Intelligent consulting system and method for aquaculture

Abstract

The invention provides an intelligent consulting system and method for aquaculture, and belongs to the technical field of artificial intelligence and aquaculture, and the system comprises a multi-modal input processing module, a context awareness inference engine, an expert cooperation platform, a self-adaptive optimization module, a predictive aquaculture management module, a knowledge graph construction module, and a continuous learning and knowledge iteration module. According to the intelligent consulting system and method for aquaculture, complex aquaculture problems can be understood and solved, accurate personalized suggestions are provided, potential risks are predicted and prevented, intelligent transmission and application of aquaculture knowledge are achieved, and powerful support is provided for improving the aquaculture efficiency and the success rate.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and aquaculture technology, and in particular relates to an aquaculture intelligent advisory system and method. Background Art

[0002] Aquaculture is a key component of global protein supply and a pillar industry in many coastal and inland regions. However, the industry currently faces numerous challenges, including uneven technology adoption, difficulty acquiring expertise, and complex risk management. With global climate change and increasing environmental pollution, aquaculture faces increasing uncertainties, placing higher demands on aquaculture technology and management.

[0003] The traditional aquaculture technical support model has the following main problems: First, professional knowledge is fragmented and difficult to access. Aquaculture involves a wide range of disciplines, including biology, water chemistry, ecology, and nutrition. Professional literature and experience are scattered across various channels, making it difficult for ordinary farmers to fully and systematically access and understand them. Furthermore, a large amount of valuable experience and knowledge exists implicitly in the minds of experienced aquaculture experts, lacking effective mechanisms for its extraction and transfer.

[0004] Secondly, untimely technical support makes it difficult to respond to emergencies. Emergency situations such as water quality changes and disease outbreaks during aquaculture often require a rapid response. However, under the traditional technical service model, limited expert resources cannot meet the real-time consultation needs of a large number of farmers, resulting in missed opportunities for intervention.

[0005] Third, there's a lack of personalized guidance. Optimal management practices vary significantly across regions, water bodies, and species, making universal technical guidelines difficult to adapt to diverse local farming conditions. Existing technical support systems generally lack the ability to provide customized advice tailored to farmers' specific circumstances.

[0006] Fourth, insufficient predictive capabilities hinder forward-looking management. The aquaculture environment is complex and ever-changing, and relying solely on current data often leads to a reactive approach. The lack of scientific forecasts of future environmental changes and aquaculture risks limits the foresight and proactive nature of aquaculture management.

[0007] Fifth, knowledge updates are slow. Aquaculture technology continues to evolve, with new research and methods constantly emerging. However, the traditional knowledge system has a long update cycle, making it difficult to integrate the latest research results and practical experience in a timely manner, resulting in lagging technology promotion.

[0008] The development of technologies such as artificial intelligence, big data, and the Internet of Things (IoT) has opened up new possibilities for addressing these challenges. Knowledge graphs can systematically organize expertise; multimodal interaction technologies enable convenient and efficient human-computer communication; deep learning-based reasoning models have the potential to simulate expert diagnosis and decision-making processes; and digital twin technology offers new approaches for predicting aquaculture environments. However, the market currently lacks comprehensive intelligent advisory systems specifically designed for aquaculture that integrate these advanced technologies. Summary of the Invention

[0009] The purpose of the present invention is to provide an intelligent advisory system and method for aquaculture. By constructing a professional knowledge map and combining multimodal interaction technology, it can realize the intelligent identification and solution of aquaculture problems; through contextual awareness technology, it can provide targeted suggestions in combination with environmental data; through a continuous learning mechanism, the system can be continuously optimized from the interaction; and through digital twin technology, it can provide predictive aquaculture management.

[0010] To achieve the above objectives, the present invention provides an intelligent aquaculture advisory system, comprising Multimodal input processing module, used to collect, analyze and digitally process user-side input and real-time data; user-side input includes breeding goals or breeding conditions; The context-aware reasoning engine is used to extract features from the implementation data and breeding targets processed by the multimodal input processing module, analyze the correlation between environmental data and breeding data, and obtain the initial reasoning results; The expert collaboration platform is used to absorb user operations collected by the multimodal input module, update the knowledge graph construction module through the continuous learning and knowledge iteration module, and assist the continuous learning and knowledge iteration module in knowledge conflict detection, knowledge integration and value assessment; The adaptive optimization module compares the consistency between user actions and system recommendations and, based on the consistency results, optimizes the model parameters of the core modules of the context-aware reasoning engine. It also adjusts the context-aware reasoning engine adaptively based on the region and season. It also caches the optimization and adjustment plans locally to ensure offline operation. The predictive farming management module is used to build a virtual farming environment model of the farming environment, predict future key farming indicators based on time series analysis, and automatically generate risk prevention and control and optimization management suggestions based on the prediction results; A personalized recommendation generation module, which uses reinforcement learning algorithms to generate implementation recommendations based on real-time collected data, knowledge graph data, and farming goals; The knowledge graph construction module provides a theoretical basis for the context-aware reasoning engine, the predictive farming management module, and the personalized recommendation generation module; The continuous learning and knowledge iteration module is used to obtain expert experience on the expert collaboration platform and use it to update the knowledge graph construction module; it is used to discover potential new knowledge points in the system or identify knowledge conflicts, and collaborate with the expert collaboration platform to provide authoritative verification and arbitration.

[0011] Preferably, the ontology modeling method adopted by the knowledge graph construction module includes the following steps: Divide the knowledge in the field of aquaculture into core categories, and set up sub-category structures under each core category; Define various entities based on core categories to establish core category attribute sets; Establish the relationship type between the core category and the core category attribute set to obtain the corresponding semantic relationship; Extract and integrate basic academic knowledge, applied knowledge, and expert experience from expert collaboration platforms to obtain professional knowledge; Taking core categories and core category attributes as knowledge graph nodes, the semantic relationship between core categories and core category attribute sets is analyzed. Based on the acquired professional knowledge, the connection relationship between knowledge graph nodes is established to form a first-level knowledge network. On the basis of the first-level knowledge network, the connection relationship between subclass structures is established to form a second-level knowledge network. A hierarchical design is used to construct general knowledge and regional specialized knowledge separately, and automatic fusion is achieved through association rules to balance the generality and specificity of knowledge, forming a knowledge graph with the first-level knowledge network and the second-level knowledge network.

[0012] Preferably, the multimodal input processing module is composed of the following units: Setting up a data acquisition unit, including user input at the user end and setting up acquisition equipment for real-time acquisition; An image processing unit is provided for processing images inputted by the user end and collected in real time; A video analysis unit is set up to process video streams from user-side input and real-time collected data, dynamically analyze the group behavior patterns of farmed organisms, and maintain video analysis performance under low-light conditions through an adaptive brightness enhancement algorithm; An audio processing unit is provided to input audio data from the user end and collect data in real time, extract acoustic features, and analyze the operating status of the aquaculture equipment and the sounds of abnormal aquatic animal activities; A text processing unit is set up to combine optical character recognition and natural language processing technologies to digitize and extract structured data from user-side input and real-time collected data, and convert unstructured text into structured data through named entity recognition and relationship extraction technologies; The multimodal fusion unit maps feature vectors from different modalities into a unified semantic space, enhancing the complementarity of information between modalities. At the same time, it uses a modality-specific confidence scoring mechanism to dynamically weight the contribution of each modality and output a unified semantic representation.

[0013] Preferably, the context-aware reasoning engine includes: Convert the data processed by the multimodal input processing module into a graph structure and align it with the knowledge graph nodes; Extract features from the graph structure, analyze the graph structure features and the association patterns of knowledge graph nodes, build causal chains for the graph structure features and knowledge graph nodes, and calculate weights; The causal chain combines the temporal and spatial information of the graph structure characteristics to construct a comprehensive state representation of the current breeding environment; Generate targeted solutions and action suggestions through reverse reasoning and template filling.

[0014] Preferably, the personalized suggestion generation module includes: A reinforcement learning algorithm is used to generate implementation recommendations based on user input, real-time collected data, and real-time graphs. Each implementation recommendation includes three levels of implementation details: the first level is core parameter settings, the second level is operational instructions, and the third level is expected results and risk assessment. The reinforcement learning algorithm adopts a policy optimization framework based on a deep Q network, uses real-time monitoring data and regional characteristics of the knowledge graph to model it as a Markov decision process, uses user input as the reward function, selects the best case in the breeding history of the knowledge graph, and forms an optimal strategy library for different situations by learning from the best cases. At the same time, the Monte Carlo tree search method is introduced to evaluate the long-term decision path to ensure the long-term effectiveness of the recommendations. The system updates the Q value from user feedback in real time and continuously optimizes the decision model.

[0015] Preferably, the continuous learning and knowledge iteration module includes: The user interaction records generated by the system, expert knowledge supplements from the expert collaboration platform, knowledge graphs, and real-time collected data are processed using a multi-stage pipeline architecture to form a knowledge base; The multi-stage pipeline architecture uses a multimodal input processing module for data cleaning and standardization, a context-aware reasoning engine for feature extraction and semantic annotation, and an expert collaboration platform for knowledge conflict detection, knowledge fusion, and value assessment. At the same time, the knowledge base is also equipped with local devices for a hierarchical storage architecture, which locally stores data processed by the multi-stage pipeline architecture; The knowledge base also has a central server that uses the data processed by the multi-stage pipeline architecture for real-time updates, supporting rollback functions based on time series or specific event sequences.

[0016] Preferably, the predictive farming management module includes: Perform spatial scanning of the farm to generate a three-dimensional terrain model, and use the data collected in real time by the multimodal input processing module to obtain the water quality parameters, meteorological parameters, and biological characteristic parameters of the farm; Using the three-dimensional terrain model and water quality parameters, meteorological parameters, and biological characteristic parameters, a coupled model of hydrological dynamics model, water quality change model, biological growth model, and meteorological impact model is constructed, namely the virtual aquaculture environment model, to achieve real-time mapping between virtual and real environments; The virtual aquaculture environment model uses a multi-scale fusion method to predict time series; The virtual farming environment model is optimized through a multi-objective decision-making algorithm, which takes the farming objectives input by the user as the optimization target and considers the constraints set by the user.

[0017] Preferably, the adaptive optimization module includes: When comparing the consistency between user actions and system recommendations, the core farming parameters output by the personalized recommendation generation module are automatically adjusted using a two-dimensional region-season matrix. The adjustment range is determined based on statistical analysis of historical data. At the same time, the warning thresholds of the predictive farming management module and the judgment thresholds of the context-aware reasoning engine are automatically updated based on the latest weather forecast data. The key parameters of each optimization, the graph neural network parameters of the context-aware reasoning engine, the reinforcement learning model weights of the personalized recommendation generation module, the time series prediction model parameters of the predictive breeding management module, and the updated content of the knowledge base are cached locally. The adaptive optimization module, predictive breeding management module, and continuous learning and knowledge iteration module all have local caching functions and automatically update data through a differentiated synchronization mechanism.

[0018] An aquaculture intelligent advisor method comprises the following steps: The multimodal input processing module collects user input and real-time data for analysis and digital processing; The context-aware reasoning engine extracts features from processed data and farming objectives, analyzes the correlation between the environment and farming data, and draws initial inference results, providing a preliminary basis for farming decisions. The expert collaboration platform absorbs user operations, updates the knowledge graph through continuous learning and knowledge iteration modules, assists in knowledge conflict detection, integration and value assessment, and ensures the accuracy and timeliness of knowledge; The adaptive optimization module compares user operations with system recommendations, optimizes context-aware inference engine model parameters, and adapts the inference engine to regional and seasonal conditions. It caches the optimization and adjustment plans locally to ensure offline operation and improve system adaptability and stability. The predictive farming management module builds a virtual farming environment model, predicts future key farming indicators based on time series analysis, and automatically generates risk prevention and control and optimization management suggestions to help users deal with potential problems in advance; The personalized suggestion generation module uses reinforcement learning algorithms to generate implementation suggestions based on real-time collected data, knowledge graph data, and farming goals to meet users' personalized needs and improve farming efficiency. The knowledge graph construction module provides a theoretical basis for other modules. The continuous learning and knowledge iteration module obtains expert experience to update the knowledge graph, collaboratively handles new knowledge points and knowledge conflicts, and ensures the continuous improvement and development of the knowledge graph.

[0019] Therefore, the present invention adopts the above-mentioned intelligent aquaculture advisory system and method, and the technical effects are as follows: First, this invention constructs a systematic aquaculture knowledge graph, integrating dispersed expertise and experience into a structured knowledge network. Using ontological modeling, it systematically covers key entities such as aquaculture species, diseases, environmental parameters, treatment options, and their complex relationships, making expertise more standardized, structured, and accessible, providing comprehensive and systematic knowledge support for aquaculture operators.

[0020] Secondly, this invention implements multimodal interaction and intelligent recognition capabilities, supporting multiple input formats such as images, videos, audio, and text. Users can intuitively describe problems by uploading images of the farming environment, videos of equipment operation, or farming records. The system can accurately identify farmed species, disease symptoms, and environmental anomalies, significantly lowering the threshold for technical communication and improving the accuracy of problem diagnosis.

[0021] Third, the present invention's context-aware reasoning engine possesses expert-like diagnostic and decision-making capabilities. Leveraging graph neural network technology, the system analyzes the complex relationships between environmental parameters, symptom manifestations, and historical data, simulating the thought process of an expert. It quickly completes the process from problem description to solution, providing farmers with timely and professional technical support.

[0022] Fourth, the present invention provides highly personalized farming recommendations, generating customized plans based on the user's farming history, regional characteristics, and farming goals. The system considers multiple farming aspects, including water quality management, feed feeding, and disease prevention and control, providing detailed parameter settings, operation steps, and expected results to help farmers develop scientific and reasonable farming strategies.

[0023] Fifth, the present invention possesses continuous learning and knowledge iteration capabilities, constantly absorbing new research findings, expert experience, and user feedback. Through a federated learning architecture, the system extracts valuable information from each interaction and updates the knowledge base, while protecting user data privacy, continuously enhancing the system's expertise.

[0024] Sixth, this invention enables predictive farming management. Using digital twin technology, it constructs a virtual model of the farming environment, predicting future changes in farming indicators and potential risks. This forward-looking management approach enables farmers to proactively prevent problems, optimize resource allocation, and improve farming efficiency and economic benefits.

[0025] Finally, the system incorporates an adaptive optimization mechanism and local caching capabilities, enabling continuous optimization of model parameters based on user feedback and farming results. It also automatically adjusts reference thresholds based on regional and seasonal variations, ensuring system adaptability in a variety of environments. Furthermore, the local caching mechanism ensures the system remains operational even in farms with unstable networks, meeting practical application requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A schematic diagram of an aquaculture intelligent advisory system according to the present invention; Figure 2 Schematic diagram of building modules for knowledge graph; Figure 3 Schematic diagram of the multimodal input processing module; Figure 4 A schematic diagram of a context-aware reasoning engine; Figure 5 Schematic diagram of the module for generating personalized recommendations; Figure 6 It is a module for continuous learning and knowledge iteration; Figure 7 It is a predictive farming management module; Figure 8 It is an adaptive optimization module. DETAILED DESCRIPTION

[0027] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0028] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0029] Example 1 like Figure 1As shown, an aquaculture intelligent advisory system includes three main functional modules: a multimodal input processing module, a context-aware reasoning engine, an expert collaboration platform, an adaptive optimization module, a predictive aquaculture management module, a knowledge graph construction module, and a continuous learning and knowledge iteration module. These modules are interconnected through data interaction interfaces to jointly complete the intelligent identification, analysis and resolution of aquaculture problems.

[0030] like Figure 2 As shown in the figure, the knowledge graph construction module is used to provide a theoretical basis for the context-aware reasoning engine, the predictive farming management module, and the personalized recommendation generation module; The ontology modeling method used in the knowledge graph construction module includes the following steps: Divide the knowledge in the field of aquaculture into core categories, and set up sub-category structures under each core category; Define various entities based on core categories to establish core category attribute sets; Establish the relationship type between the core category and the core category attribute set to obtain the corresponding semantic relationship; Extract and integrate basic academic knowledge, applied knowledge, and expert experience from expert collaboration platforms to obtain professional knowledge; Taking core categories and core category attributes as knowledge graph nodes, the semantic relationship between core categories and core category attribute sets is analyzed. Based on the acquired professional knowledge, the connection relationship between knowledge graph nodes is established to form a first-level knowledge network. On the basis of the first-level knowledge network, the connection relationship between subclass structures is established to form a second-level knowledge network. A hierarchical design is used to construct general knowledge and regional specialized knowledge separately, and automatic fusion is achieved through association rules to balance the generality and specificity of knowledge, forming a knowledge graph with the first-level knowledge network and the second-level knowledge network.

[0031] The knowledge graph construction module uses ontological modeling to systematically organize aquaculture knowledge. First, a top-level ontology framework is constructed, dividing aquaculture knowledge into five core categories: species classification, environmental parameters, disease pathology, treatment options, and aquaculture techniques. Taking species classification as an example, a three-level sub-class structure is established: species (e.g., fish, crustaceans), biological characteristics (e.g., growth cycle, feeding habits), and aquaculture characteristics (e.g., density requirements, economic value).

[0032] During the entity attribute definition phase, the system establishes standardized attribute sets for various entities. For example, the species entity for tilapia includes 15 basic attributes, including scientific name, distribution range, growth cycle, optimal water temperature, dissolved oxygen requirement, and pH tolerance range. The disease entity for streptococcal disease includes 12 attributes, including symptoms (eye bulging, body hemorrhages), pathogenic mechanism, transmission route, and susceptible stage. The environmental parameter entity for dissolved oxygen includes eight attributes, including parameter definition, measurement method, normal range, and danger threshold.

[0033] The relationship type construction defines 20 semantic relationships, including "cause" (e.g., low dissolved oxygen causes hypoxia), "apply to" (e.g., chloramphenicol is applicable to bacterial sepsis), "affect" (e.g., water temperature affects food intake), "is a type of..." (e.g., grass carp is a type of carp), and "need" (e.g., shrimp seedling cultivation requires light regulation), etc., to describe complex associations between different entities.

[0034] Knowledge extraction and integration uses natural language processing technology to automatically extract professional knowledge from three types of knowledge sources: basic academic knowledge extracted from 50 professional textbooks such as "Aquaculture" and "Fish Diseases" and 500 core journal articles published by the Chinese Society of Fisheries; applied knowledge extracted from 300 technical manuals published by the National Fisheries Technology Extension Center; and experiential knowledge extracted from 50 senior aquaculture experts through structured interviews. The system uses named entity recognition and relationship extraction algorithms to convert unstructured text into entities and relationships in the knowledge graph.

[0035] The knowledge network structure covers 100 key aquaculture species in China's major aquaculture areas (including grass carp, tilapia, shrimp, etc.), 200 common diseases (including bacterial, viral, and parasitic diseases), 50 environmental parameters (water quality, bottom sediment, meteorology, etc.) and 300 treatment options (drug, physical, biological, etc. methods), forming a knowledge network with 10,200 entity nodes and 32,500 relationship edges, providing a solid knowledge foundation for the system's reasoning and suggestion generation.

[0036] like Figure 3 As shown, the multimodal input processing module is used to collect, analyze and digitally process user-side input and real-time collected data; user-side input includes breeding goals or breeding conditions; The multimodal input processing module consists of the following units: Setting up a data acquisition unit, including user input at the user end and setting up acquisition equipment for real-time acquisition; An image processing unit is provided for processing images inputted by the user end and collected in real time; A video analysis unit is set up to process video streams from user-side input and real-time collected data, dynamically analyze the group behavior patterns of farmed organisms, and maintain video analysis performance under low-light conditions through an adaptive brightness enhancement algorithm; An audio processing unit is provided to input audio data from the user end and collect data in real time, extract acoustic features, and analyze the operating status of the aquaculture equipment and the sounds of abnormal aquatic animal activities; A text processing unit is set up to combine optical character recognition and natural language processing technologies to digitize and extract structured data from user-side input and real-time collected data, and convert unstructured text into structured data through named entity recognition and relationship extraction technologies; The multimodal fusion unit maps feature vectors from different modalities into a unified semantic space, enhancing the complementarity of information between modalities. At the same time, it uses a modality-specific confidence scoring mechanism to dynamically weight the contribution of each modality and output a unified semantic representation.

[0037] The multimodal input processing module supports multiple input forms such as images, videos, audio and text, achieving all-round perception of the farming scene.

[0038] The image processing unit uses a deep convolutional neural network based on the ResNet-152 and EfficientNet-B7 architectures to process 3-megapixel (2048×1536) resolution images of aquaculture environments. For example, if a user uploads an image of a fish in a pond, the system can identify a grass carp and detect surface red spots and fin rot, initially diagnosing it as "red skin disease" with 95% accuracy. Using a five-level multi-scale feature pyramid network, the image processing unit simultaneously analyzes visual information from three areas: the water surface (to detect floating objects and oil film), the water body (to observe transparency and color), and the pond bottom (to identify sediment status), providing a basis for a comprehensive assessment of the aquaculture environment.

[0039] The video analysis unit supports high-frame-rate video streaming at 60 frames per second, enabling analysis of group behavior patterns of aquatic organisms. For example, by analyzing the swimming speed, clustering, and floating behavior of grass carp schools, the system can assess dissolved oxygen levels in the water. By monitoring changes in feeding activity, the system can assess feed palatability and fish health. The video analysis unit utilizes an architecture that combines a 3D convolutional neural network with a long-short-term memory network. This architecture maintains excellent analysis performance even in low-light conditions through an adaptive brightness enhancement algorithm.

[0040] The audio processing unit uses 16-bit quantization and a 48kHz sampling rate to capture sounds from the aquaculture environment, enabling it to identify equipment anomalies and behavioral changes in aquatic animals. For example, by analyzing sound signatures, the system can detect abnormal noise caused by an unbalanced aerator impeller, providing an early warning of equipment failure. It can also identify the distinctive slapping sounds made by certain fish when they are deprived of oxygen, serving as an early warning of water quality issues.

[0041] The text processing unit combines optical character recognition and natural language processing technologies to digitize user-submitted farming records. For example, if a user uploads a handwritten feeding record, the system automatically recognizes information such as the date, feed type, and feeding amount, converts it into structured data, and stores it in the system, providing a foundation for subsequent analysis. The text processing unit also extracts key information from unstructured descriptions. For example, "Yesterday morning, several fish were found floating in the northwest corner of the pond, with foam on the surface" can be parsed into a structured representation of location, time, and symptoms.

[0042] The multimodal fusion unit uses a Transformer-based cross-modal attention mechanism to integrate information from different modalities into a unified semantic representation. For example, when a user provides a pond image, a video of fish behavior, and a text description of water quality, the system can comprehensively analyze this information to form a complete understanding of the aquaculture situation. Even when certain modal information is missing, the system can still analyze it using the existing modalities, maintaining over 85% system performance.

[0043] like Figure 4 As shown, the context-aware reasoning engine is used to extract features from the implementation data and breeding targets processed by the multimodal input processing module, analyze the correlation between environmental data and breeding data, and obtain the initial reasoning results; The context-aware reasoning engine includes: Convert the data processed by the multimodal input processing module into a graph structure and align it with the knowledge graph nodes; Extract features from the graph structure, analyze the graph structure features and the association patterns of knowledge graph nodes, build causal chains for the graph structure features and knowledge graph nodes, and calculate weights; The causal chain combines the temporal and spatial information of the graph structure characteristics to construct a comprehensive state representation of the current breeding environment; Generate targeted solutions and action suggestions through reverse reasoning and template filling.

[0044] The context-aware reasoning engine is based on knowledge graphs and environmental data, and uses graph neural network technology to build reasoning models to achieve intelligent diagnosis and solution generation for farming problems.

[0045] The inference model uses a hierarchical graph neural network architecture, specifically designed to solve complex problems in aquaculture. For example, a user reported that their grass carp exhibited symptoms such as slow swimming, no feeding, and pale gills. The system combined these symptoms with environmental parameters (water temperature 24°C, dissolved oxygen 3.5 mg / L, pH 8.2) and, through reasoning, determined that the likely cause was hypoxia or bacterial gill disease, providing targeted treatment recommendations.

[0046] The five-layer structure of the graph convolutional network works together: the input encoding layer converts the symptoms and environmental data described by the user into a graph structure and aligns it with the relevant nodes in the knowledge graph; the feature extraction layer learns the correlation characteristics of symptoms, environmental parameters and possible diseases through graph convolution operations; the relational reasoning layer analyzes the causal chain between different entities, such as the reasoning path of low dissolved oxygen → stress response → decreased immunity → bacterial infection; the context integration layer combines spatiotemporal information such as recent meteorological changes (such as continuous high temperatures) to construct a more complete state representation; the solution generation layer generates a comprehensive plan based on the reasoning results, including emergency oxygenation, water quality improvement, and preventive medication.

[0047] The model's 256 feature channels enable it to capture complex knowledge representations in aquaculture: 64 channels represent the physiological characteristics, growth stages, and sensitive indicators of species like grass carp; 64 channels represent the normal ranges, trends, and interactions of environmental parameters such as water temperature, dissolved oxygen, and ammonia nitrogen; 64 channels encode the symptoms, pathogenesis, and transmission mechanisms of various diseases; and 64 channels represent the applicable conditions, operational methods, and expected effects of management measures such as oxygenation, water diversion, and medication. This high-dimensional feature representation ensures the model remains sensitive to subtle but critical changes in the aquaculture process.

[0048] Correlation analysis is implemented using a multi-head graph attention mechanism, with eight attention heads focusing on different types of correlation patterns. For example, when analyzing the problem of grass carp floating, different attention heads might focus on different relationships, such as the direct impact of ammonia nitrogen concentration, the synergistic effect of water temperature and dissolved oxygen, and the indirect impact of algae overgrowth. The system uses Shapley value analysis to identify insufficient dissolved oxygen (65%), excessive water temperature (23%), and elevated ammonia nitrogen (12%) as the three main contributing factors to the problem, and prioritizes these factors in the generated solution based on their impact.

[0049] The reasoning process uses a two-stage strategy to ensure timely response. For example, if a user reports shrimp in a shrimp pond turning red and experiencing decreased activity, the system will make a preliminary assessment within 300 milliseconds: "This may be a bacterial disease, and immediate water and shrimp testing is recommended." Simultaneously, deep reasoning is initiated, completing probabilistic analysis and differential diagnosis of specific diseases such as red body disease and hepatopancreatic necrosis within a further 700 milliseconds. Ultimately, it generates comprehensive recommendations, including emergency treatment, medication regimen, and subsequent management.

[0050] like Figure 5 As shown in the figure, the personalized suggestion generation module is used to generate implementation suggestions based on real-time collected data, knowledge graph data and breeding goals using reinforcement learning algorithms; The personalized suggestion generation module includes: A reinforcement learning algorithm is used to generate implementation recommendations based on user input, real-time collected data, and real-time graphs. Each implementation recommendation includes three levels of implementation details: the first level is core parameter settings, the second level is operational instructions, and the third level is expected results and risk assessment. The reinforcement learning algorithm adopts a policy optimization framework based on a deep Q network, uses real-time monitoring data and regional characteristics of the knowledge graph to model it as a Markov decision process, uses user input as the reward function, selects the best case in the breeding history of the knowledge graph, and forms an optimal strategy library for different situations by learning from the best cases. At the same time, the Monte Carlo tree search method is introduced to evaluate the long-term decision path to ensure the long-term effectiveness of the recommendations. The system updates the Q value from user feedback in real time and continuously optimizes the decision model.

[0051] The personalized suggestion generation system uses reinforcement learning algorithms to generate targeted suggestions based on the user's breeding history, regional characteristics and breeding goals, providing scientific guidance for the entire breeding process.

[0052] The system generates personalized recommendations based on data integration from four dimensions: analyzing the user's farming data from the past three years, for example, the average survival rate of tilapia of a southern farmer was 85%, the feed coefficient was 1.6, and the main disease was streptococcosis; considering regional characteristics, such as the long high temperature period in summer, heavy rainfall, and groundwater water source in the area; combining farming goals, such as planning to release the fish from the pond within 120 days, target size of 500 grams per tail, and control cost of 12 yuan per kilogram; integrating real-time monitoring data, such as the current water temperature of 28°C, pH value of 7.8, and dissolved oxygen of 5.6mg / L.

[0053] A reinforcement learning algorithm models the farming process as a Markov decision process. By studying over 10,000 successful cases, it has developed a library of optimal strategies for different scenarios. For example, the feeding strategy developed by the system for the farmer mentioned above takes into account the dissolved oxygen risks caused by high summer temperatures in southern China. It recommends feeding twice a day, in the early morning and evening, limiting the single feeding amount to no more than 60% of the total daily feeding amount, and appropriately reducing the feeding amount when rain is predicted. These strategies are evaluated for long-term effectiveness through Monte Carlo tree search, ensuring that growth rate is achieved without increasing disease risk.

[0054] The system provides parameterized guidance for 15 key farming steps, but dynamically adjusts priorities based on current circumstances. For example, in the early stages of farming, the focus is on water quality management and seedling cultivation; in the mid-stage of growth, emphasis is placed on feed feeding and disease prevention; and in the pre-harvest period, the focus is on growth regulation and economic benefit analysis. This dynamic adjustment mechanism ensures that the system's recommendations consistently meet the actual needs of the farming cycle.

[0055] The system employs a three-tiered response mechanism, tailored to the urgency and complexity of the inquiry. For example, when a user reports an urgent issue like "sudden mass mortality in a shrimp pond," the system generates a contingency plan within 5 seconds, including specific measures and steps, such as immediate full-pond aeration, emergency 20% water changes, and the use of live bottom-modifying bacteria. For routine inquiries like "how to improve feed conversion efficiency," the system generates standardized recommendations within 10 seconds, including parameters such as feed brand selection, feeding frequency (4-6 times / day), and feed particle size (3-5mm). For strategic inquiries like "how to plan the next season's breeding program," the system generates a comprehensive plan within 30 seconds, encompassing seed selection, stocking density, vaccinations, and phased feeding.

[0056] Each recommendation is provided with three levels of implementation details. For example, water quality adjustment recommendations have the following: the first level covers core parameter settings, such as "Add 50kg of quicklime per mu (approximately 1.5 meters) of water depth." The second level provides instructions for implementation, such as "Dissolve quicklime in clean water and evenly sprinkle along the edge of the pond, avoiding areas where fish gather. The application time is between 9 and 11 a.m." The third level provides expected results and risk assessments, such as "An increase in pH of 0.5-1.0 and alkalinity of 50 mg / L within 24 hours is expected, potentially leading to a temporary decrease in plankton. Supplementation with microbial agents is recommended after two days." This layered presentation ensures that users of all experience levels can find guidance tailored to their needs.

[0057] The visual interface features a responsive design for optimal performance across different devices. The system transforms complex parameter adjustment recommendations into intuitive charts, such as water quality parameter trend graphs (showing dissolved oxygen and pH changes over the past seven days and predicted for the next three days), feed dosage curves (displaying their relationship to fish growth stages), and disease risk heat maps (visually displaying disease risk probabilities under different environmental parameter combinations). These charts allow farmers to clearly understand the scientific basis and expected outcomes of the system's recommendations.

[0058] like Figure 6 As shown in the figure, the continuous learning and knowledge iteration module is used to obtain expert experience on the expert collaboration platform and use it to update the knowledge graph construction module; it is used to discover potential new knowledge points in the system or identify knowledge conflicts, and collaborate with the expert collaboration platform to provide authoritative verification and arbitration.

[0059] The continuous learning and knowledge iteration module includes: The user interaction records generated by the system, expert knowledge supplements from the expert collaboration platform, knowledge graphs, and real-time collected data are processed using a multi-stage pipeline architecture to form a knowledge base; The multi-stage pipeline architecture uses a multimodal input processing module for data cleaning and standardization, a context-aware reasoning engine for feature extraction and semantic annotation, and an expert collaboration platform for knowledge conflict detection, knowledge fusion, and value assessment. At the same time, the knowledge base is also equipped with local devices for a hierarchical storage architecture, which locally stores data processed by the multi-stage pipeline architecture; The knowledge base also has a central server that uses the data processed by the multi-stage pipeline architecture for real-time updates, supporting rollback functions based on time series or specific event sequences.

[0060] The continuous learning and knowledge iteration module adopts a federated learning architecture, which can extract valuable information from each interaction and update the knowledge base while protecting user data privacy, thereby achieving continuous evolution of system capabilities.

[0061] New knowledge is continuously acquired from five types of data streams: approximately 12,000 user interaction records are collected each month, such as farmers' inquiries on "how to improve the survival rate of whiteleg shrimp in high temperature seasons" and subsequent operational feedback; approximately 400 high-quality knowledge supplements are obtained through the expert collaboration platform, such as "How to use and precautions of new water quality stabilizers"; the latest research and technical reports from 40 journals such as the "Journal of Fisheries of China" and 15 industry websites such as the China Fisheries Portal are automatically crawled every week; real-time data uploaded by IoT devices are integrated, and a single farm generates approximately 2GB of raw data per day, including 24-hour water temperature, dissolved oxygen, and pH value change curves; regional epidemic and environmental monitoring data are collected, such as the "Epidemic Warning of Crucian Carp Hemorrhagic Disease in South China".

[0062] Data processing uses a multi-stage pipeline architecture: the data cleaning stage filters out water quality sensor outliers and incomplete user feedback; the feature extraction stage converts descriptive text such as "water quality significantly improved after feeding EM bacteria" into a structured relationship of "EM bacteria → water quality improved (effect = significant)"; knowledge conflict detection identifies a potential contradiction between the two recommendations for "feeding probiotics during high temperature periods," one recommending "once in the morning and once in the evening" and the other recommending "feeding only at night"; a knowledge fusion assessment determines that the latter is more applicable in high-temperature southern regions, with a score of 85, while the former is scored 65. The system retains the knowledge point with the higher score and records the applicable conditions.

[0063] The federated learning architecture ensures user data privacy and security. For example, a farm's environmental data and operational records are always stored locally. The system trains models locally and extracts abstract patterns and patterns, such as the finding that under the conditions of <region = Pearl River Delta, season = summer, and species = whiteleg shrimp>, reducing daily feed intake by 15% and increasing nighttime aeration time can increase survival rate by 8%. These abstract knowledge patterns are processed using differential privacy technology and uploaded to a central server. They are then aggregated and analyzed with other users' experience patterns to form more reliable knowledge patterns while ensuring that the specific data of individual farms is not leaked.

[0064] The knowledge database features a tiered storage architecture: the hot data layer stores frequently accessed knowledge, such as "Common Disease Prevention and Control During Summer High Temperatures," and the latest environmental data; the warm data layer stores the last 90 days of user inquiry history and feedback; the cold data layer stores complete historical data and large multimedia resources such as instructional videos; and the metadata layer maintains the knowledge graph's structural information and version change history. For example, when a user inquires about issues like "How to deal with whiteleg shrimp with skin failure," the system can retrieve and match knowledge within milliseconds, providing the latest solutions.

[0065] The knowledge base is updated through a hierarchical iterative strategy: When critical information, such as a new method for detecting highly pathogenic viruses, is discovered, the system completes an emergency update within 5 minutes. Regular incremental updates are performed every 24 hours, such as adding newly discovered "water quality-disease correlation" knowledge points to the graph. Knowledge graph structure optimization is performed every 30 days, such as restructuring the "Disease Prevention and Control" classification system to improve retrieval efficiency. The system automatically retains 30 days of version history. If, for example, a newly added water treatment method is found to have a side effect, the system can quickly roll back to the knowledge state before the problem was discovered, ensuring system reliability.

[0066] like Figure 7 As shown in the figure, the predictive farming management module is used to predict future key farming indicators based on time series analysis by building a virtual farming environment model of the farming environment, and automatically generate risk prevention and control and optimization management suggestions based on the prediction results; The predictive farming management module includes: Perform spatial scanning of the farm to generate a three-dimensional terrain model, and use the data collected in real time by the multimodal input processing module to obtain the water quality parameters, meteorological parameters, and biological characteristic parameters of the farm; Using the three-dimensional terrain model and water quality parameters, meteorological parameters, and biological characteristic parameters, a coupled model of hydrological dynamics model, water quality change model, biological growth model, and meteorological impact model is constructed, namely the virtual aquaculture environment model, to achieve real-time mapping between virtual and real environments; The virtual aquaculture environment model uses a multi-scale fusion method to predict time series; The virtual farming environment model is optimized through a multi-objective decision-making algorithm, which takes the farming objectives input by the user as the optimization target and considers the constraints set by the user.

[0067] The predictive farming management module is based on digital twin technology to build a virtual model of the farming environment. It predicts future key farming indicators through time series analysis, realizing the transition from passive response to active management.

[0068] Digital twin technology first uses lidar to perform high-precision spatial scanning of the farm, reconstructing the three-dimensional topography at a resolution of 0.1 meters. This includes the precise outline and depth distribution of the shrimp pond, which is 3 meters deep and covers an area of ​​5 mu. A sensor network is then deployed, including an underwater sensor array (monitoring parameters such as temperature, dissolved oxygen, and pH at different water layers), a weather station (recording factors such as air temperature, light intensity, and wind speed), and an aquatic biomonitoring system (tracking shrimp growth and behavior). Finally, a coupled model system is constructed, incorporating hydrological dynamics, water quality changes, biological growth, and meteorological influences. Data is updated every 5 minutes, synchronizing the virtual model with the actual aquaculture environment.

[0069] The virtual aquaculture environment model consists of five sub-models: the water environment sub-model displays a 3D heat map showing the distribution of dissolved oxygen levels below the critical value of 3.5 mg / L at 6:00 AM and simulates the effects of different aerator placement options. The meteorological impact sub-model predicts the risk of heavy rainfall over the next three days leading to a 0.8 drop in pH and a 25% increase in ammonia nitrogen. The biomass sub-model shows deviations from the standard growth curve of whiteleg shrimp and correlates this with decreased feeding activity. The feed dynamics sub-model calculates the sedimentation path and decomposition rate of released feed, identifying potential dead zones. The equipment operation sub-model monitors the 20% drop in efficiency of aerator No. 2 and simulates the potential impact of further failure on the aquatic environment. This virtual environment allows farmers to intuitively understand complex system dynamics and test different management strategies without disrupting actual production.

[0070] Time series analysis uses a multi-scale fusion approach for forecasting. For short-term forecasts (within 7 days), the LSTM deep learning model accurately predicted that "the lowest dissolved oxygen level of 2.8 mg / L will occur between 2:00 and 4:00 a.m. on Wednesday." For medium-term forecasts (within 14 days), wavelet decomposition and the random forest algorithm are combined to predict that "the water temperature will steadily rise to 32°C over the next 10 days, and the algal phase will undergo a transition." For long-term forecasts (within 30 days), a seasonal ARIMA model is used to predict that "the weight of aquacultured fish will reach 24g per fish by the end of this month, necessitating adjustments to feed particle size." The system achieved prediction accuracy rates of 92%, 88%, and 85%, respectively, for 12 indicators, including dissolved oxygen, pH, and ammonia nitrogen, providing a scientific basis for aquaculture management decisions.

[0071] The risk prevention and control mechanism enables proactive management based on threshold warnings and trend analysis. The system defines three risk levels: A "Caution" alert is triggered when the lowest dissolved oxygen value is predicted to drop to 3.5-4.0 mg / L within the next 48 hours, advising appropriate control of feeding rates; a "Warning" alert is triggered when the dissolved oxygen value is predicted to drop to 2.5-3.5 mg / L, advising increased nighttime aeration and the preparation of emergency equipment; and an "Emergency" alert is triggered when the dissolved oxygen value is predicted to fall below 2.5 mg / L, advising immediate implementation of comprehensive aeration and emergency water changes. The system identified the cumulative risk of slowly rising hydrogen sulfide concentrations in the aquaculture ponds and issued recommendations for bottom soil improvement 10 days before dangerous levels were reached, averting a crisis.

[0072] Optimized management recommendations are generated through a multi-objective decision-making algorithm, balancing economic benefits, resource consumption, and risk control. The system provides three types of management recommendations. For predicted high temperatures, the system generates preventative recommendations: "Starting next week, increase shade net coverage to 40% of the water surface and adjust daily feeding times to 6:00, 19:00, and 22:00, with feed intake ratios of 20%, 30%, and 50% each time." For slower-than-expected growth rates, the system provides adjustment recommendations: "Currently, protein intake is insufficient; it is recommended to increase feed protein content from 38% to 42%, and supplement with appropriate amounts of vitamin C to enhance stress resistance." For potential typhoons, the system prepares contingency plans: "12 hours before a typhoon arrives, reinforce aeration equipment, raise the water level by 30 cm, and prepare emergency generators to ensure continuous operation of critical equipment during power outages." Simulations show that adopting the system's recommended farming plan can increase yields by 15%, reduce feed costs by 8%, and reduce disease incidence by 23% compared to traditional management.

[0073] like Figure 8 As shown, the adaptive optimization module is used to compare the consistency between user operations and system recommendations, and optimize the model parameters of the context-aware reasoning engine based on the consistency results; it is used to adaptively adjust the context-aware reasoning engine according to the region and season; and it is used to locally cache the optimization and adjustment plans to ensure offline operation. The adaptive optimization module includes: When comparing the consistency between user actions and system recommendations, the system automatically adjusts the core farming parameters (including feeding amount, feeding frequency, water quality adjustment parameters, disease prevention parameters, etc.) output by the personalized recommendation generation module using a two-dimensional region-season matrix. The adjustment range is determined based on statistical analysis of historical data. At the same time, the warning thresholds of the predictive farming management module and the judgment thresholds of the context-aware reasoning engine are automatically updated based on the latest weather forecast data. The key parameters of each optimization (water quality management thresholds, feeding control parameters, disease warning parameters), core models (graph neural network parameters of the context-aware reasoning engine, reinforcement learning model weights of the personalized recommendation generation module, time series prediction model parameters of the predictive aquaculture management module) and knowledge base update content are cached locally. The adaptive optimization module, predictive aquaculture management module, continuous learning and knowledge iteration module all have local caching functions and automatically update data through a differentiated synchronization mechanism.

[0074] The adaptive optimization module can continuously optimize model parameters and knowledge base by comparing the differences between users' actual operation results and system recommendations, thereby achieving continuous evolution of system capabilities.

[0075] Actual user operation data is collected through three methods: Intelligent sensor monitoring, through IoT devices deployed in farms, automatically records actual user parameters. For example, a smart feeding machine records a feeding amount of 25 kg per feeding, while the system recommends 20 kg. Natural behavior analysis uses edge computing cameras to identify the actions of farmers at the pond. For example, it can detect that a user has actually performed water quality adjustments but used a drug dosage 30% higher than recommended. Farming results comparison tracks changes in key indicators. For example, if the system predicts that a shrimp pond should reduce ammonia nitrogen by 50% within 7 days after implementing a recommended solution, but the actual reduction is only 30%, indirectly indicating that the user may not have fully implemented the recommended bottom soil improvement measures. This data helps the system understand the discrepancy between actual user behavior and system recommendations, providing a basis for model optimization.

[0076] User feedback is obtained by combining implicit perception and active guidance: the biological computing interface uses the camera of the smart device to capture the changes in the user's micro-expressions when viewing the suggestions. For example, if the system recognizes that the user shows a puzzled expression about the dissolved oxygen management suggestion, it will automatically add more detailed scientific explanations and operation videos; the operational behavior mirror records the user's behavioral trajectory after receiving the suggestion. For example, a user spends a long time viewing the disease prevention and control suggestion, returns to view it multiple times, and implements it quickly, indicating a high acceptance of the suggestion; the closed-loop tracking of the results compares the expected effect with the actual result. For example, the water quality control plan recommended by the system is expected to increase transparency by 15cm, and the actual effect is 17cm. The value coefficient of 1.13 is calculated, indicating that the effect of the suggestion is better than expected.

[0077] The consistency assessment is based on a five-dimensional indicator system, comprehensively evaluating the actual implementation and effectiveness of the system's recommendations: consistency in operating parameters (the user actually used 45 kg of quicklime per mu, while the recommended amount was 50 kg / mu, with a consistency of 90%); consistency in operation timing (the user performed water quality adjustment two hours after the recommended time, with a timing consistency of 80%); consistency in operation method (the user followed the recommended method of evenly spraying along the edge of the pond, with a method consistency of 95%); consistency in effect (the actual pH value increased by 0.8, while the expected increase was 1.0, with a consistency of 80%); and consistency in user acceptance (the user adopted 12 of the system's 15 recommendations, with a consistency of 80%). The system calculated an overall consistency score of 85%, which is above the trigger threshold of 80%, indicating that the current model parameters are suitable and do not require immediate update. However, the system will still record any deviations for subsequent continuous optimization.

[0078] The threshold adjustment mechanism automatically adapts to regional and seasonal variations. To address regional adaptability, the system integrates multi-source data to build a regional characteristic database. For example, the optimal dissolved oxygen range in the southern subtropics is 4-8 mg / L, while in the northern temperate zone it is 5-10 mg / L. To address seasonal adaptability, the system establishes a phenological model for aquaculture. For example, the grass carp aquaculture cycle is divided into four phases: wintering (water temperature <15°C, dissolved oxygen threshold >6 mg / L), spring transition (water temperature 15-22°C, dissolved oxygen threshold >5 mg / L), peak growth (water temperature 22-30°C, dissolved oxygen threshold >4 mg / L), and autumn adaptation (water temperature 15-22°C, dissolved oxygen threshold >5 mg / L). The system automatically adjusts parameter thresholds based on the regional-seasonal matrix. For example, the safe dissolved oxygen threshold in summer can be 0.5-1.0 mg / L lower than in winter in the same region, ensuring that system recommendations always reflect local conditions.

[0079] The local caching mechanism, based on a layered edge computing architecture, ensures the system remains functional even in farms with unstable networks. At the hardware level, edge computing nodes are deployed at user terminals and farms, equipped with sufficient storage and computing resources. At the data level, a three-level caching strategy is employed to store critical environmental data, core reasoning models, and common knowledge bases. Functionally, core functions such as basic environmental monitoring, common problem diagnosis, and emergency response are prioritized. A synchronization mechanism ensures data consistency after network recovery, prioritizing locally modified data versions. This design enables the system to adapt to the complex and changing network environments of farms, providing uninterrupted technical support.

[0080] An aquaculture intelligent advisor method comprises the following steps: The multimodal input processing module collects user input and real-time data for analysis and digital processing; The context-aware reasoning engine extracts features from processed data and farming objectives, analyzes the correlation between the environment and farming data, and draws initial inference results, providing a preliminary basis for farming decisions. The expert collaboration platform absorbs user operations, updates the knowledge graph through continuous learning and knowledge iteration modules, assists in knowledge conflict detection, integration and value assessment, and ensures the accuracy and timeliness of knowledge; The adaptive optimization module compares user operations with system recommendations, optimizes context-aware inference engine model parameters, and adapts the inference engine to regional and seasonal conditions. It caches the optimization and adjustment plans locally to ensure offline operation and improve system adaptability and stability. The predictive farming management module builds a virtual farming environment model, predicts future key farming indicators based on time series analysis, and automatically generates risk prevention and control and optimization management suggestions to help users deal with potential problems in advance; The personalized suggestion generation module uses reinforcement learning algorithms to generate implementation suggestions based on real-time collected data, knowledge graph data, and farming goals to meet users' personalized needs and improve farming efficiency. The knowledge graph construction module provides a theoretical basis for other modules. The continuous learning and knowledge iteration module obtains expert experience to update the knowledge graph, collaboratively handles new knowledge points and knowledge conflicts, and ensures the continuous improvement and development of the knowledge graph.

[0081] Therefore, the present invention adopts the above-mentioned aquaculture intelligent advisory system and method, which can understand and solve complex aquaculture problems, provide accurate personalized suggestions, predict and prevent potential risks, realize the intelligent transmission and application of aquaculture knowledge, and provide strong support for improving aquaculture efficiency and success rate.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent aquaculture advisory system, characterized in that: include Multimodal input processing module, used to collect, analyze and digitally process user-side input and real-time data; user-side input includes breeding goals or breeding conditions; The context-aware reasoning engine is used to extract features from the implementation data and breeding targets processed by the multimodal input processing module, analyze the correlation between environmental data and breeding data, and obtain the initial reasoning results; The expert collaboration platform is used to absorb user operations collected by the multimodal input module and update the knowledge graph construction module through continuous learning and knowledge iteration modules; Used to assist the continuous learning and knowledge iteration modules in knowledge conflict detection, knowledge fusion and value assessment; The adaptive optimization module is used to compare the consistency between user operations and system suggestions, and optimize the model parameters of the core modules related to the context-aware reasoning engine based on the consistency results; Used to adaptively adjust the context-aware reasoning engine based on regional and seasonal conditions; used to locally cache optimization and adjustment plans to ensure offline operation; The predictive farming management module is used to build a virtual farming environment model, predict future key farming indicators based on time series analysis, and automatically generate risk prevention and control and optimization management suggestions based on the prediction results; A personalized recommendation generation module, which uses reinforcement learning algorithms to generate implementation recommendations based on real-time collected data, knowledge graph data, and farming goals; The knowledge graph construction module provides a theoretical basis for the context-aware reasoning engine, the predictive farming management module, and the personalized recommendation generation module; Continuous learning and knowledge iteration module, used to obtain expert experience on the expert collaboration platform and use it to update the knowledge graph construction module; It is used to discover potential new knowledge points in the system or identify knowledge conflicts, and collaborates with the expert collaboration platform to provide authoritative verification and arbitration.

2. The aquaculture intelligent advisory system according to claim 1, characterized in that: The ontology modeling method adopted by the knowledge graph construction module includes the following steps: Divide the knowledge in the field of aquaculture into core categories, and set up sub-category structures under each core category; Define various entities based on core categories to establish core category attribute sets; Establish the relationship type between the core category and the core category attribute set to obtain the corresponding semantic relationship; Extract and integrate basic academic knowledge, applied knowledge, and expert experience from expert collaboration platforms to obtain professional knowledge; Taking core categories and core category attributes as knowledge graph nodes, the semantic relationship between core categories and core category attribute sets is analyzed. Based on the acquired professional knowledge, the connection relationship between knowledge graph nodes is established to form a first-level knowledge network. On the basis of the first-level knowledge network, the connection relationship between subclass structures is established to form a second-level knowledge network. A hierarchical design is used to construct general knowledge and regional specialized knowledge separately, and automatic fusion is achieved through association rules to balance the generality and specificity of knowledge, forming a knowledge graph with the first-level knowledge network and the second-level knowledge network.

3. The aquaculture intelligent advisory system according to claim 1, characterized in that: The multimodal input processing module consists of the following units: Setting up a data acquisition unit, including user input at the user end and setting up acquisition equipment for real-time acquisition; An image processing unit is provided for processing images inputted by the user end and collected in real time; A video analysis unit is set up to process video streams from user-side input and real-time collected data, dynamically analyze the group behavior patterns of farmed organisms, and maintain video analysis performance under low-light conditions through an adaptive brightness enhancement algorithm; An audio processing unit is provided to input audio data from the user end and collect data in real time, extract acoustic features, and analyze the operating status of the aquaculture equipment and the sounds of abnormal aquatic animal activities; A text processing unit is set up to combine optical character recognition and natural language processing technologies to digitize and extract structured data from user-side input and real-time collected data, and convert unstructured text into structured data through named entity recognition and relationship extraction technologies; The multimodal fusion unit maps feature vectors from different modalities into a unified semantic space, enhancing the complementarity of information between modalities. At the same time, it uses a modality-specific confidence scoring mechanism to dynamically weight the contribution of each modality and output a unified semantic representation.

4. The aquaculture intelligent advisory system according to claim 1, characterized in that: The context-aware reasoning engine includes: Convert the data processed by the multimodal input processing module into a graph structure and align it with the knowledge graph nodes; Extract features from the graph structure, analyze the graph structure features and the association patterns of knowledge graph nodes, build causal chains for the graph structure features and knowledge graph nodes, and calculate weights; The causal chain combines the temporal and spatial information of the graph structure characteristics to construct a comprehensive state representation of the current breeding environment; Generate targeted solutions and action suggestions through reverse reasoning and template filling.

5. The aquaculture intelligent advisory system according to claim 1, characterized in that: The personalized suggestion generation module includes: A reinforcement learning algorithm is used to generate implementation recommendations based on user input, real-time collected data, and real-time graphs. Each implementation recommendation includes three levels of implementation details: the first level is core parameter settings, the second level is operational instructions, and the third level is expected results and risk assessment. The reinforcement learning algorithm adopts a policy optimization framework based on a deep Q network, uses real-time monitoring data and regional characteristics of the knowledge graph to model it as a Markov decision process, uses user input as the reward function, selects the best case in the breeding history of the knowledge graph, and forms an optimal strategy library for different situations by learning from the best cases. At the same time, the Monte Carlo tree search method is introduced to evaluate the long-term decision path to ensure the long-term effectiveness of the recommendations. The system updates the Q value from user feedback in real time and continuously optimizes the decision model.

6. The aquaculture intelligent advisory system according to claim 1, characterized in that: The continuous learning and knowledge iteration module includes: The user interaction records generated by the system, expert knowledge supplements from the expert collaboration platform, knowledge graphs, and real-time collected data are processed using a multi-stage pipeline architecture to form a knowledge base; The multi-stage pipeline architecture uses a multimodal input processing module for data cleaning and standardization, a context-aware reasoning engine for feature extraction and semantic annotation, and an expert collaboration platform for knowledge conflict detection, knowledge fusion, and value assessment. At the same time, the knowledge base is also equipped with local devices for a hierarchical storage architecture, which locally stores data processed by the multi-stage pipeline architecture; The knowledge base also has a central server that uses the data processed by the multi-stage pipeline architecture for real-time updates, supporting rollback functions based on time series or specific event sequences.

7. The aquaculture intelligent advisory system according to claim 1, characterized in that: The predictive farming management module includes: Perform spatial scanning of the farm to generate a three-dimensional terrain model, and use the data collected in real time by the multimodal input processing module to obtain the water quality parameters, meteorological parameters, and biological characteristic parameters of the farm; Using the three-dimensional terrain model and water quality parameters, meteorological parameters, and biological characteristic parameters, a coupled model of hydrological dynamics model, water quality change model, biological growth model, and meteorological impact model is constructed, namely the virtual aquaculture environment model, to achieve real-time mapping between virtual and real environments; The virtual aquaculture environment model uses a multi-scale fusion method to predict time series; The virtual farming environment model is optimized through a multi-objective decision-making algorithm, which takes the farming objectives input by the user as the optimization target and considers the constraints set by the user.

8. The aquaculture intelligent advisory system according to claim 1, characterized in that: The adaptive optimization module includes: When comparing the consistency between user actions and system recommendations, the core farming parameters output by the personalized recommendation generation module are automatically adjusted using a two-dimensional region-season matrix. The adjustment range is determined based on statistical analysis of historical data. At the same time, the warning thresholds of the predictive farming management module and the judgment thresholds of the context-aware reasoning engine are automatically updated based on the latest weather forecast data. The key parameters of each optimization, the graph neural network parameters of the context-aware reasoning engine, the reinforcement learning model weights of the personalized recommendation generation module, the time series prediction model parameters of the predictive breeding management module, and the updated content of the knowledge base are cached locally. The adaptive optimization module, predictive breeding management module, and continuous learning and knowledge iteration module all have local caching functions and automatically update data through a differentiated synchronization mechanism.

9. An intelligent aquaculture advisory method, characterized in that: The following steps are involved: The multimodal input processing module collects user input and real-time data for analysis and digital processing; The context-aware reasoning engine extracts features from processed data and farming objectives, analyzes the correlation between the environment and farming data, and draws initial inference results, providing a preliminary basis for farming decisions. The expert collaboration platform absorbs user operations, updates the knowledge graph through continuous learning and knowledge iteration modules, assists in knowledge conflict detection, integration and value assessment, and ensures the accuracy and timeliness of knowledge; The adaptive optimization module compares user operations with system recommendations, optimizes context-aware inference engine model parameters, and adapts the inference engine to regional and seasonal conditions. It caches the optimization and adjustment plans locally to ensure offline operation and improve system adaptability and stability. The predictive farming management module builds a virtual farming environment model, predicts future key farming indicators based on time series analysis, and automatically generates risk prevention and control and optimization management suggestions to help users deal with potential problems in advance; The personalized suggestion generation module uses reinforcement learning algorithms to generate implementation suggestions based on real-time collected data, knowledge graph data, and farming goals to meet users' personalized needs and improve farming efficiency. The knowledge graph construction module provides a theoretical basis for other modules. The continuous learning and knowledge iteration module obtains expert experience to update the knowledge graph, collaboratively handles new knowledge points and knowledge conflicts, and ensures the continuous improvement and development of the knowledge graph.

Images