An intelligent decision-making system for a marine ranch based on self-game playing

Abstract

The present application relates to a kind of ocean ranching intelligent decision system based on self game, comprising: multi-modal data perception module, for collecting multi-modal data and constructing space-time correlation feature atlas by graph neural network;Large model decision hub module, for mapping multi-modal data to unified semantic space by cross-modal alignment technology, and variable correlation reasoning based on injected marine ecological causal diagram;Dynamic game optimization module, for taking ecological protection target and economic benefit target as game parties, and searching Pareto optimal solution;Virtual-real linkage execution module, for driving ocean ranching digital twin and physical equipment to perform corresponding operation according to decision instruction, and realizing dynamic updating of model parameters by federated learning.The present application can improve the yield of cultivation, reduce ecological disaster loss, while supporting carbon sink asset digitalization certification, providing ocean ranching with intelligent decision support with ecological adaptability and economic robustness.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and marine ecological aquaculture, and in particular to an intelligent decision-making system for marine ranches based on self-play. Background Technology

[0002] With the advancement of intelligent construction in marine ranching, multi-source data acquisition methods, such as satellite remote sensing, underwater sensors, and sonar detection, are becoming increasingly abundant. However, traditional data analysis methods rely on manual feature extraction and modality matching, making it difficult to effectively establish cross-temporal and cross-media correlation models, resulting in low efficiency in environmental monitoring and aquaculture decision-making. Simultaneously, there are significant contradictions between ecological goals such as maintaining carbon sequestration capacity and protecting biodiversity and production goals such as optimizing aquaculture density and improving economic benefits. Existing decision-making systems use fixed-weight rule engines, which cannot dynamically adjust optimization strategies based on real-time environmental changes, often leading to extreme decisions such as "protecting ecology at the expense of economic benefits" or "prioritizing production at the expense of the environment." For sudden disasters such as red tide outbreaks and typhoon attacks, current emergency decision-making mainly relies on historical experience and manual judgment, lacking support from predictive models trained on massive event databases. Especially in complex scenarios such as turbid waters and extreme weather, the response time from data acquisition to the generation of a response plan generally exceeds 6 hours, missing the golden window for disaster prevention and control. Summary of the Invention

[0003] The purpose of this invention is to at least address one of the shortcomings of the prior art and provide an intelligent decision-making system for marine ranching based on self-play.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] Specifically, a self-play-based intelligent decision-making system for marine ranching is proposed, including:

[0006] A multimodal data sensing module is used to collect multimodal data, including satellite remote sensing data, underwater sensor data, sonar signals and fishery market data, and to construct a spatiotemporal correlation feature map based on the multimodal data using a graph neural network.

[0007] The large model decision center module is built based on the pre-trained vertical domain large model OceanGPT. It is used to map multimodal data to a unified semantic space through cross-modal alignment technology, and further improve the spatiotemporal correlation feature map based on the injected marine ecological causal graph as prior knowledge to realize variable correlation reasoning. It generates a series of instruction sets to form a policy space through large model decision.

[0008] The dynamic game optimization module is used to treat ecological protection goals and economic benefit goals as two sides in a game, and search for Pareto optimal solutions from the strategy space as the final instruction set.

[0009] The virtual-physical linkage execution module connects the digital twin of the marine ranch with physical equipment. It is used to drive the digital twin of the marine ranch and physical equipment to perform corresponding operations according to decision commands, and realizes dynamic updating of parameters of the large model decision center module through federated learning.

[0010] Furthermore, specifically, based on the multimodal data, a spatiotemporal correlation feature map is constructed using a graph neural network, including:

[0011] The satellite pixel grid is divided according to a preset specification. The spatial proximity, temporal proximity, and statistical correlation between different grids are calculated to construct a spatiotemporal correlation diagram. The relevant formulas are as follows:

[0012]

[0013] in Representing spatial proximity, d(v) i ,v j ) represents node v i and v j The Euclidean distance between them, δ space Indicates spatial threshold; Indicating temporal proximity, t i and t j σ represents the timestamp of the node data. t Indicates the time decay coefficient; X represents statistical correlation. i and X j This represents the historical data sequence of a node, and Cov represents the calculation of the Pearson coefficient. and They represent variables X respectively i and X j The standard deviation of , where α, β, and γ represent weighting coefficients; A represents the adjacency matrix;

[0014] A spatiotemporal graph convolutional network ST-GCN is constructed to integrate spatiotemporal correlation features from satellite, UAV, underwater sensor, and market data to build a spatiotemporal correlation feature map. The relevant formulas are as follows:

[0015]

[0016] in This represents an adjacency matrix containing self-connections. The degree matrix nodes are used to record the number of neighbors for each node, l represents the current layer number of the graph convolutional network, L represents the total number of layers in the graph convolutional network, and H represents the number of nodes in the graph convolutional network. (l) H represents the feature matrix of the l-th layer. (0) W represents the initial features of the spatiotemporal correlation feature map. (l)Let H represent the learnable parameter matrix, and H represent the feature matrix fused from different layers.

[0017] Furthermore, specifically, when mapping multimodal data to a unified semantic space using cross-modal alignment techniques, the k-th modal feature is extracted as follows:

[0018] H k =f k (X k ;Θ k );

[0019]

[0020] Where f k Indicates a modal encoder, X k Θ represents the raw input data of the k-th mode. k f represents the dedicated encoder for the k-th mode. k The set of trainable parameters, H fused This represents the fused multimodal features, which have a total of K modes, α k represents the weight values ​​under different modalities, Q represents the query vector, and d represents the feature dimension.

[0021] Furthermore, specifically, the construction principles of the large model decision-making central module include,

[0022] Large-scale model pre-training: OceanGPT is initially trained using a corpus of hundreds of billions of tokens containing knowledge of marine physics, biochemistry and fisheries economics.

[0023] The objective function, achieved through the CLIP architecture, realizes cross-modal alignment between satellite image fragments and text descriptions, and constrains model parameters to conform to a predefined ecological causal chain using causal regularization constraints. This dual optimization of the objective function is expressed in the following formula:

[0024]

[0025] Where E (I,V) S(I,V) represents the average of all image-text pairs (I,V), where S(I,V) is the similarity score between image I and text V. ′ The text descriptions of other candidates are represented by τ, which is calculated by summing the similarity scores between the current image I and all candidate texts V, where τ is the temperature coefficient. This is a pre-defined cause-and-effect diagram of marine ecology. In the model prediction, variable X represents... i For X j The actual impact intensity, Δ ij X represents i For X j The prior value of causal strength, where λ is the regularization coefficient; Cross-modal alignment is achieved by enhancing the similarity of positive samples and reducing the similarity of negative samples. By constraining the difference between the two using the L2 norm, the causal relationships learned by the model are forced to align with pre-set ecological knowledge, thus avoiding the generation of decisions that violate scientific principles.

[0026] The CoT strategy generation algorithm is based on a large model to generate natural language policy descriptions. It takes multimodal satellite image fragments and text descriptions as input and outputs executable policies.

[0027] Furthermore, specifically, the construction principles of the dynamic game optimization module include,

[0028] The ecological utility function is constructed, and its formula is as follows:

[0029] U e =α1·CarbonCredit+α2·BiodiversityIndex;

[0030] α1 and α2 are dynamic weighting coefficients that are calculated and adjusted in real time through a large model; CarbonCredit represents carbon credits; and BiodiversityIndex represents the biodiversity index.

[0031] The economic utility function is constructed as follows:

[0032] U c =β1·NetProfit-β2·MarketRisk,

[0033] Where β1 and β2 are dynamic weighting coefficients, which are calculated and adjusted in real time through a large model; NetProfit represents net profit; MarketRisk represents the market risk index.

[0034] The Pareto optimality search employs an improved NSGA-II algorithm, with the fitness function as follows:

[0035] F(x) = [U e (x),U c (x)],

[0036] Where x represents the set of decision variables in the operation of marine ranching, and simultaneously satisfies the constraint condition g. i (x)≤0, different strategies will affect ecological and economic utility, so the NSGA-II algorithm is used to calculate the fitness of the decision set x;

[0037] When a disaster warning signal is detected, the weighting coefficients for ecological utility and economic utility are adjusted according to the following rules:

[0038] α(t)=α0·(1+k·e -t / τ ),

[0039] Where t is the duration of the disaster, k and τ are the emergency response coefficients, and α0 is the preset initial weight coefficient.

[0040] Furthermore, specifically, the construction principles of the virtual-real linkage execution module include the following:

[0041] Digital twin pre-validation requires verifying the impact of optimal decision instructions in the final instruction set on the environment and organisms, including environmental response models and biological behavior models. The environmental response model describes pollutant concentrations. As the equation varies with space (i,j) and time t, since it is a complex nonlinear partial differential equation, analytical solutions generally do not exist. Therefore, in the pre-verification of the digital twin, numerical methods are used to solve it, and the relevant formulas are as follows:

[0042]

[0043] Where C is the pollutant concentration, D is the diffusion coefficient, v is the flow velocity field, and R is the biochemical reaction term driven by temperature T. It is the discretization of the diffusion term. It is the discretization of the convection term, v x and v y These represent the components of the velocity field in the x and y directions, respectively, with Δx and Δy being the step size of the computational grid in the x and y directions. By calculating the changing trend of pollutant concentration at the current time, the impact on the environment can be determined.

[0044] The biological behavior model is represented by the fish swarm movement equation. Since the differential equation cannot be solved analytically directly, an explicit Euler method is used for discretization to obtain the time-varying positions of individual fish in the swarm. The relevant formulas are as follows:

[0045]

[0046] Where x i To initialize the individual positions of fish in the school, γ1, γ2, and γ3 represent weight values, N(i) represents the set of neighboring fish of the i-th fish, S(·) is the environmental suitability function, and x feed Let x be the spatial coordinates of the baiting point. j The location of neighboring fish schools is represented by the last two terms, which indicate the attractiveness of baiting and the schooling effect, respectively. The direction of fish movement is determined by environmental suitability, baiting attractiveness, and schooling effect, thus assessing the impact on biological behavior. Δt is the time step, where t is the distance between the fish and the school. k =kΔt is a time point, where k∈0,1,2,…

[0047] The feasibility grading mechanism uses the Eco-Economic Composite Index (EECI) output by the digital twin to categorize decisions.

[0048]

[0049] Where w1, w2, and w3 represent weight values, ΔB is the rate of change of biomass, B0 is the total biomass of the fish population at the initial time, and ρ(x,t) is the fish population density function calculated from the positions x of all individuals at the same time. i The statistical distribution is generated, where ΔP is the rate of change of profit, P0 is the economic profit at the initial moment, ΔRisk is the ecological risk increment, Risk0 is the risk value under natural conditions, and C safe This represents the critical safety concentration, where H is a step function. σ is 1 if it is true and 0 otherwise, and σ is the Sigmoid normalization function.

[0050] The grading rules are as follows:

[0051] EECI>0.7: Directly send the command to the physical device for execution;

[0052] 0.4≤EECI≤0.7: Trigger the re-optimization of the game theory module;

[0053] EECI < 0.4: Generate an alert and initiate a manual review process;

[0054] Federated learning is used to dynamically update the parameters of the decision-making central module of the large model, adding a mechanism for handling new types of pollution events. The loss function for this part is:

[0055]

[0056] in, Let μ be the loss function for predicting novel pollution events, and p be the catastrophic forgetting inhibition coefficient. new and p old These represent the distributions of the new and old data, respectively. KL represents the KL divergence, which is used to calculate the difference between the probability distributions output by the model on the new and old data.

[0057] The beneficial effects of this invention are as follows:

[0058] This invention proposes an intelligent decision-making system for marine ranching based on self-play.

[0059] 1) Establish an interpretable correlation between environmental parameters and biological behavior using a multimodal data representation learning framework based on causal reasoning;

[0060] 2) A large model thinking chain is used to generate a strategy space, and Pareto optimal decision search is achieved through self-play;

[0061] 3) Integrating the diffusion model with the dynamic evolution mechanism of federated learning to generate extreme event response plans and ensure data privacy.

[0062] This invention can increase aquaculture yield by 15%-20%, reduce ecological disaster losses by more than 40%, and support digital certification of carbon sink assets, providing intelligent decision support for marine ranches that combines ecological adaptability and economic robustness. Attached Figure Description

[0063] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments illustrated in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings:

[0064] Figure 1 The diagram shown is a structural principle diagram of an intelligent decision-making system for marine ranching based on self-play, according to the present invention. Detailed Implementation

[0065] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The same reference numerals used throughout the accompanying drawings indicate the same or similar parts.

[0066] Example 1, referring to Figure 1 This invention proposes an intelligent decision-making system for marine ranching based on self-play, comprising:

[0067] The multimodal data sensing module is used to collect multimodal data, including satellite remote sensing data (water quality parameters, surface temperature, seawater temperature, etc.), underwater sensor data (water temperature, ocean current speed and direction, dissolved oxygen content, tide and wave data, etc.), sonar signals (fish school location, quantity, etc.) and fishery market data. Based on the multimodal data, a spatiotemporal correlation feature map is constructed using a graph neural network.

[0068] The large model decision center module is built based on the pre-trained vertical domain large model OceanGPT. It is used to map multimodal data to a unified semantic space through cross-modal alignment technology, and further improve the spatiotemporal correlation feature map based on the injected marine ecological causal graph as prior knowledge (such as rising water temperature → increased algae reproduction → decreased dissolved oxygen), realize variable correlation reasoning, and generate a series of instruction sets to form a policy space through large model decision.

[0069] The dynamic game optimization module is used to treat ecological protection goals and economic benefit goals as two sides in a game, and search for Pareto optimal solutions from the strategy space as the final instruction set.

[0070] The virtual-physical linkage execution module connects the digital twin of the marine ranch with physical equipment. It is used to drive the digital twin and physical equipment of the marine ranch to perform corresponding operations according to decision commands. It also realizes the dynamic updating of parameters of the large model decision center module through federated learning, enabling the large model decision system and the digital twin to learn on their own.

[0071] In this embodiment 1, firstly, multimodal data (satellite remote sensing data, underwater sensor data, sonar signals, and fisheries market data) are collected through the data perception module, and a spatiotemporal correlation feature map is constructed as input data for the large model decision-making central module. Then, a series of decision instruction sets are obtained by variable correlation reasoning based on the marine ecological causal graph. In order to obtain the optimal decision instruction from the above instruction set, the optimal decision instruction is searched with the goal of ecological protection and economic benefits. Finally, the optimal decision instruction is input into the digital twin, and the feasibility of the current decision instruction is judged based on the impact generated by the digital simulation. After testing through the digital twin, it is input into the actual physical device. Federated learning is used to compare the effects of the two for dynamic updating of OceanGPT model parameters, realizing the self-learning of the large model decision system and the digital twin.

[0072] In a preferred embodiment of the present invention, specifically, a spatiotemporal correlation feature map is constructed based on the multimodal data using a graph neural network, and then spatiotemporal correlation features are extracted using a graph convolutional network, including:

[0073] The satellite pixel grid is divided according to a preset specification. The spatial proximity, temporal proximity, and statistical correlation between different grids are calculated to construct a spatiotemporal correlation diagram. The relevant formulas are as follows:

[0074]

[0075]

[0076] in Representing spatial proximity, d(v) i ,v j ) represents node v i and v j The Euclidean distance between them, δ space Indicates spatial threshold; Indicating temporal proximity, t i and t j σ represents the timestamp of the node data. t Indicates the time decay coefficient; X represents statistical correlation. i and Xj Represents node v i and v j The historical data sequence, where Cov represents the calculation of the Pearson coefficient. and They represent variables X respectively i and X j The standard deviation of , where α, β, and γ represent weighting coefficients; A represents the adjacency matrix;

[0077] A spatiotemporal graph convolutional network ST-GCN is constructed, which integrates the spatiotemporal correlation features of satellite, UAV, underwater sensor, and market data. The relevant formulas are as follows:

[0078]

[0079] in This represents the adjacency matrix containing self-connections I. The degree matrix nodes are used to record the number of neighbors for each node, l represents the current layer number of the graph convolutional network, L represents the total number of layers in the graph convolutional network, and H represents the number of nodes in the graph convolutional network. (l) H represents the feature matrix of the l-th layer. (0) W represents the initial features of the spatiotemporal correlation feature map. (l) Let H represent the learnable parameter matrix, and H represent the feature matrix fused from different layers.

[0080] In a preferred embodiment of the present invention, specifically, when mapping multimodal data to a unified semantic space using cross-modal alignment technology, the k-th modal feature is extracted as follows:

[0081] H k =f k (X k ;Θ k );

[0082]

[0083] Where f k Indicates a modal encoder, X k Θ represents the raw input data of the k-th mode. k f represents the dedicated encoder for the k-th mode. k The set of trainable parameters, H fused This represents the fused multimodal features, which have a total of K modes, α k represents the weight values ​​under different modalities, Q represents the query vector, and d represents the feature dimension.

[0084] As a preferred embodiment of the present invention, specifically, the construction principle of the large model decision center module includes,

[0085] Large-scale model pre-training: OceanGPT is initially trained using a corpus of hundreds of billions of tokens containing knowledge of marine physics, biochemistry and fisheries economics.

[0086] The objective function, achieved through the CLIP architecture, realizes cross-modal alignment between satellite image fragments and text descriptions, and constrains model parameters to conform to a predefined ecological causal chain using causal regularization constraints. This dual optimization of the objective function is expressed in the following formula:

[0087]

[0088] Where S(I,T) is the similarity score between image I and text T, and τ is the temperature coefficient; For the pre-defined marine ecological causal diagram, Δ ij X represents i For X j The prior value of causal strength, where λ is the regularization coefficient;

[0089] The CoT strategy generation algorithm is based on a large model to generate natural language policy descriptions. It takes multimodal satellite image fragments and text descriptions as input and outputs executable policies.

[0090] As a preferred embodiment of the present invention, specifically, the construction principle of the dynamic game optimization module includes,

[0091] The ecological utility function is constructed, and its formula is as follows:

[0092] U e =α1·CarbonCredit+α2·BiodiversityIndex;

[0093] α1 and α2 are dynamic weighting coefficients that are calculated and adjusted in real time through a large model; CarbonCredit represents carbon credits; and BiodiversityIndex represents the biodiversity index.

[0094] The economic utility function is constructed as follows:

[0095] U c =β1·NetProfit-β2·MarketRisk,

[0096] Where β1 and β2 are dynamic weighting coefficients, which are calculated and adjusted in real time through a large model; NetProfit represents net profit; MarketRisk represents the market risk index.

[0097] The Pareto optimality search employs an improved NSGA-II algorithm, with the fitness function as follows:

[0098] F(x) = [U e (x),Uc (x)],

[0099] Where x represents the set of decision variables in the operation of marine ranching, and simultaneously satisfies the constraint condition g. i (x)≤0, different strategies will affect ecological and economic utility, so the NSGA-II algorithm is used to calculate the fitness of the decision set x;

[0100] When a disaster warning signal is detected, the weighting coefficients for ecological utility and economic utility are adjusted according to the following rules:

[0101] α(t)=α0·(1+k·e -t / τ ),

[0102] Where t is the duration of the disaster, k and τ are the emergency response coefficients, and α0 is the preset initial weight coefficient. That is, α0 can refer to α1, α2, β1, and β2. In other words, α(t) represents the change of the above four parameters with the duration of the disaster.

[0103] As a preferred embodiment of the present invention, the construction principle of the virtual-real linkage execution module specifically includes the following:

[0104] Digital twin pre-validation requires verifying the impact of optimal decision instructions in the final instruction set on the environment and organisms, including environmental response models and biological behavior models. The environmental response model describes pollutant concentrations. As the equation varies with space (i,j) and time t, and because it is a complex nonlinear partial differential equation, an analytical solution generally does not exist. Therefore, numerical methods are typically used to solve it in digital twin pre-verification. The relevant formulas are as follows:

[0105]

[0106] Where C is the pollutant concentration, D is the diffusion coefficient, v is the flow velocity field, and R is the biochemical reaction term driven by temperature T. It is the discretization of the diffusion term. This involves discretizing the convection term, where Δx and Δy are the step sizes of the computational grid in the x and y directions, respectively. By calculating the changing trend of pollutant concentrations over the current time, the environmental impact can be determined.

[0107] The biological behavior model is represented by a fish swarm movement equation. Since the differential equation cannot be solved analytically directly, an explicit Euler method is used for discretization to obtain the time-varying positions of individual fish in the swarm. The relevant formulas are as follows:

[0108]

[0109] Where x i To initialize the individual positions of the fish in the school, S(·) is the environmental suitability function, xfeed Let x be the spatial coordinates of the baiting point. j The first term represents the location of neighboring fish schools, and the latter two terms represent feeding attractiveness and school aggregation effect, respectively. By considering environmental suitability, feeding attractiveness, and school aggregation effect, the direction of fish movement is determined, thereby assessing the impact on biological behavior. Δt represents the time step, where t... k =kΔt is a time point, where k∈0,1,2,….

[0110] The feasibility grading mechanism uses the Eco-Economic Composite Index (EECI) output by the digital twin to categorize decisions.

[0111]

[0112] Where ΔB is the rate of change of biomass, B0 is the total biomass of the fish population at the initial time, and ρ(x,t) is the fish population density function calculated from the positions of all individuals at the same time x. i The statistical distribution is generated, where ΔP is the rate of change of profit, P0 is the economic profit at the initial moment, ΔRisk is the increment of ecological risk, and Risk0 is the risk value under natural conditions. The solution depends on the partial differential equation C safe Represents the critical concentration for safety, where H is a step function, and C(x,t)>C safe σ is 1 if it is true and 0 otherwise, and σ is the Sigmoid normalization function.

[0113] The grading rules are as follows:

[0114] EECI>0.7: Directly send the command to the physical device for execution;

[0115] 0.4≤EECI≤0.7: Trigger the re-optimization of the game theory module;

[0116] EECI < 0.4: Generate an alert and initiate a manual review process;

[0117] The gradient aggregation formula and the local gradient update formula for federated learning are as follows:

[0118]

[0119] Where K is the number of participants, n k Let N be the k-th power of the data volume, σ be the differential privacy noise intensity, and η be the learning rate. Let be the local dataset of the kth power.

[0120] Furthermore, the reverse process of diffusion models generating extreme climate data follows:

[0121]

[0122] Where μθ and ∑ θ These are learnable parameters.

[0123] The loss function for fine-tuning novel pollution events is:

[0124]

[0125] Where μ is the catastrophic forgetting inhibition coefficient, p represents the loss function for the prediction task of novel pollution events. new and p old These represent the distributions of the new and old data, respectively. KL represents the KL divergence, which is used to calculate the difference between the probability distributions output by the model on the new and old data.

[0126] Although the description of the invention has been quite detailed and particularly of several described embodiments, it is not intended to limit it to any of these details or embodiments or any particular embodiment, but should be considered as providing a broad possible interpretation of the claims by referring to the appended claims and taking into account the prior art, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.

[0127] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention using the same means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.

Claims

1. A marine ranching intelligent decision-making system based on self-play, characterized in that, include: A multimodal data sensing module is used to collect multimodal data, including satellite remote sensing data, underwater sensor data, sonar signals and fishery market data, and to construct a spatiotemporal correlation feature map based on the multimodal data using a graph neural network. The large model decision center module is built based on the pre-trained vertical domain large model OceanGPT. It is used to map multimodal data to a unified semantic space through cross-modal alignment technology, and further improve the spatiotemporal correlation feature map based on the injected marine ecological causal graph as prior knowledge to realize variable correlation reasoning. It generates a series of instruction sets to form a policy space through large model decision. The dynamic game optimization module is used to treat ecological protection goals and economic benefit goals as two sides in a game, and search for Pareto optimal solutions from the strategy space as the final instruction set. The virtual-physical linkage execution module connects the digital twin of the marine ranch with physical devices. It is used to drive the digital twin of the marine ranch and physical devices to perform corresponding operations according to decision commands, and realizes the dynamic update of parameters of the large model decision center module through federated learning. Specifically, the construction principles of the virtual-real linkage execution module include the following: Digital twin pre-validation requires verifying the impact of optimal decision instructions in the final instruction set on the environment and organisms, including environmental response models and biological behavior models. The environmental response model describes pollutant concentrations. With space and time The change in , since the equation is a complex nonlinear partial differential equation, and an analytical solution does not exist in general, is therefore solved numerically in the digital twin pre-verification. The relevant formulas are as follows: ； ； in For pollutant concentration, The diffusion coefficient is... For the velocity field, For temperature Driven biochemical reaction term, It is the discretization of the diffusion term. It is the discretization of the convection term. and They represent the velocity field at... direction and Components in direction, and For computational grids in direction and The step size in the direction is used to calculate the changing trend of pollutant concentration over the current time, thereby determining the impact on the environment; The biological behavior model is represented by the fish swarm movement equation. Since the differential equation cannot be solved analytically directly, an explicit Euler method is used for discretization to obtain the time-varying positions of individual fish in the swarm. The relevant formulas are as follows: in To initialize the individual positions of the fish in the school, Represented as weight values, It indicates the first The fish's neighboring schools of fish gathered. Let be the environmental suitability function. The spatial coordinates of the baiting point. The first two values ​​represent the location of neighboring fish schools, while the last two represent the attractiveness of baiting and the schooling effect, respectively. By considering environmental suitability, baiting attractiveness, and schooling effect, the direction of fish movement can be determined, thereby assessing the impact on biological behavior. For time step, For time points, among which , The feasibility grading mechanism uses the Eco-Economic Composite Index (EECI) output by the digital twin to categorize decisions. ； ； ； ； in Represented as weight values, The rate of change in biomass. This represents the total biomass of the fish population at the initial moment. The fish density function is determined by the positions of all individuals at the same time. The statistical distribution is generated. For the rate of change of profit, The economic profit at the initial moment, This contributes to the increase in ecological risks. The risk value under natural conditions. Indicates the critical safety concentration. It is a step function. The value is 1 if the condition is met, and 0 otherwise. It is the Sigmoid normalization function; The grading rules are as follows: : Directly send the command to the physical device for execution; The game-playing module has been re-optimized. : Generate an alert and initiate a manual review process; Federated learning is used to dynamically update the parameters of the decision-making central module of the large model, adding a mechanism for handling new types of pollution events. The loss function for this part is: ； in, The loss function for predicting novel pollution events is represented. The coefficient of catastrophic forgetting inhibition. and The distributions of the old and new data are respectively. This represents the KL divergence, used to calculate the difference between the probability distributions output by the model on new and old data.

2. The intelligent decision-making system for marine ranching based on self-play as described in claim 1, characterized in that, Specifically, a spatiotemporal correlation feature map is constructed based on the multimodal data using a graph neural network. include, The satellite pixel grid is divided according to a preset specification. The spatial proximity, temporal proximity, and statistical correlation between different grids are calculated to construct a spatiotemporal correlation diagram. The relevant formulas are as follows: ； ； ； ； ； in Indicates spatial proximity, Represents a node and The Euclidean distance between them Indicates spatial threshold; Indicates temporal proximity. and The timestamp representing the node data. Indicates the time decay coefficient; Indicates statistical correlation. and Represents the historical data sequence of a node. This indicates the calculation of the Pearson coefficient. and Representing variables respectively and standard deviation and Indicates the weighting coefficient; Represents the adjacency matrix; A spatiotemporal graph convolutional network ST-GCN is constructed to integrate spatiotemporal correlation features from satellite, UAV, underwater sensor, and market data to build a spatiotemporal correlation feature map. The relevant formulas are as follows: in This represents an adjacency matrix containing self-connections. The degree matrix nodes are used to record the number of neighbors for each node. This indicates the current layer number of the graph convolutional network. This represents the total number of layers in the graph convolutional network. Indicates the first The feature matrix of the layer, The initial features represent the spatiotemporal correlation feature map. Represents the learnable parameter matrix, This represents the feature matrix fused from different layers.

3. The intelligent decision-making system for marine ranching based on self-play as described in claim 1, characterized in that, Specifically, when mapping multimodal data to a unified semantic space using cross-modal alignment techniques, for the first... Each modal feature is extracted as follows: ； ； ； in This indicates a modal encoder. Indicates the first The original input data for each modality, Indicates the first Modal dedicated encoder The set of trainable parameters This represents the fused multi-mode set features, which total [number missing]. One modality, This represents the weight values ​​under different modes. Represents the query vector. Indicates the feature dimension.

4. The intelligent decision-making system for marine ranching based on self-play as described in claim 1, characterized in that, Specifically, the construction principles of the large model decision-making central module include: Large-scale model pre-training: OceanGPT is initially trained using a corpus of hundreds of billions of tokens containing knowledge of marine physics, biochemistry and fisheries economics. The objective function, achieved through the CLIP architecture, realizes cross-modal alignment between satellite image fragments and text descriptions, and constrains model parameters to conform to a predefined ecological causal chain using causal regularization constraints. This dual optimization of the objective function is expressed in the following formula: ， ， ， in This represents all image-text pairs. Take the average. For image With text Similarity score, The text descriptions representing other candidates are calculated by accumulating the current image. With all candidate texts Similarity score, Temperature coefficient; This is a pre-defined cause-and-effect diagram of marine ecology. Indicating variables in model prediction right The actual intensity of the impact, express right The prior value of causal strength, The regularization coefficient is used. Cross-modal alignment is achieved by enhancing the similarity of positive samples and reducing the similarity of negative samples. By constraining the difference between the two using the L2 norm, the causal relationships learned by the model are forced to align with pre-set ecological knowledge, thus avoiding the generation of decisions that violate scientific principles. The CoT strategy generation algorithm is based on a large model to generate natural language policy descriptions. It takes multimodal satellite image fragments and text descriptions as input and outputs executable policies.

5. The intelligent decision-making system for marine ranching based on self-play as described in claim 1, characterized in that, Specifically, the construction principles of the dynamic game optimization module include: The ecological utility function is constructed, and its formula is as follows: ； in and These are dynamic weighting coefficients, calculated and adjusted in real time using a large model. Indicates carbon credits; Indicates the biodiversity index; The economic utility function is constructed as follows: ， in and These are dynamic weighting coefficients, calculated and adjusted in real time using a large model. Represents net profit; Indicates market risk index; The Pareto optimality search employs an improved NSGA-II algorithm, with the fitness function as follows: ， in This represents the set of decision variables in the operation of a marine ranch, while simultaneously satisfying the constraints. Different strategies affect ecological and economic utility, therefore the NSGA-II algorithm is used to calculate the decision set. The fitness of; When a disaster warning signal is detected, the weighting coefficients for ecological utility and economic utility are adjusted according to the following rules: ， in For the duration of the disaster, and This is the emergency response coefficient. The preset initial weighting coefficients.

Images