Prawn and fish aquaculture method and system
By using multimodal perception networks and hybrid modeling techniques, a coupled digital twin of large yellow croaker's land-sea relay aquaculture was constructed. Combined with multi-agent collaborative learning, a growth-quality collaborative prediction model was generated, which solved the problem of the timing of land-sea transition relying on experience and improved the growth efficiency and market adaptability of large yellow croaker.
Patent Information
- Application Number
- CN202610411969.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-23
AI Technical Summary
The existing technology for the sea-land transition of large yellow croaker farming lacks scientific decision-making basis, and the timing of the sea-land transition relies on experience. It fails to comprehensively consider fish size, growth potential and market price, resulting in stagnation of growth, frequent disease outbreaks and disconnect from market sales.
By acquiring aquaculture environment and biological response data through a multimodal sensing network, a multi-source heterogeneous dataset is constructed. A hybrid modeling method is used to perform coupled digital twin modeling, and a growth-quality collaborative prediction model is established. Combined with a multi-agent collaborative learning framework, an aquaculture management optimization strategy is generated to achieve intelligent management throughout the entire process.
It has improved the growth efficiency, quality consistency and economic benefits of large yellow croaker, and achieved synergistic optimization of the land-sea relay farming process and staggered market sales.
Smart Images

Figure CN122250400A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of land-sea relay aquaculture technology, and in particular to a land-sea relay aquaculture management method and system for large yellow croaker based on segmented aquaculture optimization. Background Technology
[0002] Large yellow croaker is an important economic fish species for marine aquaculture along the southeastern coast of my country, currently primarily cultured in marine cages. However, traditional cage culture is significantly constrained by the natural marine environment. During the low water temperatures of winter, large yellow croakers experience decreased feeding, stagnant growth, or even negative growth, while frequent disease outbreaks severely impact survival rates and the quality of marketable fish. To address the challenges of winter aquaculture, the industry has attempted a sea-land relay model, where fish fry are placed in cages for rapid growth in spring and summer, and then transferred to factory-style facilities for overwintering before winter, achieving off-season market availability. However, existing sea-land relay technologies rely heavily on experience and lack scientific decision-making support. The timing of the sea-land transition is often based on a fixed calendar or a single water temperature threshold, without comprehensively considering factors such as fish size, growth potential, and market price expectations. The management of the cage and factory-style stages is relatively fragmented, with a lack of coordinated optimization of environmental control parameters and feeding strategies. Factory-style recirculating aquaculture involves the coupled regulation of multiple factors such as water temperature, salinity, and light, making it difficult for human experience to simultaneously address multiple objectives such as growth rate, feed efficiency, and body color quality.
[0003] In recent years, technologies such as the Internet of Things (IoT) and artificial intelligence (AI) have been gradually applied in the aquaculture industry. However, existing solutions often focus on optimizing a single link or a single objective, failing to treat the two stages as an organic whole for coordinated optimization. They also do not fully consider the mutual influence between different aquaculture units. Furthermore, the aquaculture plan is disconnected from market sales, resulting in a misalignment between market entry time and price peaks. Therefore, there is an urgent need for a technical solution that integrates modeling and coordinated optimization of the entire process of aquaculture, from environmental perception and growth prediction to strategy optimization and market entry decision-making, to achieve intelligent management across the entire chain. Summary of the Invention
[0004] This invention overcomes the shortcomings of the prior art and provides a method and system for the integrated land-sea aquaculture management of large yellow croaker based on segmented aquaculture optimization. Its main purpose is to improve the growth efficiency, quality consistency and aquaculture economic benefits of large yellow croaker.
[0005] To achieve the above objectives, the first aspect of this invention provides a method for managing the integrated land-sea aquaculture of large yellow croaker based on segmented aquaculture optimization, comprising: A multimodal sensing network is used to acquire aquaculture environment and biological response data and construct a multi-source heterogeneous dataset. Based on the multi-source heterogeneous dataset, a hybrid modeling method is used to perform aquaculture environment-biological coupling modeling to obtain a coupled digital twin of land-sea relay aquaculture. Based on the coupled digital twin, a growth-quality synergistic prediction model is constructed and trained using a physical information neural network as the basic architecture. After training, real-time aquaculture monitoring data is extracted and input through a multi-source heterogeneous dataset to generate growth-quality synergistic prediction information. A state transition function is established based on the growth-quality co-prediction information. Each breeding unit is regarded as an independent agent to establish a multi-agent collaborative learning framework. The coupled digital twin is combined with multiple rounds of repeated trial and error training to finally obtain a multi-agent breeding strategy model. The multi-agent aquaculture strategy model is lightweighted and deployed on the edge computing node of the aquaculture site. Based on the current aquaculture density and fish status of the independent aquaculture unit, the timing of graded transfer is automatically determined, and an aquaculture management optimization strategy including feeding regulation, environmental parameter optimization and transfer planning is generated. When the large yellow croakers in a certain independent aquaculture unit meet the standards for harvesting, quality testing is conducted and a market launch plan is generated based on the current market price. If it is determined that the current time window is not the optimal market launch window, the aquaculture progress planning and management are carried out based on the aquaculture status of other independent aquaculture units.
[0006] In this solution, the step of acquiring aquaculture environment and biological response data through a multimodal sensing network and constructing a multi-source heterogeneous dataset specifically includes: Underwater autonomous cruising platforms are deployed in cage aquaculture areas and factory workshops. These platforms are equipped with hyperspectral imagers, microelectrode arrays, and acoustic Doppler current profilers, which are used to collect fish surface reflectance spectra, water vertical temperature gradients and dissolved oxygen micro-distribution, and three-dimensional flow field characteristics, respectively. Simultaneously deploy an automatic environmental DNA sampling device in the aquaculture water body. The automatic environmental DNA sampling device collects water samples at set time intervals and performs online amplification and detection through a microfluidic chip to obtain data on the abundance of pathogenic microorganisms and the concentration of stress hormone metabolites in the water body. After data collection is completed, the data is transmitted to the edge data aggregation node in real time. At the aggregation node, the original data is timestamped and outliers are removed. The color characteristics and swimming trajectories of fish groups are extracted from hyperspectral images. Environmental DNA data is converted into pathogen abundance levels and stress level indices, ultimately forming a multi-source heterogeneous dataset.
[0007] In this scheme, the hybrid modeling method based on the multi-source heterogeneous dataset is used to perform aquaculture environment-biological coupling modeling to obtain a coupled digital twin of land-sea relay aquaculture, specifically including: Based on the multi-source heterogeneous dataset, water level, flow velocity and boundary temperature data for continuous time periods are extracted, and input into a computational fluid dynamics simulator to solve the Reynolds-averaged Navier-Stokes equations. Combined with the standard k-ε turbulence model, the aquaculture area is divided into hexahedral grid cells, and the flow velocity vector, temperature scalar and dissolved oxygen flux of each grid point in three-dimensional space are output. After post-processing, the water exchange rate and temperature stratification coefficient are obtained as hydrodynamic characteristic fields. The fish population location distribution data extracted from the multi-source heterogeneous dataset during the same period were spatially registered with the hydrodynamic feature field, and the maximum likelihood estimation was used to fit the fish population's preferred flow velocity range and avoidance temperature threshold. An intelligent agent model based on individual instincts is established, and each virtual fish is given the attributes of perception radius, swimming speed and feeding preference. The preferred flow rate range and avoidance temperature threshold are used as behavioral constraints. The obstacle avoidance, gathering and escape behavior of the fish group are described by the social force model, and the time-varying distribution density and feeding activity index of the fish group are calculated and output. The daily food intake, ambient temperature, and transcriptomic metabolic pathway features of fish were extracted from the multi-source heterogeneous dataset. A physiological growth sub-model based on a biochemical reaction network was constructed. The sub-model takes ambient temperature, food intake, and activity energy consumption as inputs and outputs physiological growth data including daily weight gain and body color change data. The hydrodynamic characteristic field, the time-varying distribution density of the fish school, the feeding activity index, and the physiological growth data are input into the Bayesian assimilation engine, and the assimilation correction is performed using the ensemble Kalman filter algorithm to obtain a coupled digital twin that synchronously maps the real-time state of the physical aquaculture system.
[0008] In this scheme, the growth-quality co-prediction model is constructed and trained based on the coupled digital twin using a physical information neural network architecture. After training, real-time aquaculture monitoring data is extracted from and input into multi-source heterogeneous datasets to generate growth-quality co-prediction information, specifically including: Historical simulation data is obtained based on the coupled digital twin, including hourly hydrodynamic characteristic fields, time-varying distribution density of fish schools, feeding activity index, and daily weight gain and body color change records of each virtual fish. The actual observation data for the corresponding time period is extracted from the multi-source heterogeneous dataset. The simulation data and the actual observation data are aligned by time to form a training sample set. Each sample contains the environmental sequence of day t as input features and the average weight gain rate and body surface color value of day t+1 as prediction labels. A growth-quality co-prediction model is constructed using a physical information neural network as its basic architecture. The model is trained using the training sample set. During the training process, the bioenergetics equation of large yellow croaker growth is embedded as a physical constraint into the loss function. The ambient temperature, food intake, and current weight are extracted from the training samples. The theoretical weight gain rate is calculated using the bioenergetics equation as the theoretical value. The mean square error between the theoretical value and the neural network prediction value is used as the physical constraint term. The error between the neural network prediction value and the predicted label in the sample is used as the data-driven error term. The total loss function is formed by weighted summation of the physical constraint term and the total loss function. The weight coefficients of the total loss function are determined by the grid search method. An adaptive moment estimation optimizer is adopted and an initial learning rate is set. After the hidden layer, a multi-head attention layer is added to calculate the attention weight of each time step in the input feature sequence. Key influencing factors are highlighted by weighted summation. The model is repeatedly trained until the deviation between the output of the growth-quality co-prediction model and the preset prediction label is less than a preset threshold. The desired growth-quality co-prediction model is then output. After training, real-time aquaculture monitoring data is extracted from the multi-source heterogeneous dataset as model input. The real-time aquaculture monitoring data includes the hourly average water temperature, daily cumulative feed intake, fish activity index and current average weight for the consecutive t days before the current time. The predicted daily weight gain rate and body color of day t+1 are calculated and output through forward propagation, and finally the growth-quality co-prediction information is obtained.
[0009] In this scheme, the establishment of a state transition function based on the growth-quality co-prediction information, the establishment of a multi-agent collaborative learning framework by treating each breeding unit as an independent agent, and the multi-round trial-and-error training by combining the coupled digital twin to finally obtain a multi-agent breeding strategy model, specifically includes: Acquire growth-quality co-prediction information, using the current environmental state and fish state of each aquaculture unit as input, and construct a state transition function by combining the weight gain rate and body color prediction value for the next t days. The environmental state includes water temperature, salinity, dissolved oxygen, ammonia nitrogen concentration and light intensity, and the fish state includes current average weight, body color value, feeding activity index and stocking density. Each net cage and each factory-style aquaculture pond in the aquaculture system is defined as an independent intelligent agent. Each intelligent agent can observe the local state information of its own unit and execute a set of actions, including adjusting the feeding amount of the automatic feeder, setting the target temperature of the circulating water system, adjusting the light intensity and the proportion of red light, and providing early warning prompts for transfer. A multi-agent collaborative learning framework is constructed using a centralized training and decentralized execution paradigm. In this framework, a policy network is built for each agent to generate actions based on current observations, and an action value network is built to evaluate the quality of the joint actions of all agents. Simultaneously, a reward function is constructed based on the weighted reward of growth rate, reward of feed efficiency, penalty of survival rate, reward of body color quality, and penalty of energy consumption cost. A market condition adjustment factor is introduced to dynamically adjust the weights of each item. During the price peak, the weights of growth rate and body color quality are increased, and during the price trough, the weights of feed efficiency and energy consumption cost are increased. The state transition function, policy network, value network, and reward function are embedded into a coupled digital twin to initiate multiple rounds of repeated trial and error training. At the beginning of each training round, the initial state is randomly sampled. Each agent selects an action and executes it according to the current policy network. The digital twin updates the state according to the state transition function and calculates the immediate reward. The experience trajectory is stored in the experience replay buffer. Randomly sample small batches of data from the experience replay buffer and update the parameters of the policy network and the value network. The value network update objective is to increase the value of the next state after the reward is discounted, and the policy network update objective is to maximize the value network's score for the current action. After iterative training, when the value network loss function converges and the cumulative reward of the action sequence output by the policy network in the digital twin reaches a preset threshold, the current policy network parameters are saved as the final multi-agent farming policy model.
[0010] In this solution, the multi-agent aquaculture strategy model is lightweighted and deployed on edge computing nodes at the aquaculture site. Based on the current stocking density and fish status of individual aquaculture units, the timing of graded transfers is automatically determined, generating an aquaculture management optimization strategy that includes feeding control, environmental parameter optimization, and transfer planning. Specifically, this includes: A multi-agent farming strategy model that has been trained is obtained. The original strategy network is lightweighted using knowledge distillation technology. A student network is constructed as a lightweight strategy model. Distillation training is performed by minimizing the KL divergence between the output of the student network and the output of the original strategy network. After distillation, the student network parameters are packaged into an inference engine format file suitable for edge deployment, and deployed in the field control cabinet of each net cage aquaculture area and the distributed control unit of the factory workshop respectively; The edge computing node loads the deployed lightweight strategy model, collects real-time data from the sensor array at a fixed frequency, reads the environmental status and fish status of the current aquaculture unit in each control cycle, performs standardization processing, and inputs the data into the lightweight strategy model for forward inference calculation. The current aquaculture management optimization strategy is calculated and output through forward inference and converted into corresponding control signals and sent to each actuator to achieve real-time regulation. The aquaculture management optimization strategy includes feeding rate setpoint, water temperature setpoint, LED red light ratio adjustment value, and whether to activate transfer warning. The edge computing node determines the timing of transfer based on the transfer warning signal and the comparison between the current breeding density and the preset density threshold. When the transfer warning signal persists for n control cycles and the current breeding density exceeds the density threshold, it is determined that transfer is required, and a transfer planning scheme containing the percentage of fish to be transferred and the optimal transfer time window is generated. For independent aquaculture units where the transfer warning signal is not triggered or the aquaculture density is below the threshold, the edge node continuously executes feeding control and environmental parameter optimization strategies. After each control cycle, the current status, executed actions, and subsequent sensor feedback data are packaged to generate a decision log, which is then uploaded to the cloud digital twin platform via wireless network to adjust the original multi-agent strategy model.
[0011] In this plan, when the large yellow croakers in a certain independent aquaculture unit meet the harvest standards, quality testing is conducted and a market launch plan is generated based on the current market price. If it is determined that the current time window is not the optimal market launch window, then aquaculture progress planning and management are carried out based on the aquaculture status of other independent aquaculture units, specifically including: When the large yellow croakers in the independent breeding unit reach the preset size for harvesting, the surface reflectance spectrum data of the fish is collected by a hyperspectral imager and the depth of the characteristic absorption peak of carotenoids is extracted to calculate the color grade of the surface. The near-infrared absorption spectrum of the muscle is collected by a near-infrared spectrometer and the fat content and elasticity coefficient are predicted. Multiple quality indicators are input into the quality grading model to classify the fish grades and obtain the quality test results. Simultaneously obtain real-time price information from major trading markets through cloud market data interfaces, including the daily transaction price and weekly price trend of large yellow croaker of different specifications and grades. Combine the quality inspection results to conduct market launch decision analysis, aiming to maximize unit output revenue, and using quality grade, price prediction and storage costs as constraints, and adopting a rolling time domain optimization algorithm to calculate the theoretical revenue value of each day's market launch within a set number of days in the future. When the optimal listing time window is determined to be the current date, the listing execution process is triggered and the preliminary listing plan is converted into a shipping instruction. The origin quarantine certificate and quality certification certificate are automatically generated through smart contracts and pushed to the sales platform to complete the transaction. When the optimal listing time window is determined to be a future date, the breeding progress planning and management state is entered. Under the planning and management of breeding progress, the weight gain rate prediction sequence and body color quality evolution trajectory of all other independent breeding units in the future cycle are obtained. With the goal of maximizing the overall breeding benefits, a multi-unit collaborative planning problem is constructed. The genetic algorithm is used to solve the feeding intensity adjustment coefficient, environmental temperature compensation value and light intervention scheme of each breeding unit. The obtained feeding intensity adjustment coefficient, ambient temperature compensation value and light intervention scheme are converted into control commands and sent to the edge computing nodes of the corresponding breeding units. The feeding curve of the automatic feeder, the target temperature curve of the circulating water system and the light period and spectral ratio of the LED light source are dynamically adjusted to achieve staggered matching of the market time and price peak of different breeding units. During the implementation of the breeding progress plan, the actual growth response and market price changes of each breeding unit are continuously monitored. The original plan is re-optimized every week. When abnormal fluctuations in market prices are detected, an emergency adjustment mechanism is triggered to re-solve the planning problem and update the control instructions, so as to achieve the orderly market launch of all breeding units in batches.
[0012] The second aspect of the present invention provides a large yellow croaker sea-land relay aquaculture management system based on segmented aquaculture optimization. The system includes: a memory, a processor, and a communication interface. The memory contains a large yellow croaker sea-land relay aquaculture management method program based on segmented aquaculture optimization. When the large yellow croaker sea-land relay aquaculture management method program based on segmented aquaculture optimization is executed by the processor, it implements the steps of the large yellow croaker sea-land relay aquaculture management method based on segmented aquaculture optimization as described in any of the above claims. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.
[0014] Figure 1 A flowchart of the first method of a land-sea relay aquaculture management method for large yellow croaker based on segmented aquaculture optimization is provided in an embodiment of the present invention; Figure 2 A flowchart of the second method of a land-sea relay aquaculture management method for large yellow croaker based on segmented aquaculture optimization is provided in an embodiment of the present invention; Figure 3 This is a block diagram of a large yellow croaker sea-land relay aquaculture management system based on segmented aquaculture optimization, provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0015] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0017] Figure 1 A flowchart of the first method of a land-sea relay aquaculture management method for large yellow croaker based on segmented aquaculture optimization is provided in an embodiment of the present invention; like Figure 1 As shown, the present invention provides a first method flowchart of a large yellow croaker sea-land relay aquaculture management method based on segmented aquaculture optimization, comprising: S102, acquire aquaculture environment and biological response data through a multimodal sensing network and construct a multi-source heterogeneous dataset. Based on the multi-source heterogeneous dataset, use a hybrid modeling method to perform aquaculture environment-biological coupling modeling to obtain a coupled digital twin of land-sea relay aquaculture. S104. Based on the coupled digital twin, a growth-quality synergistic prediction model is constructed and trained using a physical information neural network as the basic architecture. After training, real-time aquaculture monitoring data is extracted and input through a multi-source heterogeneous dataset to generate growth-quality synergistic prediction information. S106. Based on the growth-quality collaborative prediction information, a state transition function is established. Each breeding unit is regarded as an independent intelligent agent to establish a multi-agent collaborative learning framework. Combined with the coupled digital twin, multiple rounds of repeated trial and error training are carried out to finally obtain a multi-agent breeding strategy model. S108, The multi-agent aquaculture strategy model is lightweighted and deployed on the edge computing node of the aquaculture site. Based on the current aquaculture density and fish status of the independent aquaculture unit, the timing of graded transfer is automatically determined, and an aquaculture management optimization strategy including feeding regulation, environmental parameter optimization and transfer planning is generated. S110: When the large yellow croaker in an independent aquaculture unit meets the standards for leaving the pond, quality testing is conducted and a market launch plan is generated based on the current market price. If it is determined that the current time window is not the optimal market launch window, the aquaculture progress planning and management is carried out based on the aquaculture status of other independent aquaculture units.
[0018] Furthermore, in a preferred embodiment of the present invention, the step of acquiring aquaculture environment and biological response data through a multimodal sensing network and constructing a multi-source heterogeneous dataset specifically includes: Underwater autonomous cruising platforms are deployed in cage aquaculture areas and factory workshops. These platforms are equipped with hyperspectral imagers, microelectrode arrays, and acoustic Doppler current profilers, which are used to collect fish surface reflectance spectra, water vertical temperature gradients and dissolved oxygen micro-distribution, and three-dimensional flow field characteristics, respectively. Simultaneously deploy an automatic environmental DNA sampling device in the aquaculture water body. The automatic environmental DNA sampling device collects water samples at set time intervals and performs online amplification and detection through a microfluidic chip to obtain data on the abundance of pathogenic microorganisms and the concentration of stress hormone metabolites in the water body. After data collection is completed, the data is transmitted to the edge data aggregation node in real time. At the aggregation node, the original data is timestamped and outliers are removed. The color characteristics and swimming trajectories of fish groups are extracted from hyperspectral images. Environmental DNA data is converted into pathogen abundance levels and stress level indices, ultimately forming a multi-source heterogeneous dataset.
[0019] It should be noted that the hyperspectral imager continuously collects the reflectance spectrum of the fish's body surface during navigation. By analyzing the absorption characteristics of specific wavebands, it quantifies the level of carotenoid deposition on the body surface, thereby indirectly reflecting the degree of fish coloration. A microelectrode array vertically penetrates the water body, recording instantaneous values of temperature and dissolved oxygen at different depths to construct a vertical stratification map of the water body. An acoustic Doppler current profiler calculates the three-dimensional velocity distribution by emitting sound waves and receiving echo frequency shifts, providing basic input for subsequent hydrodynamic modeling. A synchronously deployed automated environmental DNA sampling device extracts water samples at a set sampling frequency. The water samples are automatically processed by a microfluidic chip for nucleic acid extraction and real-time quantitative PCR amplification, outputting online the gene copy numbers of pathogenic microorganisms such as Vibrio and Pseudomonas, as well as the cortisol concentration measured by enzyme-linked immunosorbent assay (ELISA). The former is used to assess disease risk, and the latter to quantify the stress level of the fish population.
[0020] After all collected data is transmitted to the edge data aggregation node in real time, the aggregation node first timestamps the multi-source data to ensure accurate correlation of environmental, biological, and physiological data at the same time. Then, outliers caused by sensor drift or communication errors are removed. For hyperspectral images, regions of interest (ROIs) are extracted from the fish using segmentation algorithms, and the reflectance ratio of these regions in specific wavelength bands is calculated as chromaticity features. Simultaneously, optical flow is used to track the centroid displacement of the fish school in continuous images, obtaining swimming speed and turning frequency as swimming trajectory features. The raw environmental DNA data is converted from cyclic thresholds to absolute abundance based on a standard curve, and then divided into low, medium, and high abundance levels according to preset thresholds. Cortisol concentration is normalized to a stress level index of zero to one after temperature correction. After the above processing, a multi-source heterogeneous dataset is finally formed, containing a time-series sequence of physical environmental parameters, fish behavior characteristics, pathogen abundance levels, and a stress level index.
[0021] Furthermore, in a preferred embodiment of the present invention, the step of using a hybrid modeling method based on the multi-source heterogeneous dataset to perform aquaculture environment-biological coupling modeling to obtain a coupled digital twin of land-sea relay aquaculture specifically includes: Based on the multi-source heterogeneous dataset, water level, flow velocity and boundary temperature data for continuous time periods are extracted, and input into a computational fluid dynamics simulator to solve the Reynolds-averaged Navier-Stokes equations. Combined with the standard k-ε turbulence model, the aquaculture area is divided into hexahedral grid cells, and the flow velocity vector, temperature scalar and dissolved oxygen flux of each grid point in three-dimensional space are output. After post-processing, the water exchange rate and temperature stratification coefficient are obtained as hydrodynamic characteristic fields. The fish population location distribution data extracted from the multi-source heterogeneous dataset during the same period were spatially registered with the hydrodynamic feature field, and the maximum likelihood estimation was used to fit the fish population's preferred flow velocity range and avoidance temperature threshold. An intelligent agent model based on individual instincts is established, and each virtual fish is given the attributes of perception radius, swimming speed and feeding preference. The preferred flow rate range and avoidance temperature threshold are used as behavioral constraints. The obstacle avoidance, gathering and escape behavior of the fish group are described by the social force model, and the time-varying distribution density and feeding activity index of the fish group are calculated and output. The daily food intake, ambient temperature, and transcriptomic metabolic pathway features of fish were extracted from the multi-source heterogeneous dataset. A physiological growth sub-model based on a biochemical reaction network was constructed. The sub-model takes ambient temperature, food intake, and activity energy consumption as inputs and outputs physiological growth data including daily weight gain and body color change data. The hydrodynamic characteristic field, the time-varying distribution density of the fish school, the feeding activity index, and the physiological growth data are input into the Bayesian assimilation engine, and the assimilation correction is performed using the ensemble Kalman filter algorithm to obtain a coupled digital twin that synchronously maps the real-time state of the physical aquaculture system.
[0022] It should be noted that, firstly, key parameters such as water level, flow velocity, and boundary temperature over continuous time periods are extracted from the dataset and used as input conditions for the computational fluid dynamics simulator. The simulator divides the aquaculture area into millions of hexahedral grid cells, solves the Reynolds-averaged Navier-Stokes equations on each cell, and combines them with the standard k-ε turbulence model to describe the turbulent characteristics of the water body. After solving, the flow velocity vector, temperature scalar, and dissolved oxygen flux of each grid point are output. After post-processing, the water exchange rate and temperature stratification coefficient of the entire aquaculture area are calculated, which together constitute the hydrodynamic feature field to describe the dynamic distribution of the aquatic physical environment. For example, the water exchange rate reflects the renewal rate of the water inside and outside the cage, directly affecting the dilution efficiency of metabolic waste. Next, the fish location distribution data from the multi-source heterogeneous dataset of the same period are spatially registered with the hydrodynamic feature field. By statistically analyzing the correspondence between the frequency of fish occurrence in each grid cell and the corresponding flow velocity and temperature, maximum likelihood estimation is used to fit the flow velocity range preferred by the fish and the temperature threshold to be avoided, which are then used as behavioral constraints for each virtual fish in the agent model. The agent model assigns each virtual fish attributes such as perception radius, swimming speed, and feeding preference. A social force model simulates the interactions between individual fish in the school, including repulsive forces during obstacle avoidance, attractive forces during aggregation, and driving forces during escape. After parallel computation of all virtual fish, the time-varying distribution density of the fish school in three-dimensional space and the number of feeding attempts per unit time are statistically obtained as a feeding activity index.
[0023] By extracting gene expression characteristics related to growth from daily feed intake, ambient temperature, and transcriptome sequencing data of fish populations using multi-source heterogeneous datasets, a physiological growth sub-model based on a biochemical reaction network was constructed. This sub-model uses the Michael-Menten kinetic equation to describe the relationship between protein synthesis rate and amino acid concentration, the Arrhenius equation to describe the effect of temperature on basal metabolic rate, and the law of mass action to describe the transport process of carotenoids from intestinal absorption to skin deposition. Using ambient temperature, daily feed intake, and energy consumption of fish population activity as inputs, the model solves a coupled set of ordinary differential equations to output daily weight gain and body color changes for each virtual fish. Finally, the hydrodynamic feature field, time-varying distribution density of the fish population, feeding activity index, and physiological growth data are jointly input into a Bayesian assimilation engine. This engine employs an ensemble Kalman filter algorithm. First, it constructs a model state set consisting of hundreds of samples, each representing a set of possible model parameters. Then, it calculates the prediction bias based on real-time water quality sensor measurements, underwater image analysis results, and environmental DNA detection values, adjusting the sample state variables using the bias covariance matrix. After multiple iterations of assimilation, the model parameters converged, ultimately resulting in a coupled digital twin that can synchronously map the real-time state of the physical aquaculture system and has forward prediction capabilities.
[0024] Furthermore, in a preferred embodiment of the present invention, the step of constructing and training a growth-quality synergistic prediction model based on the coupled digital twin using a physical information neural network as the underlying architecture, and then extracting and inputting real-time aquaculture monitoring data through a multi-source heterogeneous dataset after training to generate growth-quality synergistic prediction information, specifically includes: Historical simulation data is obtained based on the coupled digital twin, including hourly hydrodynamic characteristic fields, time-varying distribution density of fish schools, feeding activity index, and daily weight gain and body color change records of each virtual fish. The actual observation data for the corresponding time period is extracted from the multi-source heterogeneous dataset. The simulation data and the actual observation data are aligned by time to form a training sample set. Each sample contains the environmental sequence of day t as input features and the average weight gain rate and body surface color value of day t+1 as prediction labels. A growth-quality co-prediction model is constructed using a physical information neural network as its basic architecture. The model is trained using the training sample set. During the training process, the bioenergetics equation of large yellow croaker growth is embedded as a physical constraint into the loss function. The ambient temperature, food intake, and current weight are extracted from the training samples. The theoretical weight gain rate is calculated using the bioenergetics equation as the theoretical value. The mean square error between the theoretical value and the neural network prediction value is used as the physical constraint term. The error between the neural network prediction value and the predicted label in the sample is used as the data-driven error term. The total loss function is formed by weighted summation of the physical constraint term and the total loss function. The weight coefficients of the total loss function are determined by the grid search method. An adaptive moment estimation optimizer is adopted and an initial learning rate is set. After the hidden layer, a multi-head attention layer is added to calculate the attention weight of each time step in the input feature sequence. Key influencing factors are highlighted by weighted summation. The model is repeatedly trained until the deviation between the output of the growth-quality co-prediction model and the preset prediction label is less than a preset threshold. The desired growth-quality co-prediction model is then output. After training, real-time aquaculture monitoring data is extracted from the multi-source heterogeneous dataset as model input. The real-time aquaculture monitoring data includes the hourly average water temperature, daily cumulative feed intake, fish activity index and current average weight for the consecutive t days before the current time. The predicted daily weight gain rate and body color of day t+1 are calculated and output through forward propagation, and finally the growth-quality co-prediction information is obtained.
[0025] It should be noted that when constructing the growth-quality co-prediction model, historical simulation data was first exported in batches from the coupled digital twin. This included hydrodynamic characteristic fields, fish population density, feeding activity, and growth and body color changes of each virtual fish during the aquaculture process, reflecting the fish population response patterns under different environmental conditions. The simulation data was then time-aligned with the actual observation data for the corresponding time periods to form a training sample set. Each sample used a continuous t-day environmental sequence as input features, and the average weight gain rate and body color value on day t+1 as prediction labels.
[0026] The model employs a physical information neural network as its infrastructure. Unlike traditional black-box models, this network incorporates the bioenergetics equations governing the growth of large yellow croaker as physical constraints into the loss function during training. Specifically, environmental temperature, food intake, and current weight are extracted from each training sample, and the theoretical weight gain rate is calculated using the bioenergetics equations. The mean squared error between this theoretical value and the neural network's prediction is used as the physical constraint term, while the error between the neural network's prediction and the true labels in the samples is used as the data-driven error term. The weighted sum of these two terms constitutes the total loss function. The weight coefficients are determined using a grid search method to balance the contributions of physical laws and data fitting. This design ensures that the model adheres to biological growth patterns while retaining the flexibility of data-driven approaches. Model training employs an adaptive moment estimation optimizer with a set initial learning rate. To improve prediction accuracy, a multi-head attention layer is added after the hidden layers. The attention weights at each time step in the input feature sequence are calculated, and the weighted sum highlights the contributions of key influencing factors such as environmental temperature and food intake to growth and body color. The training process iterates continuously until the deviation between the model output and the preset labels is less than a preset threshold, at which point a predictive model that meets expectations is output.
[0027] After training, real-time aquaculture monitoring data is extracted from a multi-source heterogeneous dataset as model input. This real-time monitoring data includes the hourly average water temperature, daily cumulative feed intake, fish activity index, and current average body weight for the consecutive days preceding the current time (t days). After undergoing the same normalization process as during training, the input data is used to calculate and output the predicted daily average weight gain rate and body color for day t+1 through forward propagation. This generates growth-quality co-prediction information containing multi-dimensional temporal information, providing a forward-looking basis for subsequent aquaculture strategy optimization.
[0028] Furthermore, in a preferred embodiment of the present invention, the step of establishing a state transition function based on the growth-quality collaborative prediction information, establishing a multi-agent collaborative learning framework by treating each breeding unit as an independent agent, and conducting multiple rounds of repeated trial and error training in conjunction with the coupled digital twin to finally obtain a multi-agent breeding strategy model specifically includes: Acquire growth-quality co-prediction information, using the current environmental state and fish state of each aquaculture unit as input, and construct a state transition function by combining the weight gain rate and body color prediction value for the next t days. The environmental state includes water temperature, salinity, dissolved oxygen, ammonia nitrogen concentration and light intensity, and the fish state includes current average weight, body color value, feeding activity index and stocking density. Each net cage and each factory-style aquaculture pond in the aquaculture system is defined as an independent intelligent agent. Each intelligent agent can observe the local state information of its own unit and execute a set of actions, including adjusting the feeding amount of the automatic feeder, setting the target temperature of the circulating water system, adjusting the light intensity and the proportion of red light, and providing early warning prompts for transfer. A multi-agent collaborative learning framework is constructed using a centralized training and decentralized execution paradigm. In this framework, a policy network is built for each agent to generate actions based on current observations, and an action value network is built to evaluate the quality of the joint actions of all agents. Simultaneously, a reward function is constructed based on the weighted reward of growth rate, reward of feed efficiency, penalty of survival rate, reward of body color quality, and penalty of energy consumption cost. A market condition adjustment factor is introduced to dynamically adjust the weights of each item. During the price peak, the weights of growth rate and body color quality are increased, and during the price trough, the weights of feed efficiency and energy consumption cost are increased. The state transition function, policy network, value network, and reward function are embedded into a coupled digital twin to initiate multiple rounds of repeated trial and error training. At the beginning of each training round, the initial state is randomly sampled. Each agent selects an action and executes it according to the current policy network. The digital twin updates the state according to the state transition function and calculates the immediate reward. The experience trajectory is stored in the experience replay buffer. Randomly sample small batches of data from the experience replay buffer and update the parameters of the policy network and the value network. The value network update objective is to increase the value of the next state after the reward is discounted, and the policy network update objective is to maximize the value network's score for the current action. After iterative training, when the value network loss function converges and the cumulative reward of the action sequence output by the policy network in the digital twin reaches a preset threshold, the current policy network parameters are saved as the final multi-agent farming policy model.
[0029] It should be noted that when constructing the multi-agent collaborative learning framework, a state transition function is defined to describe the evolutionary laws of the aquaculture system. This function takes the current environmental state and fish population state of each aquaculture unit as input, and combines them with the predicted weight gain rate and body color for the next t days output by the growth-quality collaborative prediction model to form a state space that reflects the dynamics of aquaculture. The environmental state includes adjustable parameters such as water temperature, salinity, dissolved oxygen, ammonia nitrogen concentration, and light intensity, while the fish population state includes biological indicators such as current average weight, body color value, feeding activity index, and stocking density. The establishment of the state transition function enables the agents to predict the impact of different actions on the future state of the system, providing a basis for decision optimization.
[0030] Each net cage and each factory-style aquaculture pond within the aquaculture system is defined as an independent agent. Each agent can only observe the local state information of its own unit and execute a specific set of actions. These actions include adjusting the feeding amount of the automatic feeder to match the feeding needs of the fish, setting the target temperature of the recirculating aquaculture system to maintain the optimal growth environment, adjusting the LED light intensity and red light ratio to promote pigment deposition on the body surface, and providing early warning of transfer. A multi-agent collaborative learning framework is constructed using a centralized training and decentralized execution paradigm. During the training phase, a policy network is built for each agent to generate actions based on current observations, while a centralized action value network is built to evaluate the merits of the joint actions of all agents. This architecture allows agents to learn collaborative strategies using global information during training and to make independent decisions based solely on local observations during execution, balancing learning effectiveness and deployment feasibility.
[0031] The design of the reward function directly influences the learning direction of the agent. This function is a weighted average of growth rate reward, feed efficiency reward, survival rate penalty, body color quality reward, and energy cost penalty. To adapt to market changes, a market condition adjustment factor is introduced to dynamically adjust the weights of each factor. During peak price periods, the weights of growth rate and body color quality are increased to pursue higher premiums, while during trough price periods, the weights of feed efficiency and energy cost are increased to control costs. For example, when the price of large yellow croaker rises before the Spring Festival, the agent will tend to adopt a more aggressive feeding strategy and stronger light intervention to improve growth rate and body color quality. After embedding the state transition function, policy network, value network, and reward function into a coupled digital twin, multiple rounds of repeated trial-and-error training are initiated. At the beginning of each training round, the initial state is randomly sampled to cover different aquaculture scenarios. Each agent selects and executes an action according to the current policy network. The digital twin updates the environment and fish swarm state according to the state transition function and calculates the immediate reward, storing the complete experience trajectory in an experience replay buffer. During training, small batches of data are randomly sampled from the buffer to update network parameters. The update objective of the value network is the next-state value after adding a discount to the reward, while the update objective of the policy network is to maximize the value network's score for the current action. After iterative training, when the value network loss function converges and the cumulative reward of the action sequence output by the policy network reaches a preset threshold, the current policy network parameters are saved as the final multi-agent aquaculture strategy model. This model can automatically generate optimal control instructions for each aquaculture unit based on real-time observations, achieving comprehensive optimization of growth efficiency, quality improvement, and energy consumption costs.
[0032] It is worth mentioning that the specific form of the reward function is as follows: , Among them, the For growth rate bonus items, This represents the current daily weight gain rate. Target weight gain rate; The This is a reward item for body color quality. This represents the current body surface color value. The target chromaticity value; The As a penalty item for survival rate, This represents the number of deaths that day. This represents the current number of animals in stock. The This is a feed efficiency bonus item. For the current feed conversion rate, As the baseline value; The As a penalty item for energy consumption costs, This represents the current energy consumption of the equipment. Maximum permissible energy consumption; The This is a market price adjustment factor, dynamically adjusted based on market price forecasts; The basic weighting coefficients can be pre-calibrated using the analytic hierarchy process and expert experience.
[0033] Furthermore, in a preferred embodiment of the present invention, the multi-agent aquaculture strategy model is lightweighted and deployed on edge computing nodes at the aquaculture site. Based on the current stocking density and fish status of independent aquaculture units, the timing of graded transfers is automatically determined, generating an aquaculture management optimization strategy that includes feeding control, environmental parameter optimization, and transfer planning. Specifically, this includes: A multi-agent farming strategy model that has been trained is obtained. The original strategy network is lightweighted using knowledge distillation technology. A student network is constructed as a lightweight strategy model. Distillation training is performed by minimizing the KL divergence between the output of the student network and the output of the original strategy network. After distillation, the student network parameters are packaged into an inference engine format file suitable for edge deployment, and deployed in the field control cabinet of each net cage aquaculture area and the distributed control unit of the factory workshop respectively; The edge computing node loads the deployed lightweight strategy model, collects real-time data from the sensor array at a fixed frequency, reads the environmental status and fish status of the current aquaculture unit in each control cycle, performs standardization processing, and inputs the data into the lightweight strategy model for forward inference calculation. The current aquaculture management optimization strategy is calculated and output through forward inference and converted into corresponding control signals and sent to each actuator to achieve real-time regulation. The aquaculture management optimization strategy includes feeding rate setpoint, water temperature setpoint, LED red light ratio adjustment value, and whether to activate transfer warning. The edge computing node determines the timing of transfer based on the transfer warning signal and the comparison between the current breeding density and the preset density threshold. When the transfer warning signal persists for n control cycles and the current breeding density exceeds the density threshold, it is determined that transfer is required, and a transfer planning scheme containing the percentage of fish to be transferred and the optimal transfer time window is generated. For independent aquaculture units where the transfer warning signal is not triggered or the aquaculture density is below the threshold, the edge node continuously executes feeding control and environmental parameter optimization strategies. After each control cycle, the current status, executed actions, and subsequent sensor feedback data are packaged to generate a decision log, which is then uploaded to the cloud digital twin platform via wireless network to adjust the original multi-agent strategy model.
[0034] It should be noted that due to the large number of parameters in the original policy network, it is difficult to run directly on resource-constrained edge devices. Therefore, knowledge distillation is used for lightweight processing. Knowledge distillation uses the original policy network as the teacher network and constructs a student network as the lightweight policy model. Distillation training is performed by minimizing the KL divergence between the student network output and the teacher network output, enabling the student network to learn the decision boundary of the teacher network. After distillation, the inference speed and model size of the student network are optimized. Subsequently, the student network parameters are packaged into an inference engine format file and deployed in the field control cabinet of each net cage aquaculture area and the distributed control unit of the factory workshop.
[0035] After the edge computing node starts, it loads a lightweight strategy model and collects real-time data from the sensor array at a fixed frequency. Within each control cycle, the node reads the environmental and fish status of the current aquaculture unit, including water quality parameters such as water temperature, dissolved oxygen, and ammonia nitrogen, as well as biological indicators such as average weight, body color, feeding activity, and stocking density. Body color is obtained by extracting the region of interest (ROI) of the fish body using a lightweight image segmentation algorithm after images are captured by an underwater camera, and calculating the average of the RGB three channels. Feeding activity is calculated by analyzing sonar data to determine the swimming speed and the number of feeding impacts. The node then outputs the current aquaculture management optimization strategy through forward inference calculations. This strategy includes continuous and discrete actions. Continuous actions include the automatic feeder's feeding rate setpoint, the target temperature setpoint for the circulating water system, and the LED red light ratio adjustment value. Discrete actions include whether to trigger graded screening and whether to activate transfer warnings.
[0036] Based on the transfer warning signal output by the lightweight strategy model, a comprehensive judgment is made by comparing the current stocking density with a preset density threshold. When the transfer warning signal persists for multiple control cycles and the current stocking density exceeds the preset threshold, transfer is deemed necessary. A transfer plan is generated, including the percentage of fish to be transferred and the optimal transfer time window, and the transfer request is sent to the cloud scheduling platform via a wireless communication module. For independent aquaculture units where the transfer warning signal is not triggered or the stocking density is below the threshold, edge nodes continuously execute feeding control and environmental parameter optimization strategies. At the end of each control cycle, the current status, executed actions, and subsequent sensor feedback data are packaged into a decision log, which is uploaded to the cloud digital twin platform periodically via a wireless network. The cloud platform uses this actual operational data to continuously fine-tune the original multi-agent strategy model. When the model is updated, a new lightweight version is generated and distributed to the edge nodes to achieve strategy iteration, forming an adaptive optimization closed loop that coordinates cloud training and edge execution.
[0037] Figure 2 A flowchart of the second method of a land-sea relay aquaculture management method for large yellow croaker based on segmented aquaculture optimization is provided in an embodiment of the present invention; like Figure 2 As shown, the present invention provides a second method flowchart for a land-sea relay aquaculture management method for large yellow croaker based on segmented aquaculture optimization, including: S202 When the large yellow croaker in the independent breeding unit reaches the preset size for harvesting, the hyperspectral imager is used to collect the surface reflectance spectrum data of the fish and extract the characteristic absorption peak depth of carotenoids to calculate the surface color grade. The near-infrared spectral analyzer is used to collect the near-infrared absorption spectrum of the muscle and predict the fat content and elasticity coefficient. Multiple quality indicators are input into the quality grading model to classify the fish grade and obtain the quality test results. S204 synchronously obtains real-time price information from major trading markets through cloud market data interfaces, including the daily transaction price and weekly price trend of large yellow croaker of different specifications and grades. Combined with quality inspection results, it conducts market launch decision analysis to maximize unit output revenue. With quality grade, price prediction and storage costs as constraints, it uses a rolling time domain optimization algorithm to calculate the theoretical revenue value of each day's market launch within a set number of days in the future. S206: When the optimal market launch time window is determined to be the current date, the market launch execution process is triggered and the preliminary market launch plan is converted into a shipping instruction. The origin quarantine certificate and quality certification certificate are automatically generated through smart contracts and pushed to the sales platform to complete the transaction. When the optimal market launch time window is determined to be a future date, the breeding progress planning and management state is entered. S208, under the state of breeding progress planning and management, obtain the future cycle weight gain prediction sequence and body color quality evolution trajectory of all other independent breeding units, construct a multi-unit collaborative planning problem with the goal of maximizing the overall breeding benefits, and use a genetic algorithm to solve the feeding intensity adjustment coefficient, environmental temperature compensation value and light intervention scheme of each breeding unit. S210 converts the obtained feeding intensity adjustment coefficient, ambient temperature compensation value and light intervention scheme into control commands and sends them to the edge computing nodes of the corresponding breeding units. It dynamically adjusts the feeding curve of the automatic feeder, the target temperature curve of the circulating water system and the light period and spectral ratio of the LED light source to achieve staggered matching of the market time and price peak of different breeding units. S212 continuously monitors the actual growth response and market price changes of each breeding unit during the execution of the breeding progress plan, and re-optimizes the original plan every week. When abnormal fluctuations in market prices are detected, an emergency adjustment mechanism is triggered to re-solve the planning problem and update the control instructions, so as to achieve the orderly market launch of all breeding units in batches.
[0038] It should be noted that when the large yellow croakers in the independent breeding unit reach the preset harvest size, the pre-market quality inspection process is initiated. A hyperspectral imager collects the fish's surface reflectance spectral data, and the surface color grade is calculated by extracting the characteristic absorption peak depth of carotenoids. A near-infrared spectroscopy analyzer collects the muscle absorption spectrum to predict fat content and elasticity coefficient. Multiple quality indicators are input into a grading model (such as a support vector machine) to classify the fish into premium, first-grade, and second-grade categories, obtaining the quality inspection results. Simultaneously, real-time price information from major trading markets is obtained from cloud-based market data interfaces, including the daily transaction price and weekly price trends of large yellow croakers of different sizes and grades. With the goal of maximizing unit output revenue, and constrained by quality grade, price prediction, and storage costs, a rolling time-domain optimization algorithm is used to calculate the theoretical daily market revenue within a set number of days. When the optimal market launch time window is determined to be the current date, the market launch execution process is triggered, and a smart contract automatically generates a certificate of origin and a quality certification, which is then pushed to the sales platform to complete the transaction.
[0039] When the optimal market launch time window is determined to be a future date, the system enters the breeding progress planning and management state. It acquires the predicted weight gain rate sequence and body color quality evolution trajectory for all other independent breeding units over future periods, constructing a multi-unit collaborative planning problem with the goal of maximizing overall breeding efficiency. A genetic algorithm is used to solve for the feeding intensity adjustment coefficient, environmental temperature compensation value, and light intervention scheme for each breeding unit, appropriately accelerating the growth rate of some units to fill the supply gap, and appropriately slowing the growth rate of others to match subsequent price peaks. The resulting control commands are sent to the edge computing nodes of the corresponding breeding units, dynamically adjusting the feeding curve of the automatic feeder, the target temperature curve of the circulating water system, and the photoperiod and spectral ratio of the LED light source to achieve staggered matching of the market launch time and price peak periods for different breeding units. During the execution of the breeding progress planning, the actual growth response and market price changes of each breeding unit are continuously monitored, and the market launch decision optimization model is re-run weekly to continuously revise the original plan. When abnormal market price fluctuations are detected, an emergency adjustment mechanism is triggered, the planning problem is re-solved, and the control commands are updated, ultimately achieving the orderly batch-by-batch market launch of all breeding units.
[0040] Figure 3 An embodiment of the present invention provides a large yellow croaker sea-land relay aquaculture management system 3 based on segmented aquaculture optimization. The system includes: a memory 301, a processor 302, and a communication interface 303. The memory 301 contains a large yellow croaker sea-land relay aquaculture management method program based on segmented aquaculture optimization. When the large yellow croaker sea-land relay aquaculture management method program based on segmented aquaculture optimization is executed by the processor 302, it implements the steps of the large yellow croaker sea-land relay aquaculture management method based on segmented aquaculture optimization as described above.
[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention 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 the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for managing the integrated land-sea aquaculture of large yellow croaker based on segmented aquaculture optimization, characterized in that, include: A multimodal sensing network is used to acquire aquaculture environment and biological response data and construct a multi-source heterogeneous dataset. Based on the multi-source heterogeneous dataset, a hybrid modeling method is used to perform aquaculture environment-biological coupling modeling to obtain a coupled digital twin of land-sea relay aquaculture. Based on the coupled digital twin, a growth-quality synergistic prediction model is constructed and trained using a physical information neural network as the basic architecture. After training, real-time aquaculture monitoring data is extracted and input through a multi-source heterogeneous dataset to generate growth-quality synergistic prediction information. A state transition function is established based on the growth-quality co-prediction information. Each breeding unit is regarded as an independent agent to establish a multi-agent collaborative learning framework. The coupled digital twin is combined with multiple rounds of repeated trial and error training to finally obtain a multi-agent breeding strategy model. The multi-agent aquaculture strategy model is lightweighted and deployed on the edge computing node of the aquaculture site. Based on the current aquaculture density and fish status of the independent aquaculture unit, the timing of graded transfer is automatically determined, and an aquaculture management optimization strategy including feeding regulation, environmental parameter optimization and transfer planning is generated. When the large yellow croakers in a certain independent aquaculture unit meet the standards for harvesting, quality testing is conducted and a market launch plan is generated based on the current market price. If it is determined that the current time window is not the optimal market launch window, the aquaculture progress planning and management are carried out based on the aquaculture status of other independent aquaculture units.
2. The method for managing the integrated land-sea aquaculture of large yellow croaker based on segmented aquaculture optimization according to claim 1, characterized in that, The process of acquiring aquaculture environment and biological response data through a multimodal sensing network and constructing a multi-source heterogeneous dataset specifically includes: Underwater autonomous cruising platforms are deployed in cage aquaculture areas and factory workshops. These platforms are equipped with hyperspectral imagers, microelectrode arrays, and acoustic Doppler current profilers, which are used to collect fish surface reflectance spectra, water vertical temperature gradients and dissolved oxygen micro-distribution, and three-dimensional flow field characteristics, respectively. Simultaneously deploy an automatic environmental DNA sampling device in the aquaculture water body. The automatic environmental DNA sampling device collects water samples at set time intervals and performs online amplification and detection through a microfluidic chip to obtain data on the abundance of pathogenic microorganisms and the concentration of stress hormone metabolites in the water body. After data collection is completed, the data is transmitted to the edge data aggregation node in real time. At the aggregation node, the original data is timestamped and outliers are removed. The color characteristics and swimming trajectories of fish groups are extracted from hyperspectral images. Environmental DNA data is converted into pathogen abundance levels and stress level indices, ultimately forming a multi-source heterogeneous dataset.
3. The method for managing the integrated land-sea aquaculture of large yellow croaker based on segmented aquaculture optimization according to claim 1, characterized in that, The method of using a hybrid modeling approach based on the multi-source heterogeneous dataset to perform aquaculture environment-biological coupling modeling, resulting in a coupled digital twin of land-sea relay aquaculture, specifically includes: Based on the multi-source heterogeneous dataset, water level, flow velocity and boundary temperature data for continuous time periods are extracted, and input into a computational fluid dynamics simulator to solve the Reynolds-averaged Navier-Stokes equations. Combined with the standard k-ε turbulence model, the aquaculture area is divided into hexahedral grid cells, and the flow velocity vector, temperature scalar and dissolved oxygen flux of each grid point in three-dimensional space are output. After post-processing, the water exchange rate and temperature stratification coefficient are obtained as hydrodynamic characteristic fields. The fish population location distribution data extracted from the multi-source heterogeneous dataset during the same period were spatially registered with the hydrodynamic feature field, and the maximum likelihood estimation was used to fit the fish population's preferred flow velocity range and avoidance temperature threshold. An intelligent agent model based on individual instincts is established, and each virtual fish is given the attributes of perception radius, swimming speed and feeding preference. The preferred flow rate range and avoidance temperature threshold are used as behavioral constraints. The obstacle avoidance, gathering and escape behavior of the fish group are described by the social force model, and the time-varying distribution density and feeding activity index of the fish group are calculated and output. The daily food intake, ambient temperature, and transcriptomic metabolic pathway features of fish were extracted from the multi-source heterogeneous dataset. A physiological growth sub-model based on a biochemical reaction network was constructed. The sub-model takes ambient temperature, food intake, and activity energy consumption as inputs and outputs physiological growth data including daily weight gain and body color change data. The hydrodynamic characteristic field, the time-varying distribution density of the fish school, the feeding activity index, and the physiological growth data are input into the Bayesian assimilation engine, and the assimilation correction is performed using the ensemble Kalman filter algorithm to obtain a coupled digital twin that synchronously maps the real-time state of the physical aquaculture system.
4. The method for managing the integrated land-sea aquaculture of large yellow croaker based on segmented aquaculture optimization according to claim 1, characterized in that, The growth-quality co-prediction model is constructed and trained based on the coupled digital twin using a physical information neural network architecture. After training, real-time aquaculture monitoring data is extracted from and input into a multi-source heterogeneous dataset to generate growth-quality co-prediction information, specifically including: Historical simulation data is obtained based on the coupled digital twin, including hourly hydrodynamic characteristic fields, time-varying distribution density of fish schools, feeding activity index, and daily weight gain and body color change records of each virtual fish. The actual observation data for the corresponding time period is extracted from the multi-source heterogeneous dataset. The simulation data and the actual observation data are aligned by time to form a training sample set. Each sample contains the environmental sequence of day t as input features and the average weight gain rate and body surface color value of day t+1 as prediction labels. A growth-quality co-prediction model is constructed using a physical information neural network as its basic architecture. The model is trained using the training sample set. During the training process, the bioenergetics equation of large yellow croaker growth is embedded as a physical constraint into the loss function. The ambient temperature, food intake, and current weight are extracted from the training samples. The theoretical weight gain rate is calculated using the bioenergetics equation as the theoretical value. The mean square error between the theoretical value and the neural network prediction value is used as the physical constraint term. The error between the neural network prediction value and the predicted label in the sample is used as the data-driven error term. The total loss function is formed by weighted summation of the physical constraint term and the total loss function. The weight coefficients of the total loss function are determined by the grid search method. An adaptive moment estimation optimizer is adopted and an initial learning rate is set. After the hidden layer, a multi-head attention layer is added to calculate the attention weight of each time step in the input feature sequence. Key influencing factors are highlighted by weighted summation. The model is repeatedly trained until the deviation between the output of the growth-quality co-prediction model and the preset prediction label is less than a preset threshold. The desired growth-quality co-prediction model is then output. After training, real-time aquaculture monitoring data is extracted from the multi-source heterogeneous dataset as model input. The real-time aquaculture monitoring data includes the hourly average water temperature, daily cumulative feed intake, fish activity index and current average weight for the consecutive t days before the current time. The predicted daily weight gain rate and body color of day t+1 are calculated and output through forward propagation, and finally the growth-quality co-prediction information is obtained.
5. The method for managing the integrated land-sea aquaculture of large yellow croaker based on segmented aquaculture optimization according to claim 1, characterized in that, The process involves establishing a state transition function based on the growth-quality co-prediction information, treating each farming unit as an independent agent to establish a multi-agent collaborative learning framework, and combining this with the coupled digital twin for multiple rounds of repeated trial and error training to finally obtain a multi-agent farming strategy model. Specifically, this includes: Acquire growth-quality co-prediction information, using the current environmental state and fish state of each aquaculture unit as input, and construct a state transition function by combining the weight gain rate and body color prediction value for the next t days. The environmental state includes water temperature, salinity, dissolved oxygen, ammonia nitrogen concentration and light intensity, and the fish state includes current average weight, body color value, feeding activity index and stocking density. Each net cage and each factory-style aquaculture pond in the aquaculture system is defined as an independent intelligent agent. Each intelligent agent can observe the local state information of its own unit and execute a set of actions, including adjusting the feeding amount of the automatic feeder, setting the target temperature of the circulating water system, adjusting the light intensity and the proportion of red light, and providing early warning prompts for transfer. A multi-agent collaborative learning framework is constructed using a centralized training and decentralized execution paradigm. In this framework, a policy network is built for each agent to generate actions based on current observations, and an action value network is built to evaluate the quality of the joint actions of all agents. Simultaneously, a reward function is constructed based on the weighted reward of growth rate, reward of feed efficiency, penalty of survival rate, reward of body color quality, and penalty of energy consumption cost. A market condition adjustment factor is introduced to dynamically adjust the weights of each item. During the price peak, the weights of growth rate and body color quality are increased, and during the price trough, the weights of feed efficiency and energy consumption cost are increased. The state transition function, policy network, value network, and reward function are embedded into a coupled digital twin to initiate multiple rounds of repeated trial and error training. At the beginning of each training round, the initial state is randomly sampled. Each agent selects an action and executes it according to the current policy network. The digital twin updates the state according to the state transition function and calculates the immediate reward. The experience trajectory is stored in the experience replay buffer. Randomly sample small batches of data from the experience replay buffer and update the parameters of the policy network and the value network. The value network update objective is to increase the value of the next state after the reward is discounted, and the policy network update objective is to maximize the value network's score for the current action. After iterative training, when the value network loss function converges and the cumulative reward of the action sequence output by the policy network in the digital twin reaches a preset threshold, the current policy network parameters are saved as the final multi-agent farming policy model.
6. The method for managing the integrated land-sea aquaculture of large yellow croaker based on segmented aquaculture optimization according to claim 1, characterized in that, The multi-agent aquaculture strategy model is lightweighted and deployed on edge computing nodes at the aquaculture site. Based on the current stocking density and fish status of independent aquaculture units, the timing of graded transfer is automatically determined, and an aquaculture management optimization strategy including feeding regulation, environmental parameter optimization, and transfer planning is generated. Specifically, this includes: A multi-agent farming strategy model that has been trained is obtained. The original strategy network is lightweighted using knowledge distillation technology. A student network is constructed as a lightweight strategy model. Distillation training is performed by minimizing the KL divergence between the output of the student network and the output of the original strategy network. After distillation, the student network parameters are packaged into an inference engine format file suitable for edge deployment, and deployed in the field control cabinet of each net cage aquaculture area and the distributed control unit of the factory workshop respectively; The edge computing node loads the deployed lightweight strategy model, collects real-time data from the sensor array at a fixed frequency, reads the environmental status and fish status of the current aquaculture unit in each control cycle, performs standardization processing, and inputs the data into the lightweight strategy model for forward inference calculation. The current aquaculture management optimization strategy is calculated and output through forward inference and converted into corresponding control signals and sent to each actuator to achieve real-time regulation. The aquaculture management optimization strategy includes feeding rate setpoint, water temperature setpoint, LED red light ratio adjustment value, and whether to activate transfer warning. The edge computing node determines the timing of transfer based on the transfer warning signal and the comparison between the current breeding density and the preset density threshold. When the transfer warning signal persists for n control cycles and the current breeding density exceeds the density threshold, it is determined that transfer is required, and a transfer planning scheme containing the percentage of fish to be transferred and the optimal transfer time window is generated. For independent aquaculture units where the transfer warning signal is not triggered or the aquaculture density is below the threshold, the edge node continuously executes feeding control and environmental parameter optimization strategies. After each control cycle, the current status, executed actions, and subsequent sensor feedback data are packaged to generate a decision log, which is then uploaded to the cloud digital twin platform via wireless network to adjust the original multi-agent strategy model.
7. The method for managing the integrated land-sea aquaculture of large yellow croaker based on segmented aquaculture optimization according to claim 1, characterized in that, When the large yellow croakers in a certain independent aquaculture unit meet the harvest standards, quality testing is conducted, and a market launch plan is generated based on the current market price. If it is determined that the current time window is not the optimal market launch window, then aquaculture progress planning and management are carried out based on the aquaculture status of other independent aquaculture units, specifically including: When the large yellow croakers in the independent breeding unit reach the preset size for harvesting, the surface reflectance spectrum data of the fish is collected by a hyperspectral imager and the depth of the characteristic absorption peak of carotenoids is extracted to calculate the color grade of the surface. The near-infrared absorption spectrum of the muscle is collected by a near-infrared spectrometer and the fat content and elasticity coefficient are predicted. Multiple quality indicators are input into the quality grading model to classify the fish grades and obtain the quality test results. Simultaneously obtain real-time price information from major trading markets through cloud market data interfaces, including the daily transaction price and weekly price trend of large yellow croaker of different specifications and grades. Combine the quality inspection results to conduct market launch decision analysis, aiming to maximize unit output revenue, and using quality grade, price prediction and storage costs as constraints, and adopting a rolling time domain optimization algorithm to calculate the theoretical revenue value of each day's market launch within a set number of days in the future. When the optimal listing time window is determined to be the current date, the listing execution process is triggered and the preliminary listing plan is converted into a shipping instruction. The origin quarantine certificate and quality certification certificate are automatically generated through smart contracts and pushed to the sales platform to complete the transaction. When the optimal listing time window is determined to be a future date, the breeding progress planning and management state is entered. Under the planning and management of breeding progress, the weight gain rate prediction sequence and body color quality evolution trajectory of all other independent breeding units in the future cycle are obtained. With the goal of maximizing the overall breeding benefits, a multi-unit collaborative planning problem is constructed. The genetic algorithm is used to solve the feeding intensity adjustment coefficient, environmental temperature compensation value and light intervention scheme of each breeding unit. The obtained feeding intensity adjustment coefficient, ambient temperature compensation value and light intervention scheme are converted into control commands and sent to the edge computing nodes of the corresponding breeding units. The feeding curve of the automatic feeder, the target temperature curve of the circulating water system and the light period and spectral ratio of the LED light source are dynamically adjusted to achieve staggered matching of the market time and price peak of different breeding units. During the implementation of the breeding progress plan, the actual growth response and market price changes of each breeding unit are continuously monitored. The original plan is re-optimized every week. When abnormal fluctuations in market prices are detected, an emergency adjustment mechanism is triggered to re-solve the planning problem and update the control instructions, so as to achieve the orderly market launch of all breeding units in batches.
8. A large yellow croaker sea-land relay aquaculture management system based on segmented aquaculture optimization, characterized in that, The system includes: a memory, a processor, and a communication interface. The memory contains a program for a sea-land relay aquaculture management method for large yellow croaker based on segmented aquaculture optimization. When the program for sea-land relay aquaculture management method for large yellow croaker based on segmented aquaculture optimization is executed by the processor, it implements the steps of the sea-land relay aquaculture management method for large yellow croaker based on segmented aquaculture optimization as described in any one of claims 1-7.
