Fishing boat sea adaptive operation coordination system based on large model agent

Abstract

The application discloses a fishing boat offshore adaptive operation cooperation system based on a large model intelligent agent, and relates to the technical field of ship cooperation operation.The system comprises the following steps: an acquisition module, which is used to acquire a current ship cooperation scheme, fish school state data and ship state data; a first determination module, which is used to determine a dynamic escape channel based on the ship state data and a cooperation correlation; a prediction module, which is used to predict a future reachable area of the fish school according to the fish school state data, and to calculate an escape overlap amount between the future reachable area of the fish school and the dynamic escape channel; a second determination module, which is used to determine a delay contribution degree corresponding to each ship task based on a decline range of the escape overlap amount after a single task delay is eliminated; and a generation module, which is used to determine a main task set to be corrected according to the delay contribution degree, and to input the main task set into a local cooperation large model to generate a corrected cooperation scheme.The application improves the stability and success rate of multi-ship cooperation operation.

Description

Technical Field

[0001] This application relates to the field of vessel collaborative operation technology, specifically to a fishing vessel maritime adaptive collaborative operation system based on a large model intelligent agent. Background Technology

[0002] Large-scale modeling techniques have been attempted to generate collaborative fishing schemes for fishing vessels. However, in actual offshore purse seine operations, fishing vessels are often affected by sea conditions, resulting in execution delays. Existing methods typically rely solely on the positional deviation of a single vessel to determine whether to replan the route, making it difficult to identify the global impact of a single vessel's delay propagating to other vessels through collaborative relationships such as "waiting" and "synchronization." Furthermore, simply relying on large-scale models for open-loop text planning is insufficient; there is an urgent need to upgrade large-scale models to agents with closed-loop capabilities of "perception-decision-action-feedback." However, existing systems lack computational mechanisms to support agents in environmental perception and feedback, making it impossible to accurately assess whether execution deviations have truly altered the purse seine boundary and provided escape routes for fish. This easily leads to frequent and ineffective global reconfigurations for minor deviations, while critical delays that truly cause fish escape are not corrected in a timely manner, severely impacting the stability and success rate of multi-vehicle collaborative operations. Summary of the Invention

[0003] To address the technical problem in related technologies where determining whether to replan a route based on the positional deviation of a single vessel fails to accurately assess whether the execution deviation has truly altered the encirclement boundary and left an escape route for fish, thus severely impacting the stability and success rate of multi-vehicle collaborative operations, this application provides a fishing vessel maritime adaptive collaborative operation system based on a large model intelligent agent.

[0004] The specific technical solution adopted is as follows: The acquisition module is used to acquire the current ship coordination scheme, fish swarm status data, and ship status data. The ship coordination scheme includes multiple ship tasks and the coordination relationships between the various ship tasks. The first determination module is used to determine the dynamic escape channels of the fish school based on ship status data and collaborative relationships. The prediction module is used to predict the future reachable area of ​​the fish based on the fish status data, and to calculate the escape overlap between the future reachable area of ​​the fish and the dynamic escape channel. The second determining module is used to determine the delay contribution of each ship mission based on the decrease in the escape overlap after eliminating the delay of a single mission when the escape overlap exceeds the preset historical normal range. The generation module is used to determine the set of main tasks to be corrected based on the contribution of delay, input the set of main tasks into the local collaborative big model, generate a corrected collaborative solution, and issue it for execution.

[0005] In one possible implementation of this application, the first determining module includes: The first determination submodule is used to determine the estimated execution time of each ship's mission based on ship status data and collaborative relationships; The reconstruction submodule is used to reconstruct the future capture boundary based on the expected execution time, and compare the future capture boundary with the planned boundary of the ship coordination scheme to obtain the dynamic escape channel of the fish school.

[0006] In one possible implementation of this application, the first determining submodule includes: The determination unit is used to determine the direct delay of each unfinished ship task based on ship status data; The propagation unit is used to propagate the direct delay along the collaborative relationship to obtain the estimated execution time of each ship's mission after propagation.

[0007] In one possible implementation of this application, the determining unit is specifically used for: Extract the mission start position, target relative position, and current relative position of any ship mission from the ship status data; Determine the task completion progress based on the task start position, the target relative position, and the current relative position; Based on the progress of the task completion and the speed at which the ship completes the task, determine the remaining execution time of the ship's task; Based on the sum of the current time and the remaining execution time, the planned completion time corresponding to the ship coordination scheme is used to determine the direct delay of each unfinished ship task.

[0008] In one possible implementation of this application, the refactoring submodule is specifically used for: Obtain the capture boundaries corresponding to each boundary task from the ship status data. The capture boundaries include the established fixed boundaries, the boundaries in formation, and the boundaries that have not yet been formed. For both formed and unformed boundaries, the positions of fishing vessels and net endpoints at the predicted future time are determined according to the planned time and expected execution time of the vessel coordination scheme. At each predicted future time, the positions of each fishing boat and the endpoints of the nets within the encirclement boundary are arranged in the order of boundary connection to obtain the future encirclement boundary.

[0009] In one possible implementation of this application, the refactoring submodule is further configured to: At each future predicted time, based on the original planned time of the ship coordination scheme, the planned execution progress of each boundary task is determined, and the planned encirclement boundary is constructed; Extract the unclosed area between adjacent endpoints in the planned enclosure boundary as the original open area; At each future prediction time, the area between adjacent boundary endpoints in the future enclosure boundary that is not yet connected by fishing boats, nets, or effective blocking conditions is regarded as the open area after propagation; The dynamic escape channels of the fish swarm are obtained by performing difference calculations between the open region after propagation and the original open region.

[0010] In one possible implementation of this application, the prediction module is specifically used for: By processing the fish swarm state data using a linear fitting method, the predicted center position of the fish swarm at any future time can be obtained. Extract the fish swarm trajectory that is the same as the current operation stage within the historical time period, and construct the historical motion extension area based on the offset point of the fish swarm trajectory relative to the prediction center corresponding to the historical time. By shifting the historical movement extension area to the prediction center location, the future reachable area of ​​the fish school can be determined. The escape overlap is calculated based on the overlap area between the future reachable area and the dynamic escape channel.

[0011] In one possible implementation of this application, the second determining module is specifically used for: For any given ship mission, determine the first escape overlap amount resulting from eliminating the direct delay of the current ship mission; The delay contribution of each ship's mission is determined based on the decrease between the escape overlap amount and the first escape overlap amount.

[0012] In one possible implementation of this application, the generation module is specifically used for: The ship missions with direct delays are sorted in descending order of their contribution to delay, and the sorting results are obtained. Based on the sorting results, the corresponding ship missions and downstream missions affected by the delay of the current ship mission in the collaborative relationship are added to the set to be corrected in turn; After each operation of adding a ship mission or downstream mission to the set to be corrected, the simulation eliminates the delay of all missions in the set to be corrected and recalculates the current simulated escape overlap. If the simulated escape overlap decreases to the preset historical normal range, then stop adding tasks and determine the current set of tasks to be corrected as the main set of tasks to be corrected; If the simulated escape overlap still does not decrease to the preset historical normal range after adding all tasks with direct delays to the set to be corrected, then the addition of tasks will be forcibly stopped, and all added tasks will be used as the main set of tasks to be corrected.

[0013] In one possible implementation of this application, the generation module is further configured to: The main task set, the current ship collaboration scheme, and ship status data are used as the model input information, and local correction constraints are set. Input the model input information and local correction constraints into the local collaborative large model to generate at least one set of candidate correction schemes; For each set of candidate correction schemes, the corresponding future capture boundary is re-deduced and the candidate escape overlap is calculated. Invalid candidate correction schemes that do not meet the local correction constraints or whose candidate escape overlap is greater than or equal to the escape overlap of the current ship coordination scheme are eliminated. From the remaining valid candidate correction schemes, the scheme with the smallest candidate escape overlap is selected as the final correction coordination scheme, and the correction coordination scheme is split and distributed to the corresponding fishing vessels for execution.

[0014] This application has, but is not limited to, the following technical effects: In this application, the dynamic escape channels of fish shoals are determined by analyzing ship status data and the collaborative relationships between ship tasks. Then, the future reachable areas of the fish shoals are predicted based on the fish status data, and the overlap between the future reachable areas and the dynamic escape channels is calculated. This accurately reflects whether deviations pose a risk of fish escape, avoiding oversensitivity to non-threatening deviations and invalid alarms. When the overlap exceeds a preset historical normal range, the delay contribution of each ship task is determined based on the decrease in overlap after eliminating the delay of a single task. Based on the delay contribution, the set of main tasks to be corrected is determined, and this set is input into the local collaborative model to generate a corrective collaborative plan and issue it for execution. This precisely identifies the key tasks that truly cause large gaps, limiting the large model to only locally correcting high-contribution tasks and their controlled downstream components. It avoids minor changes causing the large model to completely overturn and rearrange the entire fleet's tasks, significantly reducing unnecessary ship turning and command changes, and improving the stability and success rate of multi-ship collaborative operations. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system structure of the fishing vessel maritime adaptive operation collaborative system based on a large model intelligent agent, as presented in this application. Figure 2 This is a flowchart illustrating the first embodiment of the fishing vessel maritime adaptive operation collaborative system based on a large model intelligent agent in this application. Figure 3 This is a schematic diagram of the overall implementation process of the fishing vessel maritime adaptive operation collaborative system based on large model intelligent agents in this application; Figure 4 This is a schematic diagram illustrating the propagation process of various vessel tasks involved in the fishing vessel maritime adaptive operation collaborative system based on a large model intelligent agent, as described in this application. Figure 5This is a schematic diagram illustrating the prediction process of the future reachable area of ​​fish schools involved in the fishing vessel adaptive operation collaborative system based on a large model intelligent agent in this application. Figure 6 This is a schematic diagram illustrating the spatial overlap and development process of the future reachable area and escape channel of the fishing vessel's adaptive maritime operation collaborative system based on a large model intelligent agent, as described in this application. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0017] This application provides a fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent. In the first embodiment of this fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent, refer to... Figure 1 The system includes: The acquisition module 101 is used to acquire the current ship collaboration scheme, fish swarm status data, and ship status data. The ship collaboration scheme includes multiple ship tasks and the collaborative relationships between the various ship tasks.

[0018] As an example, refer to Figure 2 The corresponding execution steps of the fishing vessel adaptive operation collaborative system based on large model intelligent agents are as follows: Figure 2 As shown, a local large-scale model is first trained, and the current ship collaboration scheme is generated from the local large-scale model. Then, the current ship collaboration scheme is analyzed and corrected to obtain the final corrected collaboration scheme, which is then issued for execution, thereby improving the stability and success rate of multi-ship collaborative operations.

[0019] This embodiment achieves multi-ship adaptive collaboration by constructing a closed-loop intelligent agent architecture based on a large model. Specifically, the agent's "perception" mechanism consists of state acquisition and task progress calculation; the "decision" mechanism consists of generating initial and candidate corrective schemes through local fine-tuning of the collaborative large model; the "feedback" mechanism consists of task delay propagation, escape boundary reconstruction, overlap calculation, and contribution attribution; and the "action" mechanism consists of instruction splitting and issuance and ship execution.

[0020] As an example, scattered fish school data, fishing vessel status data, and historical coordination commands are processed into training data that can be directly used by a local large-scale model. Specifically, an independent data directory is first established according to a complete fishing operation, and fish finder data, vessel positioning and operation status data, and coordination commands are saved as structured files respectively. Then, the data processing program completes time alignment, short-term missing data completion, coordinate transformation, and relative position calculation to form a continuous multi-vessel status window with the fish school center as the reference. Finally, the moment of each coordination command is issued is used as the sample segmentation point. The continuous state before the command is issued is used as the model input, and the actual multi-vessel coordination command used at that moment is used as the model output, generating structured samples for model training, validation, and testing.

[0021] After this step, the raw equipment data is transformed as follows: raw fish school and fishing boat data → continuous state under a unified time axis → relative state of multiple boats with the fish school center as the origin → pairing of continuous state with historical coordination commands → forming large model training samples. Furthermore, the data stored in the fish finder, each boat's positioning equipment, net end-point equipment, historical coordination commands, and system operation logs are processed into structured data that can directly train a local large model and support subsequent calculations.

[0022] An open-source instruction model was selected as the base model, and an efficient parameter fine-tuning method (such as QLoRA) was employed. The two types of samples mentioned above were loaded, with the state and correction criteria used as cues, and the structured collaborative scheme used as the supervised output for supervised fine-tuning. After training, the adaptation parameters were merged with the base model.

[0023] Furthermore, the merged model is converted into a format suitable for edge computing (such as GGUF format) and low-bit quantization (such as 4-bit quantization) is performed to reduce memory usage. The quantized model is then deployed on the onboard computing equipment of the main work vessel. The model resides in memory and provides an offline inference interface through a local port, allowing it to run without an internet connection.

[0024] In actual operation, the shipborne data acquisition program automatically reads the latest status of the fish school and multiple vessels at regular intervals (e.g., every 5 seconds), calculates relative coordinates using the fish school as a reference point, and encapsulates this data into fixed-structure input data based on the current stage target. The local large-scale model inference interface is called to generate a multi-vessel collaborative scheme with a clear task number, numerical target location, fixed execution time period, and fixed collaborative relationships (e.g., independent, waiting, synchronous, pre-positioned), i.e., a vessel collaborative scheme. The system performs logical and format validation on the model output. After successful validation, a snapshot of the scheme is saved, and instructions are split and distributed to the corresponding fishing vessels for execution. A schematic diagram of the overall execution flow of this embodiment is shown below. Figure 3 As shown, where, Figure 3 The process section following step two is the main innovative part of this application.

[0025] It should be noted that when the collaborative instruction sending program in the system sends the initial ship collaboration plan, it simultaneously saves a snapshot of the current plan. The plan snapshot includes at least the plan number, plan version, task number, ship number, task role, planned start time, planned completion time, task start relative position, target relative position, target operation status, target retention duration, relationship type, associated task number, encirclement boundary number, and boundary connection order.

[0026] Among them, the relative position at the start of the task is obtained by copying the ship's status cache at the time the command is issued; the relative position of the target is obtained by the numerical field output by the large model; the relationship type is fixed as synchronous, waiting or preceding relationship; the encirclement boundary number is used to indicate which encirclement boundary is formed after the task is completed; the boundary connection sequence is used to determine the encirclement connection relationship between adjacent ships or net endpoints.

[0027] As an example, a vessel task can be a specific operational instruction assigned to a single fishing vessel, including the target's relative position, the target's operational status (e.g., casting nets on the port side, obstructing from the outside, closing off on the starboard side), and the duration of target retention. Coordination relationships can be the time and execution constraints between different vessel tasks, including synchronization, waiting, precedence, and independence relationships.

[0028] It should be noted that the shipboard processing program treats each ship task in the original scheme as a task node, and then connects the task nodes according to the relationship type and associated task number. It retains the actual collaborative constraints that exist in the original scheme of the large model. When a subsequent task is delayed, the system can determine which subsequent tasks will be postponed based on these saved relationships, rather than analyzing the deviations of each ship in isolation.

[0029] For example, vessel A can only enter the synchronous closure phase after completing the net deployment on the left side; vessel B can only close the closure phase together with vessel A after reaching the right side closure position; and vessel C needs to maintain its northeast outer perimeter blocking position before the synchronous closure. The system records the preceding tasks, waiting tasks, and synchronous task groups for each mission accordingly.

[0030] As an example, fish school status data describes the location and movement characteristics of a school of fish, including the coordinates of the school's center, the set of points on the outer edge of the school, the direction of movement, and the speed of movement. Vessel status data describes the current movement and behavior characteristics of a fishing vessel, including its relative position, heading, speed, and current operational status. Both types of data can be acquired using onboard status acquisition programs.

[0031] The first determining module 102 is used to determine the dynamic escape channel of the fish school based on ship status data and collaborative relationships.

[0032] As an example, the location of each fishing vessel is extracted from the vessel status data to calculate the direct delay that occurs when the fishing vessels perform their respective tasks. Then, based on the collaborative relationship between the fishing vessels, the future encirclement boundary is determined. Finally, the future encirclement boundary is compared with the planned boundary of the vessel collaboration scheme to obtain the dynamic escape channel of the fish school.

[0033] As an example, a dynamic escape channel refers to an unclosed spatial gap that is added or extended relative to the boundary that should have been formed in the original plan at a future predicted time due to the delayed propagation of various ship missions.

[0034] The first determining module 102 further includes: The first determination submodule is used to determine the estimated execution time of each ship's mission based on ship status data and collaborative relationships.

[0035] As an example, the estimated execution time represents the future actual completion time of a vessel's mission, calculated after considering its own physical movement progress (direct delay) and the combined effects of all constrained prerequisite tasks and waiting relationships. The estimated execution time of each vessel's mission can be determined by analyzing the mission completion progress of each fishing vessel and their collaborative relationships.

[0036] The first determining submodule includes: The determination unit, based on ship status data, determines the direct delays for each unfinished ship task. This step specifically includes: Extract the mission start position, target relative position, and current relative position of any ship mission from the ship status data.

[0037] The task completion progress is determined based on the task start position, the target relative position, and the current relative position.

[0038] As an example, for ship missions with numerical target positions such as approaching, filling in gaps, and entering closed positions, the ship's position relative to the fish school at the start of the mission is first read, which is the mission start position. Then, the target relative position and the actual current relative position are extracted. Then, the mission start position, target relative position, and current relative position are uniformly converted into coordinate points in the Earth's absolute static geographic coordinate system. Finally, the converted absolute coordinate vectors are substituted into the dot product formula to calculate the mission completion progress.

[0039] Let the task number be The task starts at the location where The relative position of the target is The current relative position is The three vectors mentioned above are derived from the task start point field, target position field, and current state cache in the snapshot of the ship collaboration scheme, respectively.

[0040] Furthermore, to avoid incorrectly including lateral displacement caused by the ship's lateral movement or the influence of waves in the effective progress, the system projects the current displacement onto the direction from the starting point of the task to the relative position of the target, thus obtaining the task completion progress. : In the formula, For the task Progress at the current moment; " represents the vector dot product;" " indicates the length of the vector; This indicates that the result will be limited to between 0 and 1. A progress of 0 indicates that progress has not yet been made in the target direction, while a progress of 1 indicates that the relative position to the target has been reached. It is only used to convert the actual ship position into mission progress; it does not directly evaluate the importance of the mission, nor does it directly trigger large-scale model corrections.

[0041] It should be noted that for non-positional tasks such as waiting, maintaining perimeter obstruction, and maintaining net deployment, the progress is determined directly based on the target operational status and target holding time: 0 when not yet in the target status; 1 when in the target status but not yet meeting the holding time requirement; and 1 when the holding requirement is met. The progress for waiting tasks is determined based on whether the task being waited for has reached the specified status.

[0042] The remaining execution time of the ship's mission is determined based on the progress of the mission completion and the speed at which the ship completes its mission.

[0043] As an example, for tasks that are not yet completed, the system retrieves the task progress from the most recent continuous analysis window from the state cache. The analysis window is determined using the median value of the historical system's complete response time. The complete response time is obtained by adding the time taken for deviation identification, model generation, command transmission, and the ship's effective maneuver response. Then, the system uses an existing least-squares fitting method to analyze the overall trend of task progress within this window, obtaining the actual increase in task progress per second. This allows the determination of the ship's task completion rate.

[0044] As an example, as the task progress continues to increase, the remaining execution time of the task is obtained by dividing the current remaining unfinished percentage by the fitted completion rate. When a task shows no significant progress (e.g., completion speed is less than or equal to a preset progress speed threshold (e.g., 0.01 / second)), select records from historical successful tasks that have the same task role and similar completion progress, and take the higher percentile value of their remaining execution time (e.g., the 80th percentile value) as the starting point. .

[0045] Based on the sum of the current time and the remaining execution time, the planned completion time corresponding to the ship coordination scheme is used to determine the direct delay of each unfinished ship task.

[0046] As an example, the original ship collaboration scheme included tasks... The planned completion time is recorded as The time is obtained from the planned completion time field in the snapshot of the plan; the current time is denoted as The estimated completion time of the task itself is the current time plus the remaining execution time, thus constructing the direct delay of the ship's mission. : In the formula, The current moment; The remaining execution time of the task itself; This represents the planned completion time of the ship coordination plan. If it is expected to be completed on time, the direct delay is 0; if it is expected to be later than originally planned, the direct delay is positive. This direct delay only represents the current time postponement of the task itself. Even if a task is significantly delayed, if it does not affect subsequent synchronization relationships or change the encirclement boundary near the fish school, it cannot be directly considered important based on this delay.

[0047] The propagation unit is used to propagate the direct delay along the collaborative relationship to obtain the estimated execution time of each ship's mission after propagation.

[0048] As an example, for tasks that have not yet started or are still subject to preconditions, the program reads its originally planned start time, all preconditions, and the normal connection interval between adjacent tasks in the original plan. The normal connection interval is obtained directly by subtracting the planned completion time of the preconditions from the planned start time of the subsequent task; if the original plan requires immediate connection, it is set to 0. A task cannot start earlier than originally planned, nor can it start earlier than all necessary preconditions are completed. The propagation flow diagrams of various ship tasks involved in the embodiments of this application are shown below. Figure 4 As shown.

[0049] As an example, let's set a task The original planned start time was The set of prerequisite tasks is Prerequisite tasks The expected completion time after dissemination is ; Prerequisite tasks in the original plan Complete the task The transition interval between normal starts. Task Estimated execution time after considering all preceding effects for: For synchronous tasks, the latest time within the group to meet the start conditions is used as the common start time; for waiting tasks, the estimated time when the task being waited for reaches the target state is used as the release time.

[0050] The completion time after propagation is recursively calculated using textual rules: completed tasks use the actual completion time; tasks currently in execution use the current time plus the remaining execution time; and tasks not yet started use the propagated start time plus the original planned execution time. By processing all nodes from beginning to end, the actual impact of a task delay on other ships' waiting, synchronization closure, and subsequent net-closing tasks can be obtained.

[0051] Therefore, a delay in B's arrival not only alters B's own timeline but could also extend A's waiting time, postpone the simultaneous closure of both sides, and delay the subsequent net-closing operation. The importance of the mission itself is thus reflected in its actual propagation range, rather than being predetermined by its name.

[0052] The reconstruction submodule is used to reconstruct the future capture boundary based on the expected execution time, and compare the future capture boundary with the planned boundary of the ship coordination scheme to obtain the dynamic escape channel of the fish school.

[0053] As an example, the projected execution times of multiple vessels are mapped onto future time slices. By interpolating the start and target points of the mission, the physical positions of each vessel at specific future moments are calculated. This allows for the determination of whether effective connections have been established between the vessels and fishing nets at these moments, thus reconstructing the future catch-up boundary affected by delays. The future catch-up boundary is then compared with the planned boundary of the vessel coordination scheme to obtain the dynamic escape channels of the fish school.

[0054] The refactoring submodule includes the step of refactoring the future capture boundary based on the expected execution time, which includes: The capture boundaries corresponding to each boundary task are obtained from the ship status data. The capture boundaries include the fixed boundaries that have been formed, the boundaries that are in formation, and the boundaries that have not yet been formed.

[0055] As an example, the system first reads the actual status of each boundary task and divides the encirclement boundary into fixed boundaries that have already formed, boundaries that are forming (boundaries that are currently forming), and boundaries that have not yet formed. Among them, boundary tasks are specific operational instructions that directly participate in forming or maintaining the encirclement, such as setting nets, closing the perimeter, and blocking the outer perimeter, and are part of the ship's tasks.

[0056] For both formed and unformed boundaries, the positions of fishing vessels and net endpoints at the predicted future time are determined according to the planned time and expected execution time of the vessel coordination scheme.

[0057] As an example, the fixed boundaries remain unchanged. Specifically, the completed net-laying edges, closing edges, and effective outer perimeter blocking edges are directly fixed and saved using the current vessel position, net endpoints, and connected net edges. Subsequent predictions and large model corrections cannot be undone. The current fishing completion status is directly reflected in the boundaries that have actually been formed.

[0058] As an example, for both forming and unformed boundaries, the system determines the portion already formed at each future prediction time based on the planned time of the ship coordination scheme and the estimated execution time after delay propagation. If task path points are saved, interpolation is performed along the path points according to the execution progress; if only the start and target points are saved, interpolation is performed from the start point to the target point (existing trajectory interpolation methods are not elaborated here). When the future prediction time is later than the task's estimated completion time, interpolation stops, and the target point is taken as the task's position at that prediction time.

[0059] At each predicted future time, the positions of each fishing boat and the endpoints of the nets within the encirclement boundary are arranged in the order of boundary connection to obtain the future encirclement boundary.

[0060] As an example, at each future predicted moment, the positions of each fishing boat and the endpoints of the nets are arranged in the order of boundary connection to obtain the future encirclement boundary.

[0061] The reconstruction submodule involves comparing the future capture boundary with the planned boundary of the ship-cooperative scheme to obtain the dynamic escape channel of the fish school. This step includes: At each future predicted time, based on the original planned time of the ship coordination scheme, the planned execution progress of each boundary task is determined, and the planned encirclement boundary is constructed.

[0062] Extract the unclosed areas between adjacent endpoints in the planned containment boundary as the original open areas.

[0063] At each future prediction time, the area between adjacent boundary endpoints in the future enclosure boundary that is not yet connected by fishing boats, nets, or effective blocking conditions is regarded as the open area after propagation.

[0064] As an example, the planned and future capture boundaries are constructed in a similar way, treating the unclosed areas between adjacent endpoints in the planned capture boundary as the original open areas. When the adjacent boundary endpoints of the future enclosure boundary are not yet connected by fishing vessels, nets, or effective barriers, that area is designated as the post-propagation open area. .in, Forecast time after the current time; superscript Indicates the original planned status; superscript This indicates the state after considering delayed propagation.

[0065] The dynamic escape channels of the fish swarm are obtained by performing difference calculations between the open region after propagation and the original open region.

[0066] As an example, since the missing connection parts obtained from the above difference calculation are geometrically represented as one-dimensional line segments or curves, they cannot be directly used for subsequent area calculations. This step expands the propagated open area and the original open area outwards (i.e., away from the center of the net) along the boundary normal direction by a preset distance threshold (which can be 50 to 150 meters, for example, 100 meters. This value is determined based on the typical swimming escape distance of the fish within the predicted time). This transforms them into two-dimensional polygonal regions with two-dimensional area attributes. Then, difference calculation is performed on these two two-dimensional polygonal regions to obtain the polygonal channel cross-section region, which is used as the final output dynamic escape channel. During the netting operation, normal openings may be retained; therefore, not all open areas can be attributed to task delay. The system only retains the portion added by the delayed scheme relative to the original scheme, the dynamic escape channel. The calculation method is expressed as follows: In the formula, For the future The dynamic escape route created by the additional task delay; The symbol "\" represents the difference between regions. In set theory, it represents the difference between sets. In addition to the net openings that normally exist in the ship coordination scheme, only the open areas that are additionally formed or extended due to the propagation of task delays are extracted. The subsequent analysis does not focus on whether there are gaps in the net, but rather on what escape channels the task delay has created relative to the original scheme.

[0067] The prediction module 103 is used to predict the future reachable area of ​​the fish based on the fish state data, and to calculate the escape overlap between the future reachable area of ​​the fish and the dynamic escape channel.

[0068] As an example, the escape overlap refers to the cumulative total of the proportion of the space area that the fish can reach in the future and enters the dynamic escape channel in the future time period from the effective date of the revised plan to the end of this phase. It is used to quantitatively determine whether the task delay has substantially caused the risk of the fish swarm escaping.

[0069] As an example, the range that the fish may swim and cover at various future times (future reachable area) is estimated by combining the current movement speed of the fish school; then, within the effective prediction time period, the polygonal overlap area of ​​this area and the dynamic escape channel is calculated and divided by the total reachable area to eliminate the influence of the fish school size. Finally, the ratio at each time point is integrated or accumulated along the time axis to determine the escape overlap amount.

[0070] The prediction module 103 is specifically used for: By using a linear fitting method to process the fish swarm state data, the predicted center position of the fish swarm at any future time can be obtained.

[0071] As an example, the most recent consecutive fish swarm center positions are extracted from the fish swarm state data, and the predicted center positions for each future prediction time are obtained using existing linear fitting methods. These predicted center positions only determine the approximate center of the future fish swarm region and do not assume the fish swarm is always moving at a uniform linear velocity.

[0072] Extract the fish swarm trajectories that are the same as those in the current operation phase within the historical time period, and construct the historical motion extension region based on the offset point of the fish swarm trajectory relative to the prediction center corresponding to the historical time.

[0073] As an example, to supplement the fish school turning, speed changes, and outer edge expansion, fish school trajectories identical to those in the current operation phase are selected from historical successful operation tasks. For the same prediction duration, the offset points of the actual outer edge of the fish school relative to the corresponding prediction center at the historical time are recorded. The main offset points are retained and an outer region is constructed to form the historical motion expansion region. Where 'b' comes from the historical job stage field. The predicted duration is shown below; this area is formed by the outer edge offset of historical real fish swarms, without setting turning weights or diffusion coefficients. A schematic diagram of the prediction process for the future reachable area of ​​the fish swarm is shown below. Figure 5 As shown.

[0074] By shifting the historical movement extension area to the prediction center location, the future reachable area of ​​the fish school can be determined.

[0075] As an example, by shifting the historical movement extension area to the current prediction center location, we can obtain the future reachable area of ​​the fish swarm. : In the formula, The area that the fish school might reach at a future time (u); The predicted center location is obtained based on the current continuous fish school center. This represents the movement extension area with the same historical stage and prediction duration; ⊕ indicates the area translation and overlay. Only when the fish-reachable area enters the delayed new channel is the task delay considered potentially affecting the capture result.

[0076] The escape overlap is calculated based on the overlap area between the future reachable area and the dynamic escape channel.

[0077] As an example, the time required from triggering the correction to the new instruction causing the fishing vessel to take effective action is obtained from historical operation logs, including model generation, instruction transmission, and vessel response time. The historical median value is then used as the correction's effective time. Simultaneously, based on the current operational phase and established boundary combinations, the median remaining time until the end of this phase is obtained from historical successful operations. Overlap calculations only cover the future range from "when the correction plan can begin to take effect" to "when the current encirclement phase is expected to end." Risks that have already occurred before the new plan takes effect cannot be eliminated through replanning, and predictions exceeding the remaining operational time are meaningless.

[0078] As an example, multiple prediction time points are established within this future range according to the state sampling interval. For each time point, the overlap area between the fish swarm's reachable area and the delayed newly added channel is calculated using an existing polygon intersection procedure, divided by the total area of ​​the dynamic escape channels, and then accumulated over time intervals to obtain the escape overlap amount. : In the formula, L represents the number of effective prediction times; For the first One predicted moment; This is the area that the fish can reach in the future; For dynamic escape routes; Indicates the first At each predicted time point, the total area of ​​the dynamic escape channels generated by all currently affected and related tasks in the two-dimensional plane. Indicates the area of ​​the region; This represents the system sampling interval. It also reflects whether overlap occurs, the proportion of the overlap to the total area of ​​the dynamic escape channel, and its duration. Delays of different vessels, cooperative dependencies, established boundaries, and fish movement are all incorporated into the same result. Specifically, a schematic diagram illustrating the spatial overlap development process between the future reachable area of ​​the fish and the escape channel is shown below. Figure 6 As shown.

[0079] The second determining module 104 is used to determine the delay contribution of each ship mission based on the decrease in the escape overlap after eliminating the delay of a single mission when the escape overlap exceeds the preset historical normal range.

[0080] As an example, using the current operational phase and established boundary combinations as matching conditions, control states are extracted from historical operations that ultimately achieved successful captures and for which no scheme modifications were made at the corresponding time. The historical escape overlap is then calculated using the same method. The historical escape overlap is sorted, and the higher quantile value determined by the validation set (e.g., the 90th percentile) is taken as the normal upper bound. If the current overlap does not exceed the normal upper bound, the original scheme continues to be executed; only when it continuously exceeds the normal upper bound within a continuous analysis window is task contribution attribution performed. When the above conditions are met, the delay contribution corresponding to each vessel's task is calculated.

[0081] The second determining module 104 is specifically used for: For any given ship mission, determine the first escape overlap amount resulting from eliminating the direct delay of the current ship mission.

[0082] The delay contribution of each ship's mission is determined based on the decrease between the escape overlap amount and the first escape overlap amount.

[0083] As an example, the importance of the tasks themselves is not assigned a fixed level. For each task q with a direct latency greater than 0, the program temporarily restores the direct latency of that task to 0, while keeping the latencies of other tasks and their real-time states unchanged, and then re-executes latency propagation, boundary reconstruction, channel extraction, and core overlap calculation. The first escape overlap amount after eliminating only the latency of task q is denoted as... The contribution of delay to ship mission q The calculation method is as follows: In the formula, The overlap amount when all actual direct delays coexist; This represents the overlap amount after eliminating only the direct delay of task q. The Max function represents the maximum value function. If eliminating this task advances the synchronization of tasks, the boundary forms earlier, and the overlap decreases significantly, then its contribution is large; if other ships have already formed alternative coverage or this task does not affect the area that the fish school may enter, then the delay contribution is close to 0.

[0084] The generation module 105 is used to determine the set of main tasks to be corrected based on the contribution of delay, input the set of main tasks into the local collaborative big model, generate a corrected collaborative scheme, and issue it for execution.

[0085] As an example, based on the contribution of delay, the main task set of the ship tasks that need to be corrected is determined. Then, the main task set is input into the local collaborative big model to generate a correction collaboration scheme and issue it for execution. This avoids reconstructing all ship tasks and only corrects some ship tasks, thereby reducing the workload and making the correction results more accurate.

[0086] As an example, the main task set represents the smallest subset of tasks that substantially contribute to the deterioration of the final escape overlap and need to be input into the large model for replanning. It includes the source delayed tasks that cause high escape risk and their directly constrained downstream tasks.

[0087] Specifically, the generation module 105 is used for: Ship missions with direct delays are ranked in descending order of their contribution to delay, and the ranking results are obtained.

[0088] Based on the sorting results, the corresponding ship missions and downstream missions affected by the delay of the current ship mission in the collaborative relationship are added to the set to be corrected.

[0089] After each operation that adds a ship mission or downstream mission to the set to be corrected, the simulation eliminates the delay of all missions in the set to be corrected and recalculates the current simulated escape overlap.

[0090] If the simulated escape overlap decreases to the preset historical normal range, then stop adding tasks and determine the current set of tasks to be corrected as the main set of tasks to be corrected.

[0091] As an example, the system sorts delayed tasks from largest to smallest contribution, starting with the task with the largest contribution, and adds that task, along with the synchronization, waiting, or subsequent tasks directly delayed by it, to the set to be corrected. For each new task, the system simulates eliminating the selected task's delay and recalculates the simulated escape overlap. The simulation stops when the escape overlap returns to a preset historical normal range, thus obtaining the main task set that can explain the current abnormal overlap, avoiding having to re-plan all minor delays for vessels in the large model. The preset historical normal range is determined by the overlap distribution under successful, uncorrected operations in the current phase, and its value is typically set so that its upper limit does not exceed a certain high quantile value (e.g., the 90th quantile) of the historical overlap distribution.

[0092] If the simulated escape overlap still does not decrease to the preset historical normal range after adding all tasks with direct delays to the set to be corrected, then the addition of tasks will be forcibly stopped, and all added tasks will be used as the main set of tasks to be corrected.

[0093] The generation module 105 is also used for: The main task set, the current ship collaboration scheme, and ship status data are used as the model input information, and local correction constraints are set.

[0094] Input the model input information and local correction constraints into the local collaborative large model to generate at least one set of candidate correction schemes.

[0095] For each set of candidate correction schemes, the corresponding future capture boundary is re-deduced and the candidate escape overlap is calculated. Invalid candidate correction schemes that do not meet the local correction constraints or whose candidate escape overlap is greater than or equal to the current ship coordination scheme are eliminated.

[0096] From the remaining valid candidate correction schemes, the scheme with the smallest candidate escape overlap is selected as the final correction coordination scheme, and the correction coordination scheme is split and distributed to the corresponding fishing vessels for execution.

[0097] As an example, the input to the local large model includes the current ship coordination scheme, the most recent consecutive fish schools and the status of multiple ships, task relationships, direct delays and propagation results, fixed encirclement boundaries, delayed new channels, overlapping areas accessible to fish schools, current escape overlap, historical normal upper bound, the set of main tasks, and the contribution of each task. It also explicitly specifies which tasks are allowed and prohibited from modification. Local correction constraints include: the large model only allows adjustments to the main delayed tasks and their directly dependent tasks, such as shortening target distances, adjusting numerical target positions, arranging temporary replacements by neighboring ships, changing synchronization closure conditions, or extending still effective outer barriers; it does not allow the cancellation of completed tasks, disruption of fixed boundaries, modification of irrelevant tasks, or output of instructions lacking target positions, task times, or inter-ship relationships.

[0098] It should be noted that after the large model generates one or more sets of candidate correction schemes, the system does not send them directly. Instead, it re-executes the same analysis chain for each candidate correction scheme: establishing candidate task relationships, updating estimated times, reconstructing future boundaries, extracting delay channels for candidate correction schemes, and recalculating candidate escape overlap. Simultaneously, it checks whether the candidate correction schemes retain completed tasks and fixed boundaries, whether they only modify tasks within the allowed range, and whether the corresponding fishing vessels can achieve the new objective within the remaining operating time. In other words, it determines whether the candidate correction schemes satisfy local correction constraints.

[0099] Invalid candidate correction schemes that do not meet the local correction constraints or whose candidate escape overlap is greater than or equal to the current ship coordination scheme are removed. Among the valid candidate correction schemes that meet the execution constraints, the scheme with the lowest overlap is selected and sent to each ship. A candidate correction scheme that cannot reduce the escape overlap of the original scheme, or that only achieves a lower result by destroying the established boundary, is deemed invalid. After the scheme is sent, the new version is saved, and the status, task delay, future boundary, and core overlap are updated every 5 seconds. Once the escape overlap returns to the historical normal range, the current scheme is maintained.

[0100] This application provides a collaborative system for adaptive fishing operations at sea based on a large-scale intelligent model. In this system, the dynamic escape channels of fish schools are determined through ship status data and the collaborative relationships between ship tasks. Then, the future reachable area of ​​the fish school is predicted based on the fish school status data, and the overlap between the future reachable area and the dynamic escape channels is calculated. This accurately reflects whether deviations pose a risk of fish escape, avoiding oversensitivity to non-threatening deviations and invalid alarms. When the overlap exceeds a preset historical normal range, the delay contribution of each ship task is determined based on the decrease in overlap after eliminating the delay of a single task. Based on the delay contribution, the main task set to be corrected is determined, and this set is input into the local collaborative large-scale model to generate a corrected collaborative scheme and issue it for execution. This system accurately identifies the key tasks that truly cause large gaps, limiting the large-scale model to only locally correcting high-contribution tasks and their controlled downstream components. It avoids minor changes causing the large-scale model to globally overturn and rearrange the entire fleet's tasks, significantly reducing unnecessary ship turning and command changes, and improving the stability and success rate of multi-ship collaborative operations.

[0101] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0102] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0103] The above are merely preferred embodiments of this application and do not limit the scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the scope of protection of this application.

Claims

1. A fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent, characterized in that, The system includes: The acquisition module is used to acquire the current ship coordination scheme, fish swarm status data, and ship status data. The ship coordination scheme includes multiple ship tasks and the coordination relationships between the ship tasks. The first determining module is used to determine the dynamic escape channel of the fish school based on the ship status data and the collaborative correlation. The prediction module is used to predict the future reachable area of ​​the fish school based on the fish school status data, and to calculate the escape overlap between the future reachable area of ​​the fish school and the dynamic escape channel. The second determining module is used to determine the delay contribution of each ship mission based on the decrease in the escape overlap amount after eliminating the delay of a single mission when the escape overlap amount exceeds the preset historical normal range. The generation module is used to determine the set of main tasks to be corrected based on the delay contribution, input the set of main tasks into the local collaborative large model, generate a corrected collaborative scheme, and issue it for execution.

2. The fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent as described in claim 1, characterized in that, The first determining module includes: The first determining submodule is used to determine the estimated execution time of each ship's mission based on the ship status data and the collaborative relationship. The reconstruction submodule is used to reconstruct the future capture boundary based on the expected execution time, and compare the future capture boundary with the planned boundary of the ship coordination scheme to obtain the dynamic escape channel of the fish school.

3. The fishing vessel maritime adaptive operation collaborative system based on a large model intelligent agent as described in claim 2, characterized in that, The first determining submodule includes: The determining unit is used to determine the direct delay of each unfinished ship task based on the ship status data; The propagation unit is used to propagate the direct delay along the collaborative association to obtain the estimated execution time of each ship's mission after propagation.

4. The fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent as described in claim 3, characterized in that, The determining unit is specifically used for: Extract the mission start position, target relative position, and current relative position of any ship mission from the ship status data; Based on the task start position, the target relative position, and the current relative position, determine the task completion progress; Based on the progress of the task completion and the speed at which the ship completes the task, the remaining execution time of the ship's task is determined; Based on the sum of the current time and the remaining execution time, the planned completion time corresponding to the ship coordination scheme is used to determine the direct delay of each unfinished ship task.

5. The fishing vessel maritime adaptive operation collaborative system based on a large model intelligent agent as described in claim 2, characterized in that, The reconstruction submodule is specifically used for: The capture boundaries corresponding to each boundary task are obtained from the ship status data. The capture boundaries include fixed boundaries that have been formed, boundaries that are in formation, and boundaries that have not yet been formed. For the formed and unformed boundaries, the positions of the fishing boats and the ends of the nets at the predicted future time are determined according to the planned time of the ship coordination scheme and the expected execution time, respectively. At each future predicted time, the positions of each fishing boat and the endpoints of the nets in the encirclement boundary are arranged in the order of boundary connection to obtain the future encirclement boundary.

6. The fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent as described in claim 5, characterized in that, The reconstruction submodule is also used for: At each future predicted time, based on the original planned time of the ship coordination scheme, the planned execution progress of each boundary task is determined, and the planned encirclement boundary is constructed; Extract the unclosed area between adjacent endpoints in the planned enclosure boundary as the original open area; At each future prediction time, the portion of the future enclosure boundary that is not yet connected by fishing boats, nets, or effective blocking is considered as the open area after propagation. The dynamic escape channel of the fish swarm is obtained by performing a difference calculation between the propagated open region and the original open region.

7. The fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent as described in claim 1, characterized in that, The prediction module is specifically used for: The fish swarm state data is processed by a linear fitting method to obtain the predicted center position of the fish swarm at any future time. Extract the fish swarm trajectory that is the same as the current operation stage within the historical time period, and construct the historical motion extension area based on the offset point of the fish swarm trajectory relative to the prediction center corresponding to the historical time. The historical movement extension area is shifted to the prediction center location to determine the future reachable area of ​​the fish school; The escape overlap amount is calculated based on the overlap area between the future reachable region and the dynamic escape channel.

8. The fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent as described in claim 1, characterized in that, The second determining module is specifically used for: For any given ship mission, determine the first escape overlap amount resulting from eliminating the direct delay of the current ship mission; Based on the decrease in the escape overlap amount and the first escape overlap amount, the delay contribution of each ship mission is determined.

9. The fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent as described in claim 1, characterized in that, The generation module is specifically used for: The ship missions with direct delays are sorted in descending order of their delay contribution, and the sorting results are obtained. Based on the sorting results, the corresponding ship missions and downstream missions affected by the delay of the current ship mission in the collaborative relationship are added to the set to be corrected in sequence; After each operation of adding a ship mission or downstream mission to the set to be corrected, the simulation eliminates the delay of all missions in the set to be corrected and recalculates the current simulated escape overlap. If the simulated escape overlap decreases to a preset historical normal range, then stop adding tasks and determine the current set of tasks to be corrected as the main set of tasks to be corrected. If the simulated escape overlap still does not decrease to the preset historical normal range after adding all tasks with direct delays to the set to be corrected, then the addition of tasks will be forcibly stopped, and all added tasks will be used as the main set of tasks to be corrected.

10. The fishing vessel adaptive maritime operation collaborative system based on a large model intelligent agent as described in claim 1, characterized in that, The generation module is further configured to: The main task set, the current ship collaboration scheme, and ship status data are used to construct the model input information, and local correction constraints are set. Input the model input information and the local correction constraints into the local collaborative large model to generate at least one set of candidate correction schemes; For each set of candidate correction schemes, the corresponding future capture boundary is re-deduced and the candidate escape overlap is calculated. Invalid candidate correction schemes that do not meet the local correction constraint conditions or whose candidate escape overlap is greater than or equal to the escape overlap of the current ship coordination scheme are eliminated. From the remaining valid candidate correction schemes, the scheme with the smallest candidate escape overlap is selected as the final correction coordination scheme, and the correction coordination scheme is split and distributed to the corresponding fishing vessels for execution.

Images