An open-ocean aquaculture system based on a large marine fishery model and digital twinning and a wind and wave resistant intelligent operation and maintenance method

Abstract

The application discloses a kind of based on ocean fishery big model and digital twinning deep sea aquaculture system and anti storm intelligent operation and maintenance method, belong to the field of ocean, including: multi-source heterogeneous information high-fidelity acquisition module, for collecting environment and working condition data;Knowledge graph and big model cognitive module, through retrieval enhancement generation realizes semantic understanding and causal reasoning;Digital twinning simulation verification module, establishes "environment-structure-operation-biology" coupling model and carries out virtual deduction and risk assessment;Multi-agent collaborative decision-making module, through interaction, cognition, coordination and four types of intelligent agents carry out task arrangement and resource scheduling;Execution and feedback closed loop module, realizes instruction issuing and online correction.The application solves the problems of data fusion difficulty, decision lag, weak collaborative control and lack of physical simulation verification in complex sea conditions, can identify extreme sea condition risk in advance and generate verified risk avoidance strategy, significantly improve the safety of deep sea aquaculture facilities.

Description

Technical Field

[0001] This invention belongs to the marine field, and in particular relates to a deep-sea aquaculture system and a wind and wave resistant intelligent operation and maintenance method based on a large marine fisheries model and digital twin. Background Technology

[0002] The marine aquaculture industry is undergoing a leapfrog development, expanding from nearshore harbors to deep-sea areas and from traditional cages to large-scale engineering facilities. The Yellow and Bohai Seas, as important marine aquaculture bases in northern my country, are significantly affected by the monsoon climate, experiencing frequent cold waves in winter and high risks of typhoons in summer. Effective wave heights can reach 6 meters once every 50 years, with high water turbidity and severe organism attachment. To cope with these harsh sea conditions, large truss-type cages and aquaculture vessels such as the "Deep Blue 1" have emerged in recent years. While their resistance to wind and waves has improved, the system complexity and maintenance difficulty have increased exponentially. Modern large-scale aquaculture facilities are equipped with a massive number of sensors and monitoring equipment, generating multi-source heterogeneous data. Existing systems are mostly based on simple threshold alarms and cannot understand the correlations between data across modalities. Meanwhile, the explosion of large language model technology has provided new ideas for semantic understanding and knowledge reasoning, but general-purpose large models lack in-depth knowledge of vertical fields such as marine engineering structures and aquatic animal behavior. Direct application can easily lead to "illusions" and is difficult to support accurate decision-making under high sea state conditions.

[0003] Currently, deep-sea aquaculture operation and maintenance still faces several key technological bottlenecks that urgently need to be overcome. Firstly, capturing operational windows under extreme sea conditions is difficult. Traditional methods based on experience or single forecast products struggle to accurately predict "operable time periods, accessible operational areas, and tolerable operational intensity," and lack a window assessment mechanism that links "environment, structure, and operational constraints." Secondly, the utilization rate of unstructured data is low. Large amounts of data, such as video surveillance, underwater images, and operational logs, lack unified semantic expression and interpretable reasoning capabilities. Fault diagnosis often remains at the rule threshold level, failing to form a reusable knowledge loop that integrates historical fault chains, environmental causes, and response measures. Thirdly, the collaborative control capabilities of complex systems are weak. The lack of unified intelligent scheduling among multiple pieces of equipment makes it difficult to achieve a closed-loop process of "inspection-diagnosis-dispatch-verification," and under high sea state conditions, it is impossible to dynamically adjust task scheduling based on risk and priority. Fourth, the lack of physical simulation and deduction capabilities for wind and wave resistant operation and maintenance makes it difficult to conduct "virtual trial runs" of candidate strategies before implementation, and makes it impossible to assess in advance the impact of strategies on structural safety margins and mooring over-limit probability, which fundamentally restricts the reliability and engineering availability of intelligent operation and maintenance. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a deep-sea aquaculture system based on a large-scale marine fisheries model and digital twin, comprising: A high-fidelity acquisition module for multi-source heterogeneous information is used to collect structured time-series data and unstructured multimedia data of the aquaculture facility and its surrounding sea area. The knowledge graph and large model cognition module is used to construct a knowledge graph in the fisheries field and connect to the operation and maintenance knowledge base through retrieval enhancement generation method to provide semantic understanding and causal reasoning capabilities for operation and maintenance tasks and generate a set of candidate operation and maintenance strategies. The digital twin simulation verification module is used to establish a coupled digital twin that includes ocean wind and wave flow field, hydrodynamic load, structural response, and fish physiological stress and behavioral response, and to perform virtual simulation and risk assessment on the strategies in the candidate operation and maintenance strategy set, and screen out executable strategies that meet safety constraints. The multi-agent collaborative decision-making module is used to decompose the executable strategy into tasks and schedule resources through the interaction and collaboration between multiple agents, and generate specific action sequences. The execution and feedback closed-loop module is used to distribute the action sequence to the on-site execution objects, perform online evaluation based on the execution feedback data, and feed the evaluation results back to the knowledge graph and big model cognition module to achieve knowledge updates.

[0005] Optionally, the multi-source heterogeneous information high-fidelity acquisition module includes a sea state prediction fusion unit, which is used to fuse on-site observation data with sea state forecast products and output wind, wave and current prediction sequences within a preset future time window.

[0006] Optionally, the multi-source heterogeneous information high-fidelity acquisition module includes a turbid water enhancement unit, which is used to perform defogging and detail enhancement processing on the acquired underwater images or videos using image enhancement algorithms, so as to improve the detection capability of underwater structures and biological anomalies.

[0007] Optionally, the knowledge graph and large model cognition module is specifically used to generate the candidate operation and maintenance strategy set by calling the operation and maintenance knowledge base, which includes operation and maintenance manuals, historical fault logs and expert experience bases, based on a general large language model and through a retrieval-enhanced generation method.

[0008] Optionally, the digital twin simulation verification module is specifically used to calculate the structural safety margin and mooring tension exceedance probability of the candidate operation and maintenance strategy under a given sea state, and use the calculation results as safety constraints for the action sequence generated by the multi-agent collaborative decision-making module.

[0009] Optionally, the multi-agent collaborative decision-making module includes an interactive agent, a cognitive agent, a coordinating agent, and a function execution agent; wherein, the interactive agent is used to receive instructions and generate operation and maintenance task sheets, the cognitive agent is used to interpret sea state evolution and operational risks, the coordinating agent is used to perform task orchestration and resource scheduling based on the output of the cognitive agent, and the function execution agent is used to generate control parameters and action sequences for specific operational tasks based on the scheduling of the coordinating agent.

[0010] Optionally, the execution and feedback closed-loop module includes an online evaluation unit, which is used to calculate the job completion degree and risk convergence based on the execution feedback data, and write the evaluation results back to the knowledge graph and large model cognition module as feedback data; The execution and feedback closed-loop module also includes an online learning unit, which is used to associate and label the execution feedback data with the corresponding work order results, and to update the retrieval library or model parameters of the knowledge graph and the large model cognition module in an incremental manner.

[0011] To address the aforementioned technical problems, this invention also provides a method for intelligent operation and maintenance against wind and waves based on a large-scale marine fisheries model and digital twin, comprising: Collect multimodal data and perform time alignment and quality management; Based on the processed data, the operation and maintenance knowledge base is invoked and semantic understanding is performed through a retrieval enhancement generation method to generate candidate operation and maintenance strategies. The candidate operation and maintenance strategies are virtually simulated and risk assessed in the digital twin to select executable strategies; Based on the executable strategy, task orchestration and resource scheduling are performed through multi-agent collaboration to generate action sequences; The action sequence is sent to the on-site execution objects, and the subsequent strategy generation process is modified based on the execution feedback.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs a closed-loop operation and maintenance system encompassing "sensing-knowledge-computation-control," achieving high-fidelity acquisition and cross-modal fusion of multi-source heterogeneous data. This significantly enhances the detection capabilities for features such as netting damage and structural anomalies in highly turbid waters. Leveraging a domain-enhanced large-scale model and knowledge graph, the system can perform causal reasoning and semantic understanding of historical fault cases and real-time operating conditions, generating interpretable fault diagnosis and handling solutions. This addresses the problem of traditional rule-based thresholds being unable to reuse experiential knowledge. Through a hydrodynamic-structural-biological coupled digital twin, this invention can perform high-fidelity simulations and risk assessments of candidate operation and maintenance strategies in virtual space. This allows for early identification of structural safety risks and mooring overrun probabilities under extreme sea conditions, screening for validated and executable strategies, effectively capturing "micro-windows" in severe sea conditions, and improving operational window utilization. Based on a multi-agent collaborative decision-making architecture, the system can achieve task orchestration, resource scheduling, and conflict resolution among multiple pieces of equipment, connecting "inspection-diagnosis-dispatch-verification" into a closed-loop process, reducing the frequency of manual intervention and operational risks. Ultimately, this invention achieves a synergistic improvement in the safety, operational efficiency, and intelligence level of deep-sea aquaculture facilities under high sea state conditions. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the overall architecture of the system according to an embodiment of the present invention; Figure 2 This is an internal interaction flowchart of the multi-agent collaborative decision-making module in an embodiment of the present invention; Figure 3 This is a logic diagram of dynamic simulation and risk assessment of net cages based on digital twins according to an embodiment of the present invention. Figure 4 This is a sequence diagram of the integrated collaborative operation of "catching, grading and transferring" under typical sea conditions according to an embodiment of the present invention; Figure 5 This is a comparison diagram of the underwater image enhancement and netting damage detection algorithm flow and effect in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a deep-sea aquaculture system based on a large-scale marine fisheries model and digital twin, including: This embodiment provides an intelligent operation and maintenance system for deep-sea aquaculture systems based on a large-scale marine fisheries model and digital twin. This system is used for state perception, risk assessment, collaborative decision-making, and closed-loop control of large-scale deep-sea aquaculture facilities under high sea state and wave disturbances. As a specific implementation, the system includes a high-fidelity acquisition module for multi-source heterogeneous information, a knowledge graph and large-scale model cognition module, a digital twin simulation verification module, a multi-agent collaborative decision-making module, and an execution and feedback closed-loop module. These modules work collaboratively to form a complete "sensing-knowledge-computation-control" closed-loop operation and maintenance system.

[0019] Specifically, the multi-source heterogeneous information high-fidelity acquisition module is deployed on the aquaculture facility itself and its surrounding sea area to collect structured time-series data and unstructured multimedia data. The structured time-series data includes at least environmental and operational parameters such as stress and strain, vibration acceleration, mooring tension, heave / roll / pitch attitude, water depth and current velocity, significant wave height, wind speed and direction, temperature, salinity, depth, dissolved oxygen, pH, turbidity, and chlorophyll content at key nodes of the cage / truss. The unstructured multimedia data includes at least underwater optical video, underwater acoustic imaging or forward-looking sonar images, sea surface video, and meteorological satellite or sea state forecast products. This module further includes a time synchronization and data governance unit for timestamp alignment, outlier removal, missing data completion, and quality labeling of the multi-source data. The processed data is then output to edge computing nodes to provide high-quality input for subsequent analysis.

[0020] Furthermore, the knowledge graph and large-scale model cognition module is used to construct a knowledge graph for the fisheries sector and provide knowledge-enhanced large-scale model reasoning capabilities. The fisheries sector knowledge graph includes at least facility structural components, failure modes, operation and maintenance procedures, environmental events, fish physiological stress indicators, and their causal relationships. Based on a general large-scale language model, this module accesses operation and maintenance manuals, historical fault logs, sensor data tags, expert experience bases, and operation records through a retrieval-enhanced generation method. This enables semantic understanding of operation and maintenance tasks, fault diagnosis, causal reasoning, and generation of disposal solutions, and outputs a set of candidate operation and maintenance strategies.

[0021] The digital twin simulation verification module is used to establish a digital twin of the aquaculture facility's "environment-structure-operation-biology" coupling, and to perform virtual simulations and risk assessments of candidate operation and maintenance strategies. The digital twin includes at least an ocean wind and wave flow field model, a hydrodynamic load model, a structural response model, and a fish physiological stress and behavioral response model. This module calculates the structural safety probability, mooring overrun risk, operational accessibility, equipment load margin, and fish stress risk index of candidate operation and maintenance strategies under given sea conditions, and selects executable strategies that meet safety constraints.

[0022] Based on this, the multi-agent collaborative decision-making module is used to perform multi-role collaborative optimization and task decomposition of executable strategies verified through digital twins. This module includes at least an interactive agent, a cognitive agent, a coordinating agent, and a function execution agent. Specifically, the interactive agent receives personnel instructions and generates traceable maintenance task sheets; the cognitive agent interprets and alerts on sea state evolution, fault chains, and operational risks; the coordinating agent performs task orchestration, resource scheduling, and priority ranking under conditions of multiple equipment, multiple operations, and multiple constraints; and the function execution agent generates control parameters and action sequences for specific tasks such as feeding, inspection, cleaning, anti-attachment, capture, levitation, mooring adjustment, and hazard avoidance.

[0023] Finally, the execution and feedback closed-loop module is used to distribute the action sequence output by the multi-agent collaborative decision-making module to the on-site execution objects and form a closed-loop adaptive correction. The on-site execution objects include at least one of the following: feeding and oxygenation equipment, net cleaning and repair equipment, harvesting and lifting equipment, mooring and tensioning adjustment equipment, and unmanned surface vessels, drones, and underwater robots. This module further includes an online evaluation unit, used to calculate the operation completion rate, risk convergence, and fish stress change trends based on the execution feedback data, and write the evaluation results back to the knowledge graph and large model cognition module to achieve model parameter updates and incremental learning of the experience base. Through the organic cooperation of the above modules, this system can realize intelligent operation and maintenance of deep-sea aquaculture facilities under high sea state conditions.

[0024] In a preferred embodiment, the multi-source heterogeneous information acquisition module further includes a sea state prediction fusion unit. This unit is used to fuse on-site observation data with sea state forecast products and output a predicted sequence of significant wave height, current velocity, wind speed, and wind direction within a preset future time window. In this embodiment, the preset time window can be set to the next 24 to 72 hours to provide forward-looking sea state input for subsequent risk assessment and decision-making.

[0025] Furthermore, considering the high turbidity of waters in the Yellow Sea and Bohai Sea, the multi-source heterogeneous information acquisition module also includes a turbidity enhancement unit. This unit is used to perform dehazing and detail enhancement processing on the acquired underwater images or videos to improve the detectability of net damage, abnormal attached organisms, abnormal structural connections, or fish behavior characteristics. As an optional implementation, the turbidity enhancement unit can use a combination model of dark channel prior and generative adversarial networks, or an equivalent model, to achieve image enhancement.

[0026] In the digital twin simulation verification module, the output includes at least one or more of the following: stress margin at key nodes, structural displacement or deformation margin, and probability of exceeding mooring tension limits. These outputs serve as safety constraints for subsequent decisions and are passed to the multi-agent collaborative decision-making module to guide the generation of action sequences, ensuring that the generated operational plans meet structural safety requirements.

[0027] Furthermore, the execution and feedback closed-loop module also includes an online learning unit. This unit is used to associate and annotate execution feedback data, work order results, and manual review conclusions, and to incrementally update the retrieval database and / or model parameters of the knowledge graph and the large model cognition module. Through this online learning mechanism, the system can continuously accumulate operational experience, achieving dynamic updates to the knowledge base and continuous optimization of model performance.

[0028] Example 2 This embodiment provides a method for intelligent operation and maintenance against wind and waves based on a large marine fisheries model and digital twin, including: This system first establishes a comprehensive three-dimensional perception network covering above-water, surface-water, and underwater environments to achieve accurate perception of multimodal and multi-element environmental and operational conditions.

[0029] In terms of the deployment of the sensing network, this embodiment deploys fiber optic stress sensors at key locations such as truss nodes and mooring lines to monitor the stress state of the structure in real time; it also deploys wave-compensated acoustic Doppler current profilers (ADCP) to monitor the flow field profile; and deploys multi-parameter water quality sensors to collect key aquaculture environment parameters such as dissolved oxygen (DO), pH value, and temperature.

[0030] To achieve advanced sea state forecasting, this system integrates meteorological satellite data and employs a Long Short-Term Memory (LSTM) network to perform refined predictions of the wind, wave, and current fields for the next 72 hours. During the forecasting process, the system pays particular attention to extreme sea state indicators at the 50-year return period level (such as wind speed of 37 m / s and wave height of 6 m), thereby providing sufficient response time for typhoon preparedness decisions.

[0031] In response to the highly turbid waters of the Yellow and Bohai Seas, this embodiment integrates an improved image enhancement algorithm that combines Dark Channel Prior (DCP) with multi-scale Retinex technology. This algorithm effectively removes fogging effects and color casts from underwater videos, significantly improving the recognition rate of key features such as damaged netting and biofouling by unmanned underwater vehicles (UUVs).

[0032] In the construction and training of the large-scale fisheries operation and maintenance model, this embodiment uses a general large-scale model as a foundation and constructs a dedicated large-scale fisheries operation and maintenance model through the following methods. First, corpus injection was performed, collecting and organizing over 100,000 maintenance manuals, fault code tables, historical operation and maintenance logs, standard operating procedures (SOPs), and historical meteorological and hydrological data from the Yellow and Bohai Seas. Second, natural language processing technology was used to extract entities and relationships, constructing a fisheries knowledge graph (Fishery-KG) containing "equipment-component-fault phenomenon-root cause-solution". Then, instruction tuning and human feedback-based reinforcement learning (RLHF) techniques were used to fine-tune and align the model, enabling it to accurately understand technical terms such as "excessive cage tilt" and "abnormal anchor chain tension" and generate decision-making schemes in accordance with safety regulations. Finally, a multimodal encoder was used to map the time-series data collected by sensors into vectors and align them with text vectors, thereby enabling the large model to achieve cross-modal understanding of log text and sensor curves, improving the accuracy and comprehensiveness of fault diagnosis.

[0033] In terms of simulation verification and risk assessment based on digital twins, this system first verifies the system in a digital twin space before performing any high-risk operations. Specifically, based on the coupling of OrcaFlex and OpenFOAM, a refined fluid-structure interaction (FSI) model of the net cage is established. This model specifically considers the hydroelastic effect, namely the deformation of the flexible net under strong current and its reaction to the flow field, thereby ensuring the simulation accuracy in the Yellow and Bohai Seas' fast-flowing environment. Simultaneously, an individual-based fish school behavior model is introduced to simulate the aggregation and escape behavior of fish during feeding and harvesting, calculate crowding and hypoxia risks, and quantitatively assess the impact of operational plans on biological welfare. Through these models, the system can simulate extreme conditions such as typhoons and cold waves in the digital twin, deduce the impact of different ballast water adjustment schemes and net cage submersion depths on the structural safety factor, and thus select the optimal risk avoidance strategy.

[0034] In terms of the multi-agent collaborative decision-making and execution architecture, this embodiment designs a collaborative architecture comprising four types of agents: interactive, cognitive, coordinating, and functional. The interactive agent acts as the human-machine interface, parsing the natural language instructions from maintenance personnel (such as "prepare for typhoon evacuation") and transforming them into structured task objectives. The cognitive agent, acting as the system's "brain," invokes large models and knowledge graphs to generate candidate diagnostic hypotheses and treatment plans. The coordinating agent, acting as the "manager," is responsible for task decomposition and resource scheduling, such as assigning which unmanned surface vessel (USV) to perform a specific task and resolving spatiotemporal conflicts between multiple devices. The functional agent, acting as the "hands and feet," is responsible for executing specific algorithms, including UUV path planning, PID control of the baiting machine, and the operation of fault diagnosis algorithms. Through the collaborative work of these four types of agents, this system achieves a complete closed loop from task understanding to resource scheduling to specific execution.

[0035] Example 3 In this embodiment, the system receives a typhoon path forecast from a meteorological satellite and, combined with locally deployed wave meter data, uses an LSTM model for predictive analysis. The results show that the effective wave height in the aquaculture area will exceed 5 meters in the next 24 hours. Based on this early warning information, the interactive agent automatically generates a "typhoon avoidance strategy generation" task and submits it to the cognitive agent for processing. The cognitive agent calls a large model and performs comprehensive reasoning based on the current net cage load status and fish stock data, generating three candidate schemes: Scheme A is to submerge the net cage to 20 meters; Scheme B is to submerge the net cage to 30 meters and increase the anchor chain tension; Scheme C is to maintain the status quo but stop feeding. Subsequently, the simulation module performs fluid-structure interaction simulations on the three schemes in a digital twin. The simulation results show that: under the condition of a wave height of 5 meters, the stress of the netting on the wave-facing side of Scheme A exceeds 85% of the yield strength, posing a risk of tearing; although Scheme B has slightly higher energy consumption, the maximum stress of the structure is controlled within a safe range, and the flow velocity in the deep water area is lower, resulting in the lowest stress response of the fish. Based on simulation results, the system automatically selects Scheme B as the optimal strategy, and coordinates the intelligent agent to generate specific control commands, including starting the ballast water pump and tightening the anchor chain winch, and sends the commands to the actuators. During execution, fiber optic grating sensors transmit stress data in real time, and the system ensures a smooth and controllable descent process through closed-loop feedback.

[0036] Example 4 In this embodiment, the biomass monitoring sonar inside the aquaculture cage first detected an abnormal gathering of fish in the northeast corner of the cage, and the underwater acoustic signal in this area contained non-biological high-frequency noise. The cognitive agent analyzed the above abnormal signal and inferred that there was a high suspicion of "fish escaping due to netting damage or predator invasion," and generated an underwater inspection task accordingly. After receiving the task, the coordinating agent dispatched a nearby underwater inspection robot (UUV) to the target area. In response to the strong current environment of the Yellow and Bohai Seas, the functional agent used a reinforcement learning algorithm to plan an approach path that could utilize the flow field, save energy, and maintain stable attitude. After the UUV arrived at the target area, it activated the camera to take pictures. Due to the turbidity of the water, the system automatically activated the dark channel defogging enhancement algorithm to process the video stream in real time. The large model vision module successfully identified a netting tear of about 20cm in length in the enhanced video stream. Based on the size of the tear and the current size of the fish, the large model determined that repair work was necessary immediately. The system automatically generates repair work orders, notifies maintenance vessels to prepare the necessary materials, and controls the UUV to deploy temporary marker beacons at the damage site. All execution data and diagnostic results are fed back to the knowledge base as diagnostic samples for similar future faults, used for incremental learning and knowledge updates of the model.

[0037] Example 5 In this embodiment, the system automatically formulates a harvesting plan based on market order demand. A coordinating agent simultaneously schedules vacuum fish-suction pumps, automatic grading machines, and live fish transport vessels according to the plan, establishing communication handshakes between the devices to ensure synchronized collaborative operations. After harvesting begins, a visual sensor installed at the suction pipe inlet monitors the density and posture of the fish passing through in real time. When excessive fish density is detected, potentially leading to congestion, the functional agent adjusts the negative pressure frequency of the suction pump within milliseconds to ensure the achievement of the "low-loss live fish" operational goal. As the sucked-up fish flow through the automatic grading machine, computer vision algorithms identify the size and surface defects of each fish in real time. Healthy fish meeting specifications are transported to the live fish hold of the transport vessel, while substandard or injured fish are returned to the aquaculture cages or quarantine chambers. After the operation is completed, the system automatically calculates key performance indicators such as harvest volume, mortality rate, and grading accuracy, and feeds the data back to a large model to optimize the suction parameters and grading strategy for the next operation, forming a continuous improvement closed-loop optimization mechanism.

[0038] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention. The diagram illustrates the end-to-end link of "multimodal full-element perception - knowledge graph and large model cognition - digital twin simulation verification - multi-agent collaborative decision-making - execution and feedback closed loop", and also shows the data flow and feedback loop between the breeding facilities, edge nodes and unmanned equipment.

[0039] Figure 2This is an internal interaction flowchart for the multi-agent collaborative decision-making module. The diagram illustrates the process of task parsing, candidate solution generation, resource constraint checking, spatiotemporal conflict resolution, and task assignment between the interactive agent, cognitive agent, coordinating agent, and functional agent, demonstrating the collaborative orchestration mechanism under complex sea conditions.

[0040] Figure 3 This is a logic diagram for dynamic simulation and risk assessment of net cages based on digital twins. The diagram shows the process of virtual simulation of candidate operation and maintenance strategies by the digital twin under typical wind, wave and current conditions in the Yellow and Bohai Seas, as well as the assessment outputs and screening criteria for structural safety risks, mooring over-limit probability, and biological stress risks.

[0041] Figure 4 This is a timeline diagram of the integrated collaborative operation of "fishing-grading-transfer" under typical sea conditions. The diagram uses a timeline to show the synchronization relationship of multiple pieces of equipment (such as fish suction pumps, graders, transport vessels / unmanned vessels, etc.) within the operation window, as well as the linkage and adjustment process of key triggering events and parameters, reflecting the closed-loop collaborative control strategy.

[0042] Figure 5 This figure compares the workflow and performance of underwater image enhancement and netting damage detection algorithms. It illustrates the image enhancement process for highly turbid water (such as dark channel prior and GAN combined enhancement) and the improvement in the detectability of damage features before and after enhancement. Damage localization and quantization outputs are also provided for subsequent inspections and work order generation.

[0043] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0044] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0045] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep-sea aquaculture system based on a large-scale marine fisheries model and digital twin, characterized in that, include: A high-fidelity acquisition module for multi-source heterogeneous information is used to collect structured time-series data and unstructured multimedia data of the aquaculture facility and its surrounding sea area. The knowledge graph and large model cognition module is used to construct a knowledge graph in the fisheries field and connect to the operation and maintenance knowledge base through retrieval enhancement generation method to provide semantic understanding and causal reasoning capabilities for operation and maintenance tasks and generate a set of candidate operation and maintenance strategies. The digital twin simulation verification module is used to establish a coupled digital twin that includes ocean wind and wave flow field, hydrodynamic load, structural response, and fish physiological stress and behavioral response, and to perform virtual simulation and risk assessment on the strategies in the candidate operation and maintenance strategy set, and screen out executable strategies that meet safety constraints. The multi-agent collaborative decision-making module is used to decompose the executable strategy into tasks and schedule resources through the interaction and collaboration between multiple agents, and generate specific action sequences. The execution and feedback closed-loop module is used to distribute the action sequence to the on-site execution objects, perform online evaluation based on the execution feedback data, and feed the evaluation results back to the knowledge graph and big model cognition module to achieve knowledge updates.

2. The system according to claim 1, characterized in that, The multi-source heterogeneous information high-fidelity acquisition module includes a sea state prediction fusion unit, which is used to fuse on-site observation data with sea state forecast products and output wind, wave and current prediction sequences within a future preset time window.

3. The system according to claim 1, characterized in that, The multi-source heterogeneous information high-fidelity acquisition module includes a turbid water enhancement unit, which uses image enhancement algorithms to perform defogging and detail enhancement processing on the acquired underwater images or videos to improve the detection capability of underwater structures and biological anomalies.

4. The system according to claim 1, characterized in that, The knowledge graph and large model cognition module is specifically used to generate the candidate operation and maintenance strategy set by calling the operation and maintenance knowledge base, which includes operation and maintenance manuals, historical fault logs and expert experience bases, based on a general large language model and through retrieval enhancement generation method.

5. The system according to claim 1, characterized in that, The digital twin simulation verification module is specifically used to calculate the structural safety margin and mooring tension exceedance probability of the candidate operation and maintenance strategy under a given sea state, and to use the calculation results as safety constraints for the action sequence generated by the multi-agent collaborative decision-making module.

6. The system according to claim 1, characterized in that, The multi-agent collaborative decision-making module includes an interactive agent, a cognitive agent, a coordinating agent, and a function execution agent. The interactive agent receives instructions and generates maintenance task lists; the cognitive agent interprets sea state evolution and operational risks; the coordinating agent orchestrates tasks and schedules resources based on the output of the cognitive agent; and the function execution agent generates control parameters and action sequences for specific operational tasks based on the scheduling of the coordinating agent.

7. The system according to claim 1, characterized in that, The execution and feedback closed-loop module includes an online evaluation unit, which is used to calculate the job completion and risk convergence based on the execution feedback data, and write the evaluation results back to the knowledge graph and large model cognition module as feedback data. The execution and feedback closed-loop module also includes an online learning unit, which is used to associate and label the execution feedback data with the corresponding work order results, and to update the retrieval library or model parameters of the knowledge graph and the large model cognition module in an incremental manner.

8. A method for intelligent operation and maintenance against wind and waves based on a large-scale marine fisheries model and digital twin, characterized in that, include: Collect multimodal data and perform time alignment and quality management; Based on the processed data, the operation and maintenance knowledge base is invoked and semantic understanding is performed through a retrieval enhancement generation method to generate candidate operation and maintenance strategies. The candidate operation and maintenance strategies are virtually simulated and risk assessed in the digital twin to select executable strategies; Based on the executable strategy, task orchestration and resource scheduling are performed through multi-agent collaboration to generate action sequences; The action sequence is sent to the on-site execution objects, and the subsequent strategy generation process is modified based on the execution feedback.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, The processor implements the method of claim 8 when executing the computing program.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of claim 8.

Images