A double-layer mission planning type unmanned ship control and mission planning method

Abstract

The present application relates to the technical field of ocean engineering, and is a kind of double-layer task planning type unmanned ship control and task planning method, comprising; based on the hardware formation of unmanned ship of multiple unmanned ships, each unmanned ship is equipped with sensor unit and is configured with propulsion execution mechanism suitable for open sea operation;The control and task planning model is pre-trained in the Unity simulation environment, and the pre-trained control and task planning model is obtained, the control and task planning model includes task allocation model, path planning model, path tracking model and PID control submodule;The pre-trained control and task planning model is to unmanned ship, and the control and task planning model is fine-tuned and optimized, and the control and task planning of unmanned ship are controlled and planned through the fine-tuned and optimized control and task planning model.The present application can solve the technical problems of task allocation imbalance, path obstacle avoidance and energy consumption in the operation of ocean ranching unmanned ship, and improve the task response efficiency of unmanned ship and reduce energy consumption.

Description

Technical Field

[0001] This invention relates to the field of marine engineering technology, and specifically to a two-layer mission planning-based unmanned surface vessel control and mission planning method. Background Technology

[0002] As an important component of my country's blue economy, marine ranching is rapidly developing towards large-scale and intelligent operations. Currently, the core operating area of ​​a single marine ranch in my country ranges from 300 to 1000 mu (approximately 20-67 hectares), with over 200 aquaculture cages. Simultaneously, diverse tasks such as cage inspection, water quality monitoring, and obstacle avoidance must be completed. Due to factors such as dynamic ocean currents, dense cage distribution, and sensor noise in salt spray environments, traditional manual inspection and single unmanned surface vessel (USV) operation modes can no longer meet the operational requirements of "high-efficiency collaboration, precise control, and low-consumption safety." Multi-USV swarm collaboration technology has become crucial for industry upgrading.

[0003] Among existing related technologies, patent CN120523200B (A Dynamic Target Tracking Control Method for Distributed Cooperative Maneuvering of Multiple Unmanned Vessels) has made improvements in the field of multi-unmanned vessel cooperative control, but it has significant limitations in the marine ranching scenario: This patent focuses on the single task of "dynamic target tracking" and uses a "virtual leader-follower" architecture to achieve formation changes, without considering the task differences between "cage inspection (high priority)" and "water quality monitoring (normal priority)" in marine ranches, and cannot solve the problem of multi-task overall allocation; the path planning stage only uses affine transformation to achieve formation rotation and scaling, without setting specific obstacle avoidance constraints for dense cages in ranches, resulting in high collision risk in scenarios with low cage spacing; although its control algorithm can compensate for external disturbances, it does not incorporate a dynamic ocean current energy consumption model for marine ranches; in addition, this patent lacks a virtual-real fusion training mechanism, making it difficult to adapt to the complex environment of ranches and unable to meet the needs of efficient collaborative operations in large-scale ranches. Therefore, a two-layer task planning method that can integrate task priority, ranch environment characteristics, and virtual-real training optimization is needed. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a two-layer mission planning-based unmanned surface vessel (USV) control and mission planning method, which can solve the technical problems of unbalanced mission allocation, difficulty in balancing path avoidance and energy consumption in USV operations in marine ranching, and improve the overall mission response efficiency of USVs and reduce energy consumption.

[0005] The purpose of this invention is to provide a two-layer mission planning-based unmanned surface vessel control and mission planning method.

[0006] The objective of this invention can be achieved by adopting the following technical solutions:

[0007] A two-layer mission planning-based unmanned surface vessel (USV) control and mission planning method includes the following steps:

[0008] S1. Build a hardware execution layer. Based on multiple unmanned surface vessels (USVs), form a USV hardware formation. Each USV is equipped with a sensor unit and a propulsion execution mechanism adapted to offshore operations. The sensor unit is used to collect USV status data and marine ranch environmental data.

[0009] S2. Construct a control and task planning model. Pre-train the control and task planning model in the Unity simulation environment to obtain a pre-trained control and task planning model. The control and task planning model includes a task allocation model, a path planning model, a path tracking model, and a PID control sub-module.

[0010] The task allocation model is used to output the task execution order of each unmanned surface vessel (USV) based on its initial position, battery level, a series of task points, and task priorities.

[0011] The path planning model uses a particle swarm optimization-A fusion algorithm to avoid obstacles in aquaculture cages and reduce ocean current energy consumption. It generates the globally optimal path for each unmanned surface vessel (USV) based on the mission execution order.

[0012] The path tracking model is used to output a coordinated reference heading angle and velocity in real time based on the global optimal path of each unmanned surface vessel and the local observation information received from each unmanned surface vessel.

[0013] The PID control submodule is used to generate a PWM drive signal for controlling the propulsion actuator based on the reference heading angle and speed output by the path tracking module, after PID closed-loop adjustment to compensate for actual ocean current interference.

[0014] S3. Transfer the pre-trained control and mission planning model to the unmanned surface vessel (USV), fine-tune and optimize the control and mission planning model, and use the fine-tuned and optimized control and mission planning model to control and plan the USV based on the USV's state data and marine ranch environment data.

[0015] Specifically, the construction of the control and task planning model, and the pre-training of the control and task planning model in the Unity simulation environment to obtain the pre-trained control and task planning model, includes:

[0016] Based on the PPO reinforcement learning algorithm, a task allocation model is constructed by combining marine ranching task priority, unmanned surface vessel remaining battery power, and ranching no-navigation zone constraints. The task allocation model is pre-trained in the Unity simulation environment, and the task execution order of each unmanned surface vessel is output through the task allocation model.

[0017] Based on particle swarm optimization and A* algorithm, a path planning model is constructed with the goal of avoiding obstacles in aquaculture cages and reducing ocean current energy consumption. According to the task execution order of the unmanned surface vessel, the path planning model generates the global optimal path for each unmanned surface vessel.

[0018] A path tracking model is constructed based on the MAPPO algorithm and trained in the Unity simulation environment. The path tracking model is used to output the coordinated reference heading angle and velocity in real time based on the global optimal path of each unmanned surface vessel and the local observation information received from each unmanned surface vessel.

[0019] Specifically, the sensor unit includes: a positioning sensor, an attitude sensor, and an obstacle detection sensor; the unmanned surface vessel's status data includes its position and attitude data; and the marine ranch environment data includes: the location of the net cages, ocean current speed, and obstacle distribution. The propulsion actuator includes: a thruster and a servo motor.

[0020] Specifically, the task allocation model constructed based on the PPO reinforcement learning algorithm, combined with marine ranching task priorities, unmanned surface vessel remaining battery power, and ranching no-navigation zone constraints, includes:

[0021] Based on the PPO reinforcement learning algorithm, the objective function of the task allocation model is set as a multi-objective reward function. This multi-objective function includes: task priority reward, energy consumption optimization reward, power balance reward, and constraint penalty reward. The multi-objective reward function R... total Represented as:

[0022] R total = R1 + R2 + R3 + R4;

[0023] Among them, R1 is the task priority reward, R2 is the power balance reward, R3 is the energy consumption reward, which is set according to the energy consumption per unit path of the unmanned surface vessel, and a positive reward is given when the energy consumption is lower than the preset threshold; R4 is the constraint penalty reward.

[0024] The task allocation model is pre-trained in the Unity simulation environment. The pre-trained task allocation model is used to output the task execution order of each unmanned surface vessel based on its initial position, battery level, a series of task points and task priorities.

[0025] Specifically, based on particle swarm optimization and A* algorithm, a path planning model is constructed with the goal of avoiding obstacles in aquaculture cages and reducing ocean current energy consumption. According to the task execution order of the unmanned surface vessels (USVs), the path planning model generates the globally optimal path for each USV, including:

[0026] First, the initial path node set is selected using the particle swarm optimization algorithm. The fitness function is set and the constraint is processed to ensure that the distance between the path and the edge of the net cage meets the safety obstacle avoidance requirements of the marine ranch. Then, the particle is encoded and the particle swarm is initialized to calculate the particle fitness and remove nodes that may touch the no-navigation zone. The initial path node set that meets the obstacle avoidance requirements is then selected.

[0027] Using the initial path node set as the starting / ending point, a heuristic function for fusing target distance and ocean current resistance is designed based on the A* fusion algorithm. By weighting the path length and energy consumption cost, a smooth, collision-free path is generated.

[0028] Specifically, the step of constructing a path tracking model based on the MAPPO algorithm and training the path tracking model in the Unity simulation environment includes:

[0029] Based on the MAPPO algorithm, the global state is defined as including the position, speed, heading angle, reference path point sequence, ocean current vector of key points, and formation state of the vessels. The local observation includes the sensor information of each vessel and limited neighborhood information. The action space includes the output of each agent as continuous control commands: heading angle and desired speed.

[0030] The design incorporates a centralized Critic network and a distributed Actor network. During the training phase, for multi-agent systems requiring strict spatiotemporal coordination, a shared centralized Critic network is used, which takes the global state as input and outputs a scalar value. Each agent, on the other hand, has an independent Actor network, whose input is its own local observations and output is an action probability distribution.

[0031] Set a reward function that includes tracking reward, collision avoidance reward, smoothness reward, and formation reward.

[0032] Based on the Unity simulation environment, state data and environmental perception data of all unmanned surface vessels were collected to construct training scenarios with various ocean current interference modes. The MAPPO algorithm was trained in a centralized manner, and a formation tracking strategy was learned through multi-agent collaborative decision-making.

[0033] Specifically, the step of generating a PWM drive signal for controlling the propulsion actuator's operation after compensating for actual ocean current interference by PID closed-loop adjustment based on the desired heading angle and desired velocity output by the path tracking module includes:

[0034] After receiving the desired heading angle and desired velocity output by MAPPO, the actual motion state and ocean current disturbance characteristics of the unmanned surface vessel are analyzed by combining multi-sensor data with the state observer.

[0035] The PID parameters are dynamically adjusted according to the intensity of the disturbance. When a strong ocean current causes the deviation to exceed the limit, the proportional coefficient is increased to speed up the response and the integral coefficient is reduced to prevent overshoot. When the state is stable, the parameters are optimized to achieve zero steady-state error tracking, and an advance compensation amount is generated and superimposed with the basic PID control amount.

[0036] The PWM drive signal is generated by coordinating the frequency and duty cycle. The thruster adapts the frequency according to the speed and load, and the servo uses a fixed high frequency in conjunction with a fine duty cycle to finely adjust the heading.

[0037] Specifically, the step of transferring the pre-trained control and mission planning model to the unmanned surface vessel and fine-tuning and optimizing the control and mission planning model includes:

[0038] The control and mission planning model is loaded into the actual computing unit of the unmanned surface vessel (USV), and the parameter of the feature extraction layer of the control and mission planning model is frozen. Based on the sensor unit of the actual vessel, the state data of the USV and the marine ranch environment data are collected to construct a real environment training dataset. The control and mission planning model is trained on the real environment training dataset, and the fully connected layer parameters of the control and mission planning model are iteratively optimized. By pre-setting the number of iterations and the learning rate, the control and mission planning model is adapted to the environmental differences of the real marine ranch, resulting in a finely tuned and optimized control and mission planning model.

[0039] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0040] This embodiment provides a two-layer task planning-based unmanned surface vessel (USV) control and task planning method. By constructing a control and task planning model, and pre-training the model in a Unity simulation environment, a pre-trained control and task planning model is obtained. The control and task planning model includes a task allocation model, a path planning model, a path tracking model, and a PID control submodule. Sensor units ensure data acquisition accuracy, the task allocation model realizes intelligent coordination of multiple objectives, the path planning model and the path tracking model take into account both path safety and formation coordination, the PID control submodule ensures execution accuracy under ocean current interference, and the virtual-real fusion supported by the ROS system enables seamless model transfer. This method can solve the technical problems of unbalanced task allocation, difficulty in balancing path avoidance and energy consumption in USV operations in marine ranching, and improve overall task response efficiency and reduce energy consumption. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0042] Figure 1 This is a flowchart of a two-layer mission planning unmanned surface vessel control and mission planning method in an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the upper-layer task allocation model training process in an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the centralized training of the MAPPO algorithm for the path tracking model in this embodiment of the invention;

[0045] Figure 4 This is a schematic diagram of the model parameter fine-tuning optimization strategy in an embodiment of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments, and the implementation of the present invention is not limited thereto. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1:

[0048] like Figure 1 The diagram shows a flowchart of a two-layer mission planning unmanned surface vessel (USV) control and mission planning method. This invention provides a two-layer mission planning USV control and mission planning method applicable to marine ranching, comprising the following steps:

[0049] S1. Build a hardware execution layer. Based on multiple unmanned surface vessels (USVs), form a USV hardware formation. Each USV is equipped with a sensor unit and a propulsion execution mechanism adapted to offshore operations. The sensor unit is used to collect USV status data and marine ranch environmental data.

[0050] Specifically, the system comprises a coordinated formation of multiple unmanned surface vessels (USVs). Each USV is equipped with sensor units including a positioning sensor, an attitude sensor, an obstacle detection sensor, and a water quality and flow sensor. These sensors collect real-time status data of the USVs, including their positioning and attitude data, as well as environmental data related to the marine ranching environment, such as the location of fish cages, ocean current speed, and obstacle distribution. The system is equipped with low-energy propulsion actuators, including thrusters and servo motors, to meet the requirements of long-duration operations in the open ocean. Generally, the positioning sensor is a GPS sensor, used to detect the USVs' positioning data, with positioning accuracy suitable for centimeter-level obstacle avoidance requirements in marine ranching. The attitude sensor is an IMU sensor, used to detect the USVs' attitude data, with attitude accuracy meeting the micro-amplitude attitude control requirements of coordinated USV formation operations. The obstacle detection sensor is a radar sensor, used to detect the location of fish cages and the distribution of obstacles, with a detection range covering the early warning range for safe obstacle avoidance by the USVs. The water quality and flow sensor can be used to detect ocean current speed and water quality parameters. The propulsion actuator is a low-energy brushless thruster, with rated power and maximum thrust suitable for the power and endurance requirements of long-duration operations in the open ocean. In this embodiment, six unmanned surface vessels are selected to form a formation. Each vessel is equipped with a GPS sensor, an IMU sensor, a radar sensor, and a water quality and flow sensor. It is also equipped with a low-energy propulsion system and a servo motor. The sensor network is built through CAN bus integration.

[0051] S2. Construct a control and task planning model, which includes a task allocation model, a path planning model, a path tracking model, and a PID control submodule. Pre-train the control and task planning model in the Unity simulation environment to obtain a pre-trained control and task planning model.

[0052] S21. Based on the PPO (Proximal Policy Optimization) reinforcement learning algorithm, a task allocation model is constructed by combining the marine ranch task priority, the remaining battery power of the unmanned surface vessel, and the ranch no-navigation zone constraints. The task allocation model is pre-trained in the Unity simulation environment. The task allocation model outputs the task execution order of each unmanned surface vessel, realizing intelligent coordination of multiple tasks.

[0053] Specifically, based on the PPO reinforcement learning algorithm, the reward function is designed as a multi-objective collaborative mode, with multiple objectives including task priority, power balance, and energy consumption optimization. The priority difference between cage inspection and water quality monitoring tasks is differentiated, with cage inspection having higher priority than regular water quality monitoring. Completing high-priority tasks yields higher base rewards, with extra points awarded for early completion and penalties for failure to complete within the time limit. Positive and negative rewards are set based on the balance of remaining power in the unmanned surface vessel fleet, with higher rewards for more balanced power distribution. Rewards are also given based on the energy consumption level per unit path, with positive incentives given when energy consumption is below a preset threshold. Ranch no-navigation zone constraints are implemented, with a penalty mechanism for entering the no-navigation zone around the cages, strengthening obstacle avoidance constraints.

[0054] S211. Based on the PPO reinforcement learning algorithm, the objective function of the task allocation model is set as a multi-objective reward function. The multi-objective function includes: task priority reward, energy consumption optimization reward, power balance reward, and constraint penalty reward. The task allocation reward function is key to guiding the PPO reinforcement learning algorithm to achieve multiple objectives; it consists of multiple reward sub-items that collectively form the "task priority-power balance-energy consumption optimization" objective. The multi-objective reward function R... total It can be represented as:

[0055] R total = R1 + R2 + R3 + R4;

[0056] Among them, R1 is the task priority reward, which sets differentiated reward values ​​based on the urgency of marine ranching tasks, with the reward value for cage inspection tasks being higher than that for regular water quality monitoring tasks; R2 is the power balance reward, which sets positive and negative rewards based on the difference in the remaining power of the unmanned surface vessel (USV) formation, with higher reward values ​​for more balanced power distribution; R3 is the energy consumption reward, which sets rewards based on the energy consumption per unit path of the USV, with positive rewards given when energy consumption is lower than a preset threshold; R4 is the constraint penalty reward, which applies a negative reward if the assigned action would cause the USV to cross a restricted area or be assigned to an USV with insufficient power to return.

[0057] In this implementation, the reward function includes task priority reward R1, energy consumption optimization reward R2, power balancing reward R3, and constraint penalty reward R4. The purpose of the task priority reward is to ensure that high-priority tasks are executed first.

[0058] The task priority reward R1 can be represented as:

[0059] ;

[0060] Among them, P j ω represents the priority of the task. p It is a weighting coefficient; whenever a task is successfully assigned, a positive reward is given, the size of which is proportional to the priority of the task. This incentivizes the agent to prioritize assigning high-value tasks.

[0061] The energy consumption optimization bonus R2, used to reduce the overall fleet's sailing energy consumption and extend operational time, can be expressed as:

[0062] ;

[0063] in, These are weighting coefficients. This is the distance from the unmanned surface vessel (USV) to the mission point. The reward is inversely proportional to the distance the USV is assigned to the mission point (i.e., a negative reward / penalty is imposed). This incentivizes the agent to assign tasks based on proximity, reducing the total travel distance.

[0064] The power balancing bonus R3 is used to prevent some ships from being overloaded while others are idle, thus balancing the fleet load. It can be expressed as:

[0065]

[0066] in, These are weighting coefficients. The standard deviation of the entire fleet's battery level serves as a global reward. It also penalizes the overall battery level standard deviation. A larger standard deviation indicates a more uneven battery distribution, resulting in a greater penalty. This forces the agent to consider the battery status of all ships when assigning tasks, proactively assigning tasks to ships with high battery levels and allowing ships with low battery levels to rest.

[0067] The constraint penalty reward R4, used to ensure that the allocation scheme conforms to physical and security constraints, can be represented as:

[0068] ;

[0069] in, This is a weighting coefficient. If assigning an action would cause the unmanned surface vessel (USV) to cross a restricted area, or if it is assigned to an USV with insufficient battery power to return, a large negative reward (penalty) is applied to force the agent to avoid such behavior. Specifically, the ranch restricted area constraint is a preset safe zone around the aquaculture cages and reefs. The task execution order output by the upper-level task allocation module must ensure that the USV's planned path does not enter the preset safe zone to avoid collision risks.

[0070] S22. Pre-train the task allocation model in the Unity simulation environment. The pre-trained task allocation model is used to output the task execution order of each unmanned surface vessel based on its initial position, battery level, a series of task points and task priorities.

[0071] like Figure 2 The diagram illustrates the training process of the upper-layer task allocation model. The Proximal Policy Optimization (PPO) reinforcement learning algorithm is used in the upper-layer task allocation training process. First, the simulation environment is initialized. In a high-fidelity scene built with Unity, the initial positions, battery levels, and a series of task points and priorities of the unmanned surface vessels (USVs) are randomly generated. The agent, acting as the central task planner, observes the environmental state (including the positions, battery levels, and task information of all USVs) based on its current policy π(a|s) (parameterized by the Actor network) and outputs a task allocation action. After the environment executes this action (e.g., assigning a USV to a specific task point), the state transitions, and a multi-objective reward signal r, which integrates task priority, navigation energy consumption, and fleet battery balance, is returned. t At the same time, a new environmental state s is generated.t+1 This interaction process is stored as an experience tuple (s t ,a t , r t , s t+1 Finally, during the network update phase, the critic network minimizes the mean squared error between its predicted state value and the actual reward through temporal difference learning to accurately assess the state value. Simultaneously, the actor network optimizes according to the PPO-Clip objective function, which, by introducing the clipping probability ratio, constrains the magnitude of policy updates, ensuring training stability and avoiding drastic degradation of policy performance. This interaction and update process iterates until the policy performance converges, thereby obtaining an optimal task allocation policy that can adaptively balance multiple objectives and possesses strong generalization capabilities.

[0072] S22. Based on particle swarm optimization and A* algorithm, a path planning model is constructed with the goal of avoiding obstacles in aquaculture cages and reducing ocean current energy consumption. According to the task execution order of the unmanned surface vessel, the global optimal path for each unmanned surface vessel is generated through the path planning model.

[0073] Particle Swarm Optimization (PSO) is a population-based stochastic optimization algorithm that simulates the social behavior of flocks of birds or schools of fish. A swarm of "particles" flies through the search space. Each particle represents a potential solution, and its position is determined by its velocity. The particles update themselves by tracking two "extremes," aiming to find a solution that optimizes the objective function. The A* (A-Star) algorithm is a classic algorithm for finding efficient paths in a graphical plane. It combines the advantages of Dijkstra's algorithm and greedy best-first search.

[0074] Specifically, based on particle swarm optimization and the A* algorithm, a path planning model is constructed with the goals of avoiding obstacles in aquaculture cages and reducing ocean current energy consumption. This model generates the globally optimal path for each unmanned surface vessel (USV) according to its task execution order, including:

[0075] First, the initial path node set is selected using the particle swarm optimization algorithm. The fitness function is set and the constraint is applied to ensure that the distance between the path and the edge of the net cage meets the safety obstacle avoidance requirements of the marine ranch. Then, the particle is encoded and the particle swarm is initialized. The particle fitness is calculated and nodes that may touch the no-navigation zone are removed. Finally, the initial path node set that meets the obstacle avoidance requirements is selected.

[0076] Using the initial path node set as the starting / ending point, a heuristic function is designed based on the A* fusion algorithm to fuse target distance and ocean current resistance. By balancing path length and energy consumption cost through weight allocation, a smooth, collision-free path is generated. The heuristic function comprehensively considers the basic distance from the node to the target point and the influence of ocean currents at the current location. It balances path length and energy consumption cost through dynamic weighting—when ocean current interference is strong, the proportion of the resistance coefficient in the function is appropriately increased to guide the algorithm to generate a low-energy path that better reflects actual sea conditions; when the ocean current is stable, the focus is on optimizing path directness. The final output global path avoids all net cage obstacles and conforms to the direction of ocean currents to reduce ineffective energy consumption.

[0077] In this example, step S22 is implemented based on the following prior environment model: an environment map (Map), a set of no-navigation zones (O), feasible channels (C), and an ocean current field (V(x, y)). Where: the environment map (Map) is a two-dimensional raster map constructed from electronic charts and known cage coordinates. The set of no-navigation zones (O) is defined by expanding outwards from the coordinates of all cages by at least R. avoid The set of closed polygons formed by meters represents the absolute obstacle avoidance zone. Feasible path C is defined by all paths on the map that satisfy Distance(p,o). i R init (For all obstacles) i A connected region consisting of points p ∈ O, where R init >R avoid (e.g. R) init =10m). This channel serves as the initialization space for path nodes. Ocean current field V(x, y): A two-dimensional vector field defined on map space, representing the ocean current velocity and direction at each location.

[0078] Based on electronic charts and known cage coordinates, a 'feasible channel' at a certain distance from the boundaries of all cages is constructed as an initialization region. The particle swarm optimization algorithm runs within this channel, with the fitness function of its particles optimized by the distance to the target point. A constraint mechanism is introduced that imposes a huge penalty if the distance from the cage is less than the safe distance, thereby quickly selecting a set of initial nodes that meet the safety obstacle avoidance conditions.

[0079] First, particle encoding is performed, and the position X of each particle i is determined. i Represents a potential sequence of path nodes, namely X i =(Pos {i,1} Pos {i,2} , ..., Pos {i, K-1} ), where each Pos {i,j}Let C be a two-dimensional coordinate system. All particles are initially positioned randomly within feasible channels C. A fitness function is set, and constraints are applied. To force particles to satisfy obstacle avoidance constraints, the following fitness function is designed to evaluate particle performance: The fitness function is expressed as follows:

[0080] ;

[0081] Where: L(X) i The path represented by the particle is the total length, i.e., Path = S→Pos. {i,1} →...→Pos {i,K-1} →The sum of the Euclidean distances of each segment of G. P(X) i ) represents the penalty item, and its calculation formula is:

[0082] ;

[0083] Here, M is a very large positive penalty coefficient (e.g., 1e6), and I(·) is an indicator function that has a value of 1 when the condition within the parentheses is met (i.e., the node enters the no-fly zone), and 0 otherwise. This design ensures that any particle that violates the safety constraints will have its fitness drastically deteriorate, thus being eliminated in the evolutionary process.

[0084] Then, particle updates are performed. Particles iterate according to the standard PSO velocity and position update formulas. However, after the position update, a projection operation is required. If the new position falls outside the feasible path C, it is pulled back to the nearest boundary of C to ensure the search always proceeds within a safe space. The velocity update calculation formula is:

[0085] ;

[0086] The location update calculation formula is:

[0087] ;

[0088] The final output is an optimal particle position X. best The node sequence it represents PSO This refers to the set of high-quality initial path nodes that satisfy obstacle avoidance constraints.

[0089] Secondly, the output set of high-quality initial path nodes that satisfy the obstacle avoidance constraints is used. PSO To guide the algorithm, the A* algorithm is run on a refined raster map. The core of the algorithm lies in designing a heuristic function that incorporates ocean current dynamics costs to achieve energy optimization. First, node costs are evaluated; the evaluation function of the A* algorithm... The calculation formula is:

[0090] ;

[0091] Where g(n) is the actual cumulative cost from the starting point S to the current node n, and h(n) is the heuristically estimated cost from the current node n to the target point G. In this invention, g(n) not only includes the geometric path length but also incorporates the work done by ocean current resistance, and its calculation formula is as follows:

[0092]

[0093] Where |n - parent(n)| is the Euclidean distance to the parent node, V(n) is the ocean current vector at node n, and u {p→n} It is the unit direction vector pointing from the parent node to n, and κ is the ocean current influence coefficient. The term (V(n) · u) {p→n} ) is the dot product of the ocean current and the direction of navigation. When the current is reversed, it is negative, which leads to an increase in (1 - κ*(negative value)), that is, the cost of the reverse section increases. When the current is downstream, it is positive, which reduces the cost, thus reflecting the difference in energy consumption in the cumulative cost.

[0094] The formula for calculating the heuristic estimation cost h(n) is as follows:

[0095] ;

[0096] Where |n - G| is the Euclidean distance to the target, V avg It is the average ocean current vector from region n to region G, u {n→G} It is the unit direction vector from n to G, and λ is the heuristic drag coefficient. Term |V avg · u {n→G} | Reflects the average influence intensity of the ocean current in the remaining path direction. It is always positive, ensuring that h(n) does not exceed the actual cost, and can effectively guide the algorithm to avoid strong countercurrent areas.

[0097] Finally, a global path from S to G is generated by evaluating the function f(n) using the A* algorithm. This path minimizes navigation energy consumption by dynamically balancing path length and ocean current effects, while maintaining strict obstacle avoidance.

[0098] S23. Construct a path tracking model based on the MAPPO algorithm and train the path tracking model in the Unity simulation environment; based on the global optimal path of each unmanned surface vessel and the local observation information received from each unmanned surface vessel, output the coordinated reference heading angle and velocity in real time.

[0099] MAPPO (Multi-Agent Proximal Policy Optimization) is a multi-agent reinforcement learning algorithm based on PPO, specifically designed to solve decision-making problems in multi-agent cooperative and competitive environments. MAPPO employs a centralized training and distributed execution architecture. A path-tracking model is built based on the MAPPO algorithm. When a task sequence requires strict spatiotemporal coordination among multiple vessels, each vessel activates its path-tracking model. This model, based on the cooperative strategy obtained through centralized training, receives local observation information from each vessel and outputs real-time reference heading angles and velocities for coordination, maintaining formation and avoiding collisions while tracking the path.

[0100] Specifically, a path tracking model is constructed based on the MAPPO algorithm, and the path tracking model is trained in the Unity simulation environment, including:

[0101] S231. Based on the MAPPO algorithm, the global state is defined as including the position, speed, heading angle, reference path point sequence, ocean current vector of key points, and formation state of the vessels. The local observation includes the sensor information of each vessel and limited neighborhood information. The action space includes the output of each agent as continuous control commands: the change in heading angle and the desired speed.

[0102] S232. Design a centralized Critic network and a distributed Actor network. During the training phase, for multi-ships that require strict spatiotemporal coordination, a shared centralized Critic network is used. The Critic network takes the global state as input and outputs a scalar value. Each agent has an independent Actor network, whose input is its own local observations and whose output is an action probability distribution.

[0103] S233. Set the reward function, which includes tracking reward, collision avoidance reward, smoothness reward, and formation reward.

[0104] S233. Based on the Unity simulation environment, state data and environmental perception data of all unmanned surface vessels (USVs) are collected to construct a training scenario containing multiple ocean current interference modes. The MAPPO algorithm is then trained centrally, and a formation tracking strategy is learned through multi-agent collaborative decision-making. After the path tracking model converges to a stable formation cooperative strategy, it can be used to output the cooperative reference heading angle and velocity in real time based on the globally optimal path of each USV and the local observation information received from each USV.

[0105] During centralized training of the MAPPO algorithm, state data (position, velocity, heading angle, etc.) and environmental perception data (cage location, current velocity, obstacle distribution, etc.) of all unmanned surface vessels (USVs) are collected. Training scenarios with various current disturbance modes are constructed and used to train the MAPPO algorithm. Through multi-agent collaborative decision-making and learning of formation tracking strategies, the MAPPO algorithm masters the collaborative logic of maintaining formation and correcting deviations in complex flow fields. During centralized training, a preset number of iterations and learning rate are used to ensure the model converges to a stable formation coordination strategy. During distributed execution, each USV autonomously outputs a desired heading angle and velocity that meet global coordination requirements based on local perception data and interaction information with neighboring USVs. The adjustment range of the desired heading angle and velocity is adapted to the power performance and operational needs of the USVs in the marine ranching area. This achieves accurate tracking of the planned path while adapting to dynamic environmental changes in the ranching area.

[0106] like Figure 3 As shown, the MAPPO algorithm for the path tracking model is trained in a centralized manner. First, after the Unity simulation environment is initialized, each unmanned surface vessel (USV) samples and executes actions based on its current policy and local observations. After the environment state transitions, new observations and global rewards are generated. At the same time, empirical data containing the global state, observations of each USV, actions, and rewards are stored in a buffer. Then, data is sampled from the buffer. First, the centralized critic network is updated to minimize the mean square error of value prediction. Then, the actor network of each USV is updated using the PPO-Clip objective function. The training stability is ensured by constraining the policy update magnitude. After training, only the policy network of each USV is retained in the actual ship deployment, so that it can generate control commands independently based entirely on its own local observations, thereby achieving decentralized autonomous collaborative control.

[0107] In this embodiment, the MAPPO algorithm used in the path tracking model adopts a "centralized training-distributed execution" architecture. First, the state space, observation space, and action space are defined, and the global state s t It contains all the real-world information from the simulation environment and is only available to the critic network during intensive training:

[0108] s t = [P1, V1, ψ1, Path1, ..., P n V n , ψ n Path n V current1 , ..., V currentk ,F];

[0109] Where P1, V1, and ψ1 represent the position, velocity, and heading angle of the i-th vessel, respectively. Path iLet V be the sequence of reference path points for the i-th vessel. currentk Let F be the ocean current vector at k key points, and F be the formation state (such as the distance between boats).

[0110] Local observation o t i This includes sensor information from each vessel itself and limited neighborhood information, which serve as input to the actor network.

[0111] o t i = [e i , ψ ei V i , ψ i V currentlocal Neighbor Infoi ];

[0112] Among them, e i This represents the path tracking error (e.g., the distance to the current target path point). ψ ei This represents the heading error (the difference between the current heading and the desired heading). V currentlocal This refers to locally sensed ocean current vectors (obtainable through model estimation or sensors). Neighbor Infoi It provides communication information (such as relative position and speed) from neighboring vessels for coordinated collision avoidance and formation maintenance.

[0113] Action space a t i The output of each agent consists of a series of control commands.

[0114] a t i = [Δψ desired V desired ];

[0115] Where, Δψ desired V represents the desired change in heading angle. desired The desired speed.

[0116] Next, a centralized Critic network and a distributed Actor network are designed. During the training phase, for multi-ship systems requiring strict spatiotemporal coordination, a shared centralized Critic network V(s) is used. t It takes the global state s as input. t It outputs a scalar value to evaluate the expected total reward that all agents can obtain in the future under the current global state. Each agent has an independent Actor network π. i (a t i | o ti It is a policy function whose input is its own local observations. t i The output is an action probability distribution.

[0117] In the MAPPO algorithm, a reward function is set. It must include tracking rewards, collision avoidance rewards, smoothness rewards, and formation rewards, encompassing multiple objectives such as tracking accuracy, coordination, energy consumption, and smoothness. The reward function... It can be represented as:

[0118] ;

[0119] r trackingi For tracking rewards:

[0120] ;

[0121] Here, ω1 and ω2 are weights, which encourage the reduction of path and heading errors.

[0122] collisioni Collision avoidance reward:

[0123] ;

[0124] Among these, a severe penalty is imposed when the distance to other boats or obstacles is less than the safe distance d_safe.

[0125] r smoothi For smoothness rewards:

[0126] ;

[0127] Excessive changes in heading are penalized, resulting in smoother control actions, reduced energy consumption, and less equipment wear.

[0128] r formation Formation rewards:

[0129] ;

[0130] Where, d actual It is the actual distance between the boats, d desired This is the desired formation distance. This incentive maintains the predetermined formation.

[0131] S24, PID control submodule, is used to generate a PWM drive signal for controlling the propulsion actuator after compensating for actual ocean current interference by PID closed-loop adjustment based on the desired heading angle and desired speed output by the path tracking module.

[0132] Specifically, the PID control submodule adopts a composite strategy of "dynamic parameter self-tuning + ocean current feedforward compensation". After receiving the desired heading angle and desired speed output by MAPPO, it first analyzes the actual motion state of the unmanned surface vessel and the characteristics of ocean current interference through a state observer combined with multi-sensor data. Then, it dynamically adjusts the PID parameters according to the interference intensity: when strong ocean currents cause deviations to exceed limits, it increases the proportional coefficient to speed up the response and decreases the integral coefficient to prevent overshoot. When the state is stable, it optimizes the parameters to achieve zero steady-state error tracking. At the same time, it calls the ocean current feedforward model to generate advance compensation, which is superimposed with the PID basic control quantity. Then, it generates a PWM drive signal through coordinated adjustment of frequency and duty cycle: the propeller adapts the frequency according to speed and load (low-speed frequency reduction for energy saving, high-speed frequency increase to maintain power), and the servo uses a fixed high frequency with a fine duty cycle to fine-tune the heading. It can also collect feedback signals from the actuator in real time, triggering safety correction when deviations exceed limits, ensuring control accuracy and operational safety under complex ocean currents. The frequency and duty cycle adjustment range of the PWM drive signal are adapted to the drive requirements of the unmanned surface vessel's propulsion actuator, achieving smooth speed regulation and steering.

[0133] S3. Transfer the pre-trained control and mission planning model to the unmanned surface vessel (USV), fine-tune and optimize the control and mission planning model, and use the fine-tuned and optimized control and mission planning model to control and plan the USV based on the USV's state data and marine ranch environment data.

[0134] Specifically, the control and mission planning model is loaded into the actual ship's computing unit, and the parameters of the feature extraction layer of the control and mission planning model are frozen. State data of the unmanned surface vessel and marine ranch environmental data are collected through the sensor units of the actual ship to construct a real-environment training dataset. The control and mission planning model is trained using this real-environment training dataset, iteratively optimizing the fully connected layer parameters. By pre-setting the number of iterations and the learning rate, the control and mission planning model is adapted to the environmental differences of the real marine ranch, resulting in a finely tuned and optimized control and mission planning model. This method only performs lightweight iterative optimization on the fully connected network of the decision layer to achieve efficient and stable transfer of model knowledge from the simulation environment to the real scene.

[0135] like Figure 4 As shown in the diagram, the model parameter fine-tuning optimization strategy in this embodiment of the invention is implemented as follows:

[0136] First, model initialization and parameter freezing are performed. The control and task planning model, which has been pre-trained and converged in the Unity simulation environment, is loaded into the actual ship computing unit. Then, the parameters of all feature extractors, including convolutional layers and recurrent layers, are frozen to ensure they remain unchanged during subsequent fine-tuning. This aims to preserve the general environmental representation capabilities (such as cage shape features and ocean current dynamics patterns) learned by the model during pre-training, avoiding catastrophic forgetting of the model due to limited sample data from the real environment.

[0137] Secondly, by collecting state data of the unmanned surface vessel (USV) and marine ranching environment data using sensor units deployed on the actual vessel, a real-world training dataset was constructed, forming a fine-tuning training set. Key samples in the dataset include: real ocean current velocity vectors, the actual relative positions of the USV and the net cages (including positioning noise), and control actions and reward signals calculated from the actual navigation trajectory. This dataset is used to replace simulated data in the simulation environment, serving as a realistic basis for model fine-tuning.

[0138] Finally, constrained parameter space optimization is performed. Supervised / reinforcement learning fine-tuning is performed only on the parameters of the fully connected layers at the end of the model using a real-world training dataset. The optimization objective is to maximize the expected return (for PPO / MAPPO) or minimize the control error under the real data distribution. The optimization process is strictly constrained: the preset number of fine-tuning iterations is 500 epochs to prevent overfitting; a significantly reduced learning rate (typically 1 / 10 to 1 / 100 of the pre-training rate) is used to ensure smooth and controllable parameter updates.

[0139] Through this constrained optimization, the model can quickly adapt general feature representations to the specific noise and dynamic characteristics of real-world scenarios, thereby significantly reducing the performance differences between virtual and real systems caused by modeling mismatch, and ultimately improving the robustness and control accuracy of unmanned surface vessel systems in real marine ranches.

[0140] In a preferred embodiment, a virtual training environment highly consistent with the real-world scenario can be set up within the Unity simulation environment. Dynamic ocean currents utilize a fluid dynamics model calibrated based on actual pasture hydrological data to simulate flow field characteristics at different times (e.g., high / low tide). Dense net cages are modeled using 3D coordinates from laser scanning of the actual ship, preserving the spatial topology between cages. The salt spray sensor noise model is constructed by analyzing historical data (GPS drift, sonar echo attenuation) collected by the actual ship's ROS system, ensuring the noise distribution of the virtual sensor output matches reality. Simultaneously, the module has a built-in ROS communication interface, allowing direct reception and publication of actual ship ROS topics, injecting real-world environmental parameters into the simulation scene in real time, ensuring a high degree of consistency between the distribution characteristics of virtual training data and actual ship operational data. During the simulation phase, the virtual unmanned surface vessel's state and environmental data in Unity are mapped to standard ROS topics via ROS-Bridge, fully compatible with the message format of the actual ship's ROS system, enabling algorithm debugging based on the real communication protocol within the simulation environment. During algorithm deployment, the pre-trained model is encapsulated as a ROS node, acquiring real-world environmental data by subscribing to actual ship sensor topics, and publishing control command topics to drive the actuators. During the model fine-tuning phase, ROS nodes record error data from actual ship operations (such as path tracking deviation and measured energy consumption) in real time and feed it back to the simulation system. By comparing the differences between virtual and real data, the parameters of the fully connected layer are adjusted accordingly. The entire process does not require reconstruction of the communication architecture. Seamless migration from simulation to actual ship can be achieved simply through ROS topic adaptation, which greatly reduces deployment costs and improves algorithm reliability.

[0141] This embodiment provides a two-layer task planning-based unmanned surface vessel (USV) control and task planning method. Through a design of "upper-layer intelligent allocation - lower-layer precise execution - full-process virtual-real optimization," a control and task planning model is constructed. This model is pre-trained in a Unity simulation environment to obtain a pre-trained control and task planning model. The control and task planning model includes a task allocation model, a path planning model, a path tracking model, and a PID control submodule. This method can solve technical problems in marine ranching USV operations, such as unbalanced task allocation, difficulty in balancing path avoidance and energy consumption, and performance disconnect between virtual and real scenarios. Sensor units ensure data acquisition accuracy; the upper-layer task allocation model achieves intelligent coordination of multiple objectives; the lower-layer path planning and path tracking models balance path safety and formation coordination; the PID control submodule ensures execution accuracy under ocean current interference; and virtual-real fusion supported by the ROS system enables seamless migration of the control and task planning model, thereby improving overall task response efficiency and reducing energy consumption and deployment costs.

[0142] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A two-layer mission planning-based unmanned surface vessel (USV) control and mission planning method, characterized in that, Includes the following steps: S1. Based on multiple unmanned surface vessels (USVs), an unmanned surface vessel (USV) hardware formation is formed. Each USV is equipped with a sensor unit and a propulsion actuator adapted to offshore operations. The sensor unit is used to collect the status data of the USV and the marine ranch environment data. S2. Construct a control and task planning model. Pre-train the control and task planning model in the Unity simulation environment to obtain a pre-trained control and task planning model. The control and task planning model includes a task allocation model, a path planning model, a path tracking model, and a PID control sub-module. The task allocation model is used to output the task execution order of each unmanned surface vessel (USV) based on its initial position, battery level, a series of task points, and task priorities. The path planning model employs particle swarm optimization and A* algorithm to avoid obstacles in aquaculture cages and reduce ocean current energy consumption. It generates the globally optimal path for each unmanned surface vessel (USV) based on the task execution order. The path tracking model is used to output a coordinated reference heading angle and velocity in real time based on the global optimal path of each unmanned surface vessel and the local observation information received from each unmanned surface vessel. The PID control submodule is used to generate a PWM drive signal for controlling the propulsion actuator based on the reference heading angle and speed output by the path tracking module, after PID closed-loop adjustment to compensate for actual ocean current interference. The construction of the control and task planning model, and the pre-training of the control and task planning model in the Unity simulation environment to obtain the pre-trained control and task planning model, includes: Based on the PPO reinforcement learning algorithm, a task allocation model is constructed by combining marine ranching task priority, unmanned surface vessel remaining battery power, and ranching no-navigation zone constraints. The task allocation model is pre-trained in the Unity simulation environment, and the task execution order of each unmanned surface vessel is output through the task allocation model. Based on particle swarm optimization and A* algorithm, a path planning model is constructed with the goal of avoiding obstacles in aquaculture cages and reducing ocean current energy consumption. According to the task execution order of the unmanned surface vessel, the path planning model generates the global optimal path for each unmanned surface vessel. First, the initial path node set is selected using the particle swarm optimization algorithm. The fitness function is set and the constraint is processed to ensure that the distance between the path and the edge of the net cage meets the safety obstacle avoidance requirements of the marine ranch. Then, the particle is encoded and the particle swarm is initialized to calculate the particle fitness and remove nodes that may touch the no-navigation zone. The initial path node set that meets the obstacle avoidance requirements is then selected. Using the initial path node set as the starting / ending point, a heuristic function for fusing target distance and ocean current resistance is designed based on the A* fusion algorithm. By balancing path length and energy consumption cost through weight allocation, a smooth, collision-free path is generated. A path tracking model is constructed based on the MAPPO algorithm and trained in the Unity simulation environment. The path tracking model is used to output the coordinated reference heading angle and velocity in real time based on the global optimal path of each unmanned surface vessel and the local observation information received from each unmanned surface vessel. S3. Transfer the pre-trained control and mission planning model to the unmanned surface vessel (USV), fine-tune and optimize the control and mission planning model, and use the fine-tuned and optimized control and mission planning model to control and plan the USV based on the USV's state data and marine ranch environment data.

2. The dual-layer mission planning unmanned surface vessel control and mission planning method according to claim 1, characterized in that, The sensor unit includes: a positioning sensor, an attitude sensor, and an obstacle detection sensor. It collects state data of the unmanned surface vessel, including the unmanned surface vessel's position and attitude data. It also collects marine ranch environmental data, including the location of net cages, ocean current speed, and obstacle distribution. The propulsion actuator includes: a thruster and a servo motor.

3. The dual-layer mission planning unmanned surface vessel control and mission planning method according to claim 1, characterized in that, The task allocation model, constructed based on the PPO reinforcement learning algorithm and incorporating marine ranching task priorities, unmanned surface vessel remaining battery power, and ranching no-navigation zone constraints, includes: Based on the PPO reinforcement learning algorithm, the objective function of the task allocation model is set as a multi-objective reward function. This multi-objective function includes: task priority reward, energy consumption optimization reward, power balance reward, and constraint penalty reward. The multi-objective reward function R... total Represented as: R total = R1+ R2+ R3+ R4; Among them, R1 is the task priority reward, R2 is the power balance reward, R3 is the energy consumption reward, which is set according to the energy consumption per unit path of the unmanned surface vessel, and a positive reward is given when the energy consumption is lower than the preset threshold; R4 is the constraint penalty reward. The task allocation model is pre-trained in the Unity simulation environment. The pre-trained task allocation model is used to output the task execution order of each unmanned surface vessel based on its initial position, battery level, a series of task points and task priorities.

4. The dual-layer mission planning unmanned surface vessel control and mission planning method according to claim 1, characterized in that, The fitness function is expressed as follows: ； Among them, X i For each particle i at its position, X i L(X) represents a potential sequence of path nodes. i P(X) represents the total length of the path represented by the particle. i () is a penalty item; The formula for calculating the evaluation function is expressed as follows: ； Where g(n) is the actual cumulative cost from the starting point S to the current node n, and h(n) is the heuristically estimated cost from the current node n to the target point G.

5. The dual-layer mission planning unmanned surface vessel control and mission planning method according to claim 1, characterized in that, The path tracking model based on the MAPPO algorithm is constructed and trained in the Unity simulation environment; including: Based on the MAPPO algorithm, the global state is defined as including the position, speed, heading angle, reference path point sequence, ocean current vector of key points, and formation state of the vessels. The local observation includes the sensor information of each vessel and limited neighborhood information. The action space includes the output of each agent as continuous control commands: heading angle and desired speed. The design incorporates a centralized Critic network and a distributed Actor network. During the training phase, for multi-agent systems requiring strict spatiotemporal coordination, a shared centralized Critic network is used. The Critic network takes the global state as input and outputs a scalar value. Each agent, on the other hand, has an independent Actor network, whose input is its own local observations and whose output is an action probability distribution. Set a reward function, which includes tracking reward, collision avoidance reward, smoothness reward and formation reward; Based on the Unity simulation environment, state data and environmental perception data of all unmanned surface vessels were collected to construct training scenarios with various ocean current interference modes. The MAPPO algorithm was trained in a centralized manner, and a formation tracking strategy was learned through multi-agent collaborative decision-making.

6. The dual-layer mission planning unmanned surface vessel control and mission planning method according to claim 1, characterized in that, The step of generating a PWM drive signal for controlling the propulsion actuator's operation, based on the desired heading angle and desired velocity output by the path tracking module, after PID closed-loop adjustment to compensate for actual ocean current interference, includes: After receiving the desired heading angle and desired velocity output by MAPPO, the actual motion state and ocean current disturbance characteristics of the unmanned surface vessel are analyzed by combining multi-sensor data with the state observer. The PID parameters are dynamically adjusted according to the intensity of the disturbance. When a strong ocean current causes the deviation to exceed the limit, the proportional coefficient is increased to speed up the response and the integral coefficient is reduced to prevent overshoot. When the state is stable, the parameters are optimized to achieve zero steady-state error tracking, and an advance compensation amount is generated and superimposed with the basic PID control amount. The PWM drive signal is generated by coordinating the frequency and duty cycle. The thruster adapts the frequency according to the speed and load, and the servo uses a fixed high frequency in conjunction with a fine duty cycle to finely adjust the heading.

7. The dual-layer mission planning unmanned surface vessel control and mission planning method according to claim 1, characterized in that, The process of transferring the pre-trained control and mission planning model to the unmanned surface vessel (USV) and fine-tuning and optimizing the model includes: The control and mission planning model is loaded into the actual computing unit of the unmanned surface vessel (USV), and the parameter of the feature extraction layer of the control and mission planning model is frozen. Based on the sensor unit of the actual vessel, the state data of the USV and the marine ranch environment data are collected to construct a real environment training dataset. The control and mission planning model is trained on the real environment training dataset, and the fully connected layer parameters of the control and mission planning model are iteratively optimized. By pre-setting the number of iterations and the learning rate, the control and mission planning model is adapted to the environmental differences of the real marine ranch, resulting in a finely tuned and optimized control and mission planning model.

Images