issued by the unmanned boat during various missions, such as steering angles and propulsion power adjustments, are recorded, along with the actual effects of these commands.

[0040] Data cleaning and labeling: Collected data often contains noise and outliers and requires cleaning. Statistical methods, such as the 3σ criterion, are used to identify and remove data points that significantly deviate from the normal range. Missing data is filled in using methods such as linear interpolation and the K-nearest neighbor (KNN) algorithm. After cleaning, the data is labeled, adding labels for different task types, environmental conditions, and control strategies so that the model can accurately learn the corresponding relationships in different scenarios.

[0041] 2) Transfer learning knowledge fusion:

[0042] Domain knowledge extraction: From the field of ocean dynamics, we extract the characteristics of the solutions to the Navier-Stokes equations that describe the movement of seawater, as well as the parameters of the mathematical models of wave propagation under different sea conditions. For example, through numerical simulation, we obtain characteristic data such as wave height, period, and direction under different wind speeds. In the field of navigation, we review the International Regulations for Preventing Collisions at Sea (COLREGs) and convert the conditional judgments and response strategies in the regulations into quantifiable feature vectors, such as safe distances and avoidance angles under different ship encounter situations.

[0043] Knowledge Fusion: Utilizing an attention-based fusion method, the model automatically focuses on the relevance of different domain knowledge to the UAV control task during training. Specifically, at the model's input layer, UAV operational data is combined with extracted domain knowledge to form new input features. At the model's hidden layer, an attention weight matrix dynamically adjusts the influence of different knowledge sources on the model output. This allows the model to fully leverage ocean dynamics and navigation knowledge when learning the UAV swarm formation control task, accelerating convergence and improving generalization.

[0044] 3) Reinforcement learning strategy optimization:

[0045] Virtual Environment Construction: Utilizing game development engines such as Unity or Unreal Engine, a highly realistic virtual ocean environment is constructed. Within this environment, various ocean phenomena are accurately simulated, including ocean currents of varying strength and direction, including the characteristics of typical currents such as the Equatorial Warm Current and the Westerly Drift Current. A variety of weather conditions, from gentle breezes to destructive hurricanes, are simulated, along with changes in visibility under varying weather conditions. Furthermore, various complex scenarios are implemented, such as navigation through narrow straits and obstacle avoidance when approaching islands.

[0046] Reward function design: Design a comprehensive reward function that comprehensively considers multiple key indicators of task execution. For the accuracy of task completion, taking the seabed mapping task as an example, rewards are given based on the degree of deviation between the mapping data and the actual seabed topography. The smaller the deviation, the higher the reward. In terms of task efficiency, if the unmanned boat can complete the task ahead of schedule within the specified time, additional rewards will be given according to the proportion of the time in advance. For formation stability, it is measured by calculating the relative position deviation and speed consistency between the unmanned boats. When the formation can remain stable in complex sea conditions, positive rewards are given; if there is a risk of collision or the formation is seriously confused, negative rewards are given. In addition, energy consumption factors are also taken into consideration, and strategies with high energy efficiency are rewarded to encourage the model to learn energy-saving control methods.

[0047] Training process optimization: The Proximal Policy Optimization (PPO) algorithm is used for reinforcement learning training. During training, the exploration-exploitation balance parameter is dynamically adjusted. As training progresses, the exploration rate is gradually reduced, allowing the model to focus more on utilizing the learned optimal policy. Furthermore, the model's training status is regularly saved to enable rapid recovery in the event of training interruptions or anomalies. By continuously training and optimizing policies in a virtual environment, the large language model can learn the optimal cooperative control strategy for unmanned vehicle swarm formations in various complex marine environments.

[0048] The design of the reinforcement learning reward function is the core of the above process. The design process of the present invention is as follows:

[0049] 1) Rewards for task completion accuracy:

[0050] Taking the seabed mapping task as an example, by comparing the seabed topography data obtained by the actual mapping of the unmanned boat with the known high-precision reference data (such as historical mapping data or authoritative topography data of the area), the root mean square error (RMSE) is used to measure the degree of deviation between the two. Assume that the reference data is y true , the actual surveying data is y pred , the sample size is n, then the calculation formula of RMSE is: Reward value R accuracy It is inversely proportional to RMSE, that is In this way, the closer the survey data is to the reference data, the smaller the RMSE is, and the higher the reward value is.

[0051] This award aims to encourage large language models to learn control strategies that can accurately complete tasks, ensuring that the data obtained by unmanned vehicles when performing tasks such as seabed mapping is highly accurate, and providing a reliable data foundation for subsequent marine research and development.

[0052] 2) Task efficiency rewards:

[0053] According to the preset time T of the task 预设 and the time T when the unmanned boat actually completes the mission 实际 To calculate the reward. First determine an efficiency reward coefficient k (for example, k = 10, which can be adjusted according to actual conditions). Reward value R efficiency The calculation formula is: When the unmanned boat completes its mission ahead of schedule: T 实际 <T 预设 , the reward value is positive, and the more time in advance, the higher the reward; if the task is completed overtime, T 实际 >T 预设，奖励值为 Negative to penalize inefficient strategies.

[0054] This incentive model optimizes the unmanned boat's route of action and task allocation plan, enabling the unmanned boat cluster formation to complete the task in the shortest time, improve resource utilization efficiency, and reduce operating costs. It is especially suitable for time-sensitive marine operations.

[0055] 3) Formation Stability Reward:

[0056] The formation stability is evaluated by monitoring the relative position deviation and speed consistency between the UAVs. For the relative position deviation, the sum of the Euclidean distances D between each UAV and its theoretical position in the formation is calculated. Assuming that there are m UAVs in the formation, the actual position coordinates of the jth UAV are (x j ,yj ), the theoretical position coordinates are but For speed consistency, calculate the standard deviation σ of each unmanned boat speed and the average speed of the formation v .

[0057] Formation stability bonus R stability With D and σ v Inversely proportional, for example, it can be expressed as

[0058]

[0059] This reward function ensures that the unmanned boat cluster formation can maintain a stable formation in a complex ocean environment, avoids the risk of collision and reduced mission execution efficiency caused by formation chaos, and improves the collaborative operation capability between unmanned boats.

[0060] 4) Energy consumption rewards:

[0061] Record the total energy consumption E of the unmanned boat during the mission 实际 , and the theoretical minimum energy consumption E to complete the same task 理论 ,

[0062] (can be estimated through simulation or empirical data). Set an energy incentive coefficient c (for example,

[0063] c=5), reward value R e The calculation formula for energy is: When the actual energy consumption is lower than the theoretical value, the reward value is positive, and the more energy saved, the higher the reward; if the actual energy consumption is too high, the reward value is negative to encourage the model to learn energy-saving control strategies.

[0064] Considering that unmanned boats usually rely on a limited energy supply to operate, this reward can guide the large language model to optimize the control strategy, reduce energy consumption, and extend the endurance of the unmanned boat, thereby expanding its operating range and time.

[0065] By comprehensively considering the reward factors in the above aspects, a comprehensive and scientific reinforcement learning reward function is constructed, which enables the large language model to continuously optimize the collaborative control strategy of the unmanned boat cluster formation during the learning process to achieve efficient, accurate, stable and energy-saving operation goals.

[0066] 2. Model-based Intelligent Formation Planning and Task Allocation: A large language model receives real-time status information from the UAV swarm, as well as mission requirements and ocean environment data. Through in-depth understanding and analysis of this information, it first accurately determines the mission type, such as seafloor mapping, marine scientific research, or maritime patrol. Then, based on the mission characteristics and the actual conditions of the UAVs, such as their performance and location, a scientific and reasonable collaborative control strategy is formulated. This includes planning UAV formations, such as linear, diamond, and circular formations, to suit different mission requirements and ocean environments. Furthermore, each UAV is assigned a specific task, determining its route, operating time, and coordination with other UAVs. For example, in a seafloor mapping mission, the optimal formation and route for each UAV are planned based on the shape, size, and complexity of the survey area, as well as the UAVs' accuracy and speed, ensuring comprehensive and efficient completion of the mission.

[0067] 1) Multi-source information fusion and analysis: The large language model receives the status information of the unmanned boat cluster formation in real time through a dedicated communication protocol. The location information is obtained by the high-precision global satellite navigation system (GNSS) and can be expressed as geographic coordinates (x i ,y i ), i represents the vth i Unmanned boats; the speed information is measured by Doppler velocimeter, denoted as θ i ; The heading information is determined by the electronic compass, using T = [t1, t2, ..., t n ] represents the task requirement information, including task objectives, task priorities, etc., in the form of vector t j Indicates that j represents the vth c Current mission attributes. Ocean environment data, such as ocean current velocity θ measured by Acoustic Doppler Current Profiler (ADCP) current and direction v wind , and the wind speed θ collected by the meteorological sensor wind , wind direction, etc. The model utilizes an attention mechanism to fuse this multi-source information, assigning different weights to different types of information to achieve in-depth understanding and analysis. For example, for seabed mapping tasks, location information and the extent of the mapping area are given higher weight; while in adverse sea conditions, ocean currents and weather information are given higher weight.

[0068] 2) Task type identification and analysis: The model uses a classification model based on a combination of convolutional neural networks (CNN) and recurrent neural networks (RNN) to determine the task type. The fused information is converted into a feature matrix and input into the classification model. CNN extracts local features, and RNN captures sequence features. Taking the seabed mapping task as an example, when the input features include features such as a large continuous seabed topography mapping range and high-precision mapping requirements, the model outputs the task type as seabed mapping. Assuming that the output of the classification model is y, y∈{seabed mapping, marine scientific research, maritime patrol, ...}, through the cross entropy loss function To optimize the model parameters, where K is the total number of task types, y k is the true label, is the predicted probability.

[0069] 3) Formation planning: Based on the mission type and ocean environment conditions, the model uses optimization to plan the formation of unmanned boats. In the seabed mapping mission, if the mapping area is long and narrow, a linear formation is used. The Hungarian algorithm is used to plan the order of the unmanned boats so that the mapping ranges of adjacent unmanned boats have minimal overlap and can cover the entire area. Assuming that there are na and angle α to maximize the mapping efficiency, the efficiency function can be expressed as By continuously iteratively optimizing a and α, E is maximized.

[0070] 4) Task allocation decision: Comprehensively consider the performance of the unmanned boat (endurance E range , speed v, equipment accuracy a, etc.), current position and task requirements, a linear programming model is constructed to allocate tasks. Taking the shortest task completion time as the objective function, set policy x ij is the decision variable. When the i-th unmanned boat is assigned to the j-th task, x ij =1, otherwise x ij = 0. The objective function is where t ij For the The time required for an unmanned boat to complete the jth task. The constraints include that each unmanned boat can undertake at most one task, and each task can be performed by only one unmanned boat. And the performance constraints of the unmanned boat, such as endurance constraints is the distance from the i-th unmanned boat to the j-th mission location. By solving the linear programming model, the specific mission of each unmanned boat is determined.

[0071] 3. Dynamic Collaboration Strategy Adjustment and Optimization: During the UAV swarm's mission, the large language model continuously monitors mission progress, UAV status, and changes in the ocean environment. If actual conditions deviate from expectations, such as unexpected severe weather, new obstacles, or changes in mission requirements, the model immediately reassesses the current situation and rapidly adjusts the collaborative control strategy.

[0072] 1) Real-time monitoring and deviation detection: The large language model continuously collects the position, speed, heading and other status information of the unmanned boat, as well as the task execution progress data, through the real-time communication link between the unmanned boat and the unmanned boat. Assume that the expected position of the unmanned boat at time t is The actual location is Using the Euclidean distance formula To calculate the position deviation. For the speed deviation, let the expected speed be v expected The actual speed is v actual , then the speed deviation At the same time, monitor changes in ocean environmental parameters, such as obtaining ocean current speed in real time through sensors and Xiangfanghai wind speed and wind direction Once the position deviation d exceeds the preset threshold D, or the speed deviation Δv exceeds the preset threshold V, or the ocean environment parameter changes exceed a certain range, the model determines that the actual situation deviates from the expectation and triggers the strategy adjustment mechanism.

[0073] 2) Strategy adjustment decision: When a deviation is detected, the model quickly analyzes the current situation. Taking a sudden storm as an example, the model first analyzes the storm's intensity I (which can be calculated by comprehensively calculating parameters such as wind speed and air pressure, such as I = αv) based on meteorological sensor data and satellite cloud images. wind +βΔPα, β is the weight coefficient, ΔP is the pressure difference between the storm center and the area where the unmanned boat is located), range R (the area covered by the storm is estimated by satellite cloud images) and moving direction θ storm Then, based on the current position of the unmanned boat (x i ,y i ) and storm related information to calculate the distance between the unmanned boat and the storm center Where (x storm ,y storm ) is the location of the storm center. If d i-storm ≤R, the unmanned boat is judged to be within the storm’s influence range. The model is based on the unmanned boat’s endurance E range , speed v and the position of the safe area (x ssfe ,y safe ), formulate a new course of action. New direction of action θ new is determined by calculating the angle with the storm's moving direction and the safe distance, and the formula is θ new =arctan2(y safe -y i ,x safe -x i ), while considering the influence of ocean current and wind, corrected After correction, the corrected direction can be expressed as θ corrected =θ new +γθ current +δθ wind , where γ and δ are coefficients related to the degree of influence of ocean current and wind.

[0074] 3) Formation and task adjustment: While adjusting the route of the unmanned boats, the model re-plans the formation according to the task requirements and the distribution of the unmanned boats. If the original linear formation is used to perform seabed mapping tasks, it may be adjusted to a more flexible dispersed formation when encountering a storm to reduce the risk of collision. For task allocation, if a new special geological structure or marine biological community is discovered, the model will re-plan the formation according to the location of the new target (x new ,y new ) and the distribution set of unmanned boats Reassign tasks. Calculate the distance from each unmanned boat to the new target The nearest unmanned boat is selected to perform the detection task of the new target first. At the same time, coordinate other unmanned boats to continue to complete the original tasks or adjust the task priority to ensure the efficient advancement of the overall task. For example, for marine scientific research missions, if a new biological community is discovered, the priority of the biological observation task will be increased, and some unmanned boats responsible for collecting seawater samples will be temporarily adjusted to conduct biological observations. After obtaining sufficient information, the task allocation will be re-planned. For example, when the unmanned boat formation originally performing marine scientific research missions encounters a sudden storm, the model quickly analyzes the intensity, range and movement direction of the storm, adjusts the formation form and route of the unmanned boats, and guides the unmanned boats to a safe area to avoid the storm. After the storm passes, the tasks and formations are re-planned to continue completing the scientific research mission.

[0075] Example 2:

[0076] Taking marine scientific research as an example, the specific implementation process of this method is explained:

[0077] 1. Model Training Preparation: We collected extensive historical marine scientific research data, covering marine ecology and meteorological conditions in various sea areas, as well as collaborative control strategies for unmanned vehicles used in various scientific research missions. We pre-trained a large language model, incorporating knowledge from marine biology and meteorology. Through reinforcement learning, we established a reward mechanism, such as positive rewards for successfully obtaining rare marine samples and accurately recording special meteorological phenomena, and negative rewards for failures, to encourage the model to learn optimal strategies.

[0078] 2. Scientific expedition mission execution: When a new marine scientific expedition mission is launched, the model receives mission information and the status of the unmanned boat. The scientific expedition mission is broken down into multiple subtasks, such as marine biological observation, seawater sample collection, and meteorological monitoring. The corresponding subtasks are assigned according to the equipment and performance of the unmanned boat. For example, an unmanned boat equipped with high-definition cameras and biometric equipment is responsible for the marine biological observation area, an unmanned boat carrying professional sampling equipment goes to the designated sea area to collect samples, and an unmanned boat equipped with meteorological monitoring instruments collects meteorological data. The route of action and formation of each unmanned boat are planned so that they can complete their respective tasks efficiently while maintaining the coordination of the formation. For example, in the biological observation area, a decentralized formation is adopted to facilitate comprehensive observation by the unmanned boat; in the sampling area, a compact formation is adopted to facilitate centralized operations and data sharing.

[0079] 3. Dynamic adjustment of strategies: During the marine scientific expedition, if severe weather (such as heavy rain, strong winds, etc.) suddenly affects the navigation safety and mission execution of the unmanned boat, the model will immediately evaluate the environmental changes and dynamically adjust the navigation strategy of the unmanned boat. For areas with higher risks, the model can command some unmanned boats to evacuate to safe waters, and after the weather conditions improve, re-optimize the mission planning and formation strategy to ensure the smooth progress of the scientific expedition mission. If new marine biological communities or special meteorological phenomena are discovered, the model reallocates unmanned boat resources and focuses on research on the new discoveries. For example, when a new marine biological community is discovered, the tasks and routes of some unmanned boats are adjusted to allow them to approach the biological community for detailed observation and sampling, while coordinating other unmanned boats to continue to complete their original tasks to ensure the smooth progress of the entire scientific expedition mission and maximize scientific value.

[0080] Those skilled in the art will understand that the discussion of any of the above embodiments is merely illustrative and is not intended to limit the scope of the present invention to these examples. Within the spirit and principles of the present invention, the technical features of the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and many other variations exist for the various aspects of the present invention described above, which are not provided in detail for the sake of clarity. Any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A collaborative control method for an unmanned watercraft swarm formation using a large language model, characterized in that: The following steps are involved: S1. Collect multi-source information, including historical data on UAV operations and marine environmental data, perform multi-source information fusion, and use transfer learning technology to integrate relevant domain knowledge into the large language model. Combined with reinforcement learning, rewards are given based on task completion accuracy, efficiency, and formation stability indicators in a virtual marine environment, and the large language model is deeply customized and trained. S2. The large language model receives the status information, mission requirements, and ocean environment data of the UAV swarm in real time. It analyzes and determines the mission type, uses an optimization algorithm to plan the UAV formation, and constructs a linear programming model to assign specific tasks, taking into account the UAV performance, location, and mission requirements. S3. The large language model continuously monitors the progress of the mission, the status of the unmanned boats, and changes in the ocean environment. When the actual situation deviates from expectations, it calculates the position deviation and speed deviation, re-evaluates the current operating situation, adjusts the unmanned boat formation structure and navigation route, and optimizes task allocation.

2. The method for cooperative control of unmanned watercraft swarm formations using a large language model according to claim 1, characterized in that: When fusing multi-source information, the attention mechanism is used to assign different weights to different types of information, achieving a deep understanding and analysis of the UAV status, mission requirements, and marine environment data.

3. The cooperative control method of unmanned watercraft swarm formation using a large language model according to claim 1 is characterized in that: The analysis and judgment task type adopts a classification model based on a combination of convolutional neural networks and recurrent neural networks, and optimizes model parameters through a cross-entropy loss function.

4. The method for cooperative control of a swarm of unmanned boats using a large language model according to claim 1, characterized in that: The method further comprises: The reward function considers multiple key indicators of task execution, including: Rewards are given based on the accuracy of task completion, with smaller deviations resulting in higher rewards; In terms of mission efficiency, if the unmanned boat can complete the mission ahead of schedule, additional rewards will be given according to the proportion of the time ahead of schedule; The formation stability is measured and rewarded by calculating the relative position deviation and speed consistency between the unmanned boats.

5. The method for cooperative control of a swarm of unmanned boats using a large language model according to claim 4 is characterized in that: The key indicators for evaluating task execution also include: Record the total energy consumption of the unmanned boat during the mission and the theoretical minimum energy consumption to complete the same mission, and give rewards based on the actual energy consumption savings.

6. The method for cooperative control of a swarm of unmanned boats using a large language model according to claim 1, characterized in that: The specific tasks of constructing the linear programming model include: A linear programming model is constructed with the shortest task completion time as the objective function and the capability limitation and task priority of the unmanned boat as the constraints to determine the specific tasks of each unmanned boat.

7. The method for cooperative control of a swarm of unmanned boats using a large language model according to claim 1, characterized in that: Step S3 further comprises: When encountering severe weather such as sudden storms, the unmanned boat's route is adjusted by calculating the distance from the storm center and the new direction of movement formula based on the storm's intensity, range, direction of movement, and the location and performance of the unmanned boat. At the same time, the direction of movement is corrected by considering the influence of ocean currents and wind.

8. The method for cooperative control of a swarm of unmanned boats using a large language model according to claim 1, characterized in that: The method further comprises: When a new target is discovered, the distance from each unmanned boat to the new target is calculated based on the new target location and the distribution of unmanned boats, the task is assigned to the unmanned boat closest to the target, and other unmanned boats are coordinated to adjust the task priority.

Images