Deep sea intelligent aquaculture control method, electronic device, medium and program product

Abstract

The application provides a deep-sea intelligent aquaculture control method, an electronic device, a medium and a program product, and relates to the technical field of aquaculture. The method comprises the following steps: receiving a standardized multi-modal data sequence; performing multi-modal fusion with an aquaculture causal knowledge graph as a prior constraint to generate an aquaculture digital state vector; performing state evolution through a state evolution model based on the aquaculture digital state vector and historical time series data to obtain an aquaculture state evolution result; performing multi-objective collaborative optimization through multiple groups of field agents according to a marine environment disturbance coefficient and the aquaculture state evolution result, and generating an aquaculture control strategy and hierarchical early warning information based on a risk threshold; issuing the aquaculture control strategy to an aquaculture execution terminal and outputting the hierarchical early warning information; receiving operation feedback data of the aquaculture execution terminal and aquaculture state change data after an execution control operation is performed, so as to realize continuous optimization of state evolution model parameters, the risk threshold and the aquaculture control strategy.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, specifically to a deep-sea intelligent aquaculture control method, electronic equipment, media, and program products. Background Technology

[0002] As the aquaculture industry expands into deep seas, cage aquaculture is being conducted further and further offshore, and the scale of aquaculture is continuously expanding. Compared to nearshore and land-based aquaculture, deep sea areas have larger waves and more complex and changeable sea conditions. The traditional management model that relies on manual inspections and single-point sensors is no longer suitable for the actual needs of large-scale deep-sea aquaculture.

[0003] The limitations of traditional aquaculture management models are mainly reflected in several aspects: First, the underwater environment is closed, and data on fish status, water quality parameters, and equipment operating conditions are collected from multiple sources in a scattered manner, making it difficult to form a unified status judgment. Most decisions are made based on the experience of management personnel, which has limited accuracy and timeliness. Second, the distance from the shore in deep-sea areas is far, and manual inspection and on-site treatment are costly and time-consuming. When faced with sudden risks such as hypoxia and disease, the response is often delayed, which can easily lead to aquaculture losses. Third, the management process does not form a complete closed loop, and the effect of the control strategy cannot be used to optimize the decision-making logic, making it difficult for the system to iterate autonomously.

[0004] In recent years, some AI-based aquaculture management solutions have emerged, but most are designed for nearshore or land-based aquaculture scenarios, exhibiting significant limitations in adaptability to deep-sea environments. On one hand, most solutions rely solely on single water quality parameters or video images for identification, and multimodal data is simply pieced together without considering the inherent transmission logic of the aquaculture process. This results in poor interpretability of state judgments and accuracy easily affected by environmental disturbances. On the other hand, existing solutions are mostly passive responses to the current state, lacking trend prediction capabilities and failing to proactively address slowly developing risks such as hypoxia and disease. Furthermore, decision-making logic is largely based on fixed rules, neglecting the unstable communication in deep-sea environments and lacking collaborative optimization of multiple objectives such as aquaculture efficiency, equipment wear and tear, and production safety, leading to insufficient engineering feasibility for the overall system. Therefore, there is an urgent need for a smart aquaculture management solution that can adapt to the complex environment of deep-sea areas and possesses state prediction and multi-objective collaborative capabilities. Summary of the Invention

[0005] This invention provides a method for controlling intelligent aquaculture in deep sea, electronic equipment, media, and program products.

[0006] According to one aspect of the present invention, a method for controlling deep-sea intelligent aquaculture is provided, comprising: receiving a standardized multimodal data sequence uploaded by an edge acquisition terminal; the standardized multimodal data sequence being obtained by the edge acquisition terminal after local preprocessing of multi-source monitoring data collected at the aquaculture site; performing multimodal fusion on the standardized multimodal data sequence using an aquaculture causal knowledge graph as a prior constraint to generate an aquaculture digital state vector; performing state evolution through a state evolution model based on the aquaculture digital state vector and historical time-series data composed of historical aquaculture digital state vectors to obtain a predicted aquaculture state evolution result covering multiple time scales; performing multi-objective collaborative optimization through multiple groups of domain agents based on the marine environmental disturbance coefficient and the aquaculture state evolution result, and generating an aquaculture control strategy and graded early warning information based on a risk threshold; distributing the aquaculture control strategy to an aquaculture execution terminal to drive the corresponding aquaculture equipment to complete control operations and outputting the graded early warning information; receiving operational feedback data and executing control operations from the aquaculture execution terminal. The data on subsequent changes in aquaculture status are used to continuously optimize the parameters of the state evolution model, risk thresholds, and aquaculture control strategies. Specifically, using an aquaculture causal knowledge graph as a prior constraint, multimodal fusion is performed on the standardized multimodal data sequence to generate an aquaculture digital state vector. This includes: based on the aquaculture causal knowledge graph, constructing a transmission correlation chain containing changes in fish behavior, feeding status, oxygen consumption levels, water quality parameters, and disease risk, based on the physiological logic and environmental transmission laws of aquaculture; extracting initial features of each data modality from the standardized multimodal data sequence; mapping each initial feature to the aquaculture causal knowledge graph to obtain the feature nodes corresponding to each initial feature; determining the causal relationships between each feature node according to the transmission correlation chain; and learning the feature weights of each feature node along the transmission correlation chain using a causal attention mechanism, weakening the feature weights of noisy features without causal relationships, completing the cross-modal feature weighted fusion of each data modality, and outputting the fused features to generate the aquaculture digital state vector.

[0007] According to at least one embodiment of the present invention, the deep-sea smart aquaculture control method includes local preprocessing, which includes: cleaning multi-source monitoring data to obtain multimodal time-series data, wherein the data cleaning includes: timestamp alignment, data missing completion, abnormal sampling point removal, multi-source sampling frequency unification, and invalid sensor channel removal.

[0008] According to at least one embodiment of the deep-sea intelligent aquaculture control method of the present invention, the local preprocessing further includes temporal local normalization processing, wherein the temporal local normalization processing includes: splitting the multimodal time-series data into visual modal data, acoustic modal data, and environmental modal data according to data modality; calculating the local mean and local variance of the visual modal data, the acoustic modal data, and the environmental modal data respectively using an exponential moving average algorithm; performing a time-by-time normalization operation on the visual modal data, the acoustic modal data, and the environmental modal data based on the local mean and the local variance; and combining the visual modal data, the acoustic modal data, and the environmental modal data after performing the time-by-time normalization operation to generate the standardized multimodal data sequence.

[0009] According to at least one embodiment of the present invention, the deep-sea intelligent aquaculture control method includes the following: fish health status, water environment status, aquaculture equipment status, real-time risk status, and short-term trend representation.

[0010] According to at least one embodiment of the present invention, a deep-sea intelligent aquaculture control method, based on the marine environmental disturbance coefficient and the evolution result of the aquaculture state, performs multi-objective collaborative optimization through multiple sets of built-in domain agents, and generates aquaculture control strategies and hierarchical early warning information based on risk thresholds. The method includes: generating corresponding local control strategies through each agent in the multiple sets of domain agents; using a multi-objective optimization algorithm, with the goal of maximizing overall aquaculture benefits, resolving conflicts and balancing weights among the local control strategies to obtain a globally optimal preliminary control strategy; and analyzing the aquaculture digital state vector to obtain real-time risk levels in each dimension. Risk probability; analyze the evolution results of the aquaculture state to obtain the predicted risk probability of each dimension; perform weighted summation of the real-time risk probabilities of each dimension to obtain the current risk weighted value; perform weighted summation of the predicted risk probabilities of each dimension to obtain the predicted risk weighted value; calculate the comprehensive risk value based on the marine environmental disturbance coefficient, the current risk weighted value, and the predicted risk weighted value; determine the risk level based on the comprehensive risk value and the risk threshold; generate corresponding graded early warning information based on the risk level, and simultaneously adaptively modify the preliminary control strategy based on the risk level to generate an aquaculture control strategy.

[0011] According to at least one embodiment of the present invention, a deep-sea intelligent aquaculture control method calculates a comprehensive risk value based on the marine environmental disturbance coefficient, the current risk weighting value, and the predicted risk weighting value, comprising: adding the current risk weighting value and the predicted risk weighting value, and multiplying the sum by the marine environmental disturbance coefficient to obtain the comprehensive risk value.

[0012] According to at least one embodiment of the present invention, the deep-sea intelligent aquaculture control method comprises multiple sets of domain intelligent agents, including: biological intelligent agents, environmental intelligent agents, equipment intelligent agents, economic intelligent agents, and safety intelligent agents. The biological intelligent agents are used to propose feeding and disease control strategies; the environmental intelligent agents are used to calculate the marine environmental disturbance coefficient and propose environmental adaptation and management suggestions; the equipment intelligent agents are used to propose equipment scheduling and maintenance suggestions; the economic intelligent agents are used to propose efficiency optimization suggestions; and the safety intelligent agents are used to propose emergency response and safety protection suggestions.

[0013] According to another aspect of the present invention, an electronic device is provided, comprising: a memory storing execution instructions; and a processor executing the execution instructions stored in the memory, such that the processor performs a deep-sea intelligent aquaculture control method according to any embodiment of the present invention.

[0014] According to another aspect of the present invention, a readable storage medium is provided, wherein the readable storage medium stores execution instructions, which, when executed by a processor, are used to implement the deep-sea intelligent aquaculture control method of any embodiment of the present invention.

[0015] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the deep-sea intelligent aquaculture control method of any embodiment of the present invention.

[0016] The present invention has the following beneficial effects: (1) By constructing a physiological and environmental transmission chain based on the aquaculture causal knowledge graph as a prior constraint, and combining it with the causal attention mechanism to perform cross-modal feature weighted fusion and noise reduction, the rationality and accuracy of multimodal fusion can be improved, the business interpretability of state judgment can be enhanced, and environmental disturbance interference under complex sea conditions can be effectively resisted.

[0017] (2) By predicting the state evolution at multiple time scales based on the digital state vector of aquaculture and historical time series data, it is possible to upgrade the management from passive response to active prediction, identify the slow-developing aquaculture risks such as hypoxia and disease in advance, and reserve sufficient time for treatment. (3) By using the marine environment disturbance coefficient and multiple domain agents to perform multi-objective collaborative optimization and generate strategies based on risk thresholds, it is possible to take into account multiple dimensions of objectives such as aquaculture survival, equipment life, input cost and production safety, and generate control strategies that meet actual aquaculture needs and have a safety net.

[0018] (4) By receiving operational feedback and status change data from the execution terminal, the model parameters and strategies can be continuously optimized, forming a complete self-optimization closed loop, enabling the system to have adaptive iteration capabilities to continuously improve long-term control accuracy. Attached Figure Description

[0019] The accompanying drawings illustrate exemplary embodiments of the invention and, together with the description thereof, serve to explain the principles of the invention. These drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification.

[0020] Figure 1 This is a flowchart illustrating a deep-sea intelligent aquaculture control method according to one embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating a time local normalization processing method according to an embodiment of the present invention.

[0022] Figure 3 This is a flowchart illustrating the method corresponding to step S2 according to an embodiment of the present invention.

[0023] Figure 4 This is a flowchart illustrating the method corresponding to step S4 according to an embodiment of the present invention.

[0024] Figure 5 This is a comparative simulation diagram of risk evolution of a cloud-edge collaborative system for deep-sea smart aquaculture according to one embodiment of the present invention.

[0025] Figure 6 This is a simulation comparison diagram of multimodal causal fusion ablation according to one embodiment of the present invention.

[0026] Figure 7 This is a schematic structural block diagram of an electronic device employing a processor-based hardware implementation according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings.

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. The technical solution of this invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] Figure 1A schematic diagram of the overall process of a deep-sea intelligent aquaculture control method according to one embodiment of the present invention is shown. Figure 1 The method shown includes steps S1 to S6. This method can be performed by an electronic device.

[0030] S1: Receives standardized multimodal data sequences uploaded by the edge acquisition terminal.

[0031] The standardized multimodal data sequence is obtained by local preprocessing of multi-source monitoring data collected at the aquaculture site by the edge acquisition terminal.

[0032] As one possible implementation, the edge acquisition terminal connects to five types of monitoring equipment, covering comprehensive data. These five types of monitoring equipment include an underwater video acquisition unit, an acoustic sensing unit, a marine environment sensing unit, a meteorological and sea condition monitoring unit, and an equipment status acquisition unit. For example, the underwater video acquisition unit uses underwater high-definition cameras deployed on the inner wall of the net cage to capture images of fish activity, used to identify fish swimming behavior, feeding status, and body surface characteristics. The acoustic sensing unit uses underwater acoustic transducers to collect underwater acoustic signals, used to help determine the distribution density and activity level of the fish school. The marine environment sensing unit deploys probes for dissolved oxygen, salinity, water temperature, and pH value, collecting core water quality parameters in real time. The meteorological and sea condition monitoring unit connects to the meteorological station and marine forecast data of the net cage platform to obtain information on on-site wind speed, wave height, ocean currents, and surface water temperature. The equipment status acquisition unit connects to the control systems of each aquaculture device, collecting equipment operating status, working parameters, and fault alarm information.

[0033] As one possible implementation, local preprocessing includes: cleaning the multi-source monitoring data to obtain multimodal time-series data. For example, the data cleaning process may include: (1) Timestamp alignment: using the local clock of the edge acquisition terminal as a reference, timestamp calibration is performed on the data collected by different monitoring devices to achieve time synchronization of multi-source monitoring data. (2) Data missing completion: missing data caused by short-term communication interruption is completed using linear interpolation to maintain the continuity of the data sequence. (3) Abnormal sampling point removal: abnormal values ​​and sudden noise points that exceed the physical range are removed to reduce the interference of invalid data on subsequent judgment. (4) Unification of multi-source sampling frequency: data sources with different sampling frequencies are resampled and unified to the standard sampling frequency. (5) Removal of invalid sensor channels: invalid sensor channels that have no data output for a long time are automatically identified to eliminate invalid computing power occupation.

[0034] As one possible implementation, local preprocessing also includes temporal local normalization. Exemplarily, the temporal local normalization process may include, for example,... Figure 2 Steps S210 to S240 are shown.

[0035] S210: Multimodal time series data is split into visual modal data, acoustic modal data, and environmental modal data according to data modality.

[0036] S220: The exponential moving average algorithm is used to calculate the local mean and local variance of the visual modal data, acoustic modal data, and environmental modal data, respectively.

[0037] S230: Perform time-by-time normalization on visual modal data, acoustic modal data, and environmental modal data based on local mean and local variance.

[0038] S240: Combine the visual modal data, acoustic modal data and environmental modal data after performing time-by-time normalization to generate a standardized multimodal data sequence.

[0039] The aforementioned time-local standardization process effectively eliminates dimensional differences and local environmental noise interference in multimodal data, highlighting the temporal evolution characteristics of each modality and providing a high-quality data foundation for subsequent causal multimodal fusion.

[0040] S2: Using the causal knowledge graph of aquaculture as a prior constraint, multimodal fusion is performed on the standardized multimodal data sequence to generate aquaculture digital state vector.

[0041] Regarding step S2, it may include... Figure 3 Steps S310 to S350 are shown.

[0042] S310: Based on the causal knowledge graph of aquaculture, and based on the physiological logic and environmental transmission laws of aquaculture, a transmission chain is constructed that includes changes in fish behavior, feeding status, oxygen consumption level, water quality parameters, and disease risk.

[0043] By constructing a transmission and correlation chain, the causal logic and influence weights between each modality can be clarified, serving as a prior constraint for multimodal fusion.

[0044] S320: Extract the initial features of each data mode in the standardized multimodal data sequence.

[0045] For example, the initial features may include fish behavior features corresponding to video frames, spectral features corresponding to acoustic signals, and time series features corresponding to water quality parameters.

[0046] S330: Map each initial feature to the aquaculture causal knowledge graph to obtain the feature nodes corresponding to each initial feature.

[0047] S340: Determine the causal relationships between each feature node based on the transmission chain.

[0048] It should be noted that in step S340, the relationship between feature nodes is determined based on the prior transmission chain, which replaces the neighbor association method based on data statistical correlation commonly used in traditional multimodal fusion. This forces the physical and physiological transmission logic of the aquaculture field into the model topology.

[0049] S350: Combining causal attention mechanism, it learns the feature weights of each feature node along the transmission correlation chain, weakens the feature weights of noisy features without causal correlation, completes cross-modal feature weighted fusion of each data modality, and outputs fused features to generate aquaculture digital state vector.

[0050] Through the multimodal fusion process under causal constraints in steps S310 to S350 above, data-driven feature fusion can be combined with the inherent transmission logic in the aquaculture field. This approach retains the powerful feature extraction capabilities of artificial intelligence models while effectively improving the business interpretability of state judgments (i.e., the system can clearly trace the upstream physiological or environmental changes from which the risk is transmitted), making the model more adaptable and resistant to environmental disturbances under complex sea conditions, and avoiding the "black box" misjudgment problem caused by traditional simple feature splicing.

[0051] As one possible implementation, the aquaculture digital state vector includes: fish health status, aquatic environment status, aquaculture equipment status, real-time risk status, and short-term trend representation. Fish health status encompasses indicators such as fish activity level, feeding intensity, and the percentage of fish with abnormal body surfaces; aquatic environment status includes core water quality parameters such as dissolved oxygen, salinity, water temperature, and pH; aquaculture equipment status includes indicators such as the operating conditions, remaining lifespan, and failure probability of each piece of equipment; real-time risk status includes the probability of diseases, hypoxia, and environmental anomalies at the current moment; and short-term trend representation includes the rate of change and development trend of each status indicator. The five dimensions of data are stored aligned with a unified timestamp, forming a complete aquaculture digital state space.

[0052] S3: Based on the aquaculture digital state vector and historical time series data composed of historical aquaculture digital state vectors, state evolution is carried out through a state evolution model to obtain the predicted aquaculture state evolution results covering multiple time scales.

[0053] As one possible implementation, the state evolution model takes the current aquaculture digital state vector as the input state at the current moment, and uses historical time-series data composed of aquaculture digital state vectors from multiple historical moments arranged in chronological order as the temporal context. It employs a temporal prediction architecture to perform multi-step state deduction and obtain aquaculture state evolution results covering multiple time scales.

[0054] For example, the evolution results of the farming state can cover the following three time scales: (1) 6-hour scale hypoxia risk prediction: Focus on predicting the trend of dissolved oxygen change and the probability of hypoxia risk in water bodies, so as to schedule the start-up and shutdown and power adjustment of oxygenation equipment in advance. (2) Disease development prediction on a 12-hour scale: Focus on predicting the development trend of diseases in fish populations and the spread of abnormalities on the body surface, so as to activate disease prevention and control plans in advance; (3) Prediction of feeding and growth status on a 24-hour scale: Focus on predicting changes in the feeding intensity and growth trend of fish populations to optimize feeding plans and reduce feed waste.

[0055] It should be noted that the state evolution model, based on the time-series prediction architecture, combines the transmission logic constraints of the aquaculture causal knowledge graph for prediction, so that the prediction results follow the physical and physiological transmission laws in the aquaculture field, avoiding the generation of prediction results without physical basis, thereby improving the accuracy and rationality of trend prediction at multiple time scales, and providing a forward-looking state input basis for subsequent multi-agent collaborative decision-making.

[0056] For example, the state evolution model can employ a temporal prediction architecture based on deep learning models such as Long Short-Term Memory Networks (LSTM), Gated Recurrent Units (GRU), or Temporal Transformers.

[0057] S4: Based on the marine environmental disturbance coefficient and the evolution of aquaculture status, multi-objective collaborative optimization is carried out through multiple groups of domain agents, and aquaculture control strategies and hierarchical early warning information are generated based on risk thresholds.

[0058] Regarding step S4, it may include, for example: Figure 4 Steps S410 to S470 are shown.

[0059] S410: Generate corresponding local control strategies through each agent in multiple groups of domain agents.

[0060] As one possible implementation, the multiple sets of domain-specific intelligent agents may include: biological intelligent agents, environmental intelligent agents, equipment intelligent agents, economic intelligent agents, and safety intelligent agents. Each intelligent agent, based on the evolution results of the aquaculture status, independently generates local control strategies for its corresponding domain. For example, the biological intelligent agent is responsible for assessing the health status of the fish population, judging its growth trend, and proposing strategies related to feeding and disease prevention; the environmental intelligent agent is responsible for analyzing the water environment and sea conditions, calculating the marine environmental disturbance coefficient, and proposing environmentally adapted management and control suggestions; the equipment intelligent agent is responsible for assessing the operating conditions, predicting the lifespan, and predicting the failure of aquaculture equipment, and proposing suggestions for equipment scheduling and maintenance; the economic intelligent agent is responsible for calculating aquaculture costs such as feed input and equipment wear and tear, and proposing suggestions for optimizing efficiency based on market conditions; and the safety intelligent agent is responsible for assessing production safety risks such as extreme weather and equipment failure, and proposing suggestions for emergency response and safety protection.

[0061] S420: Through a multi-objective optimization algorithm, with the optimal comprehensive benefits of aquaculture as the global optimization guide, conflict resolution and weight balancing are performed on various local control strategies to obtain the preliminary control strategy with global optimum.

[0062] The local control strategies output by each group of intelligent agents may have conflicting objectives. For example, increasing oxygen supply can reduce the risk of hypoxia, but it will increase equipment wear and energy consumption.

[0063] As one possible implementation method, with the overall optimization guided by the goal of maximizing the comprehensive benefits of aquaculture, the corresponding multi-objective optimization dimensions include: aquaculture survival rate, equipment lifespan, feed input costs, and production safety in extreme weather.

[0064] By using a multi-objective optimization algorithm, under the premise of ensuring the bottom line of aquaculture safety, the algorithm can balance and resolve the conflicts between the survival rate of aquaculture, equipment lifespan, feed input costs and production safety in extreme weather, thereby outputting a globally optimal preliminary control strategy. This provides a globally optimal decision-making basis for generating the final aquaculture control strategy by combining risk thresholds.

[0065] S430: Analyze the digital state vector of aquaculture to obtain the real-time risk probability of each dimension. Analyze the evolution results of aquaculture states to obtain the predicted risk probability of each dimension.

[0066] For example, the dimensions include four dimensions: disease, hypoxia, environmental abnormalities, and abnormal fish behavior.

[0067] S440: The real-time risk probabilities of each dimension are weighted and summed to obtain the current risk weighted value. The predicted risk probabilities of each dimension are weighted and summed to obtain the predicted risk weighted value.

[0068] As one possible implementation method, the weighting coefficients used in the weighted summation can be adaptively adjusted according to the species and growth stage of the farm to suit the risk preferences of different farming scenarios.

[0069] S450: Calculate the comprehensive risk value based on the marine environmental disturbance coefficient, the current risk weighting value, and the predicted risk weighting value.

[0070] The marine environmental disturbance coefficient can be calculated from data on wind and waves, ocean currents, thermocline, and extreme weather warnings. The more complex the sea conditions, the higher the marine environmental disturbance coefficient.

[0071] As one possible implementation method, the comprehensive risk value = environmental disturbance coefficient × (current risk weighted value + predicted risk weighted value). By introducing the environmental disturbance coefficient as a multiplier factor, the basic risk value can be dynamically amplified under severe sea conditions, thereby improving the system's sensitivity to extreme environments.

[0072] S460: Determine the risk level based on the comprehensive risk value and risk threshold.

[0073] As one possible implementation method, the risk level includes three levels: low, medium, and high. The current risk level can be determined by comparing the comprehensive risk value with the risk threshold.

[0074] S470: Generates corresponding graded early warning information based on risk level, and adaptively corrects the initial control strategy based on risk level to generate aquaculture control strategy.

[0075] For example, aquaculture control strategies may include four categories: adjustment of feeding amount and frequency, start-up and shutdown and power control of aeration equipment, adjustment of cage depth and location, and disease classification, early warning and treatment plans.

[0076] S5: Distribute the aquaculture control strategy to the aquaculture execution terminal to drive the corresponding aquaculture equipment to complete the control operation and output hierarchical early warning information.

[0077] For example, the aquaculture execution terminal can be deployed at the net cage site. The aquaculture equipment may include: an automatic feeding system, an aeration equipment system, an underwater inspection robot, and an aquaculture net cage control system. After receiving the aquaculture control strategy, the aquaculture execution terminal drives the corresponding aquaculture equipment to complete the control operations, specifically including: driving the automatic feeding system to adjust the feeding amount and frequency; driving the aeration equipment system to start and stop the equipment and control its power; driving the aquaculture net cage control system to adjust the depth and position of the net cage; and driving the underwater inspection robot to perform underwater inspection operations.

[0078] As one possible implementation method, the graded early warning information can be output from the cloud-based management and control platform to the human-machine interaction terminal or the on-site audible and visual alarm equipment to prompt on-site management personnel to carry out graded early warning and handling of diseases, extreme weather safety protection or emergency duty, so as to realize a closed-loop management and control system with human-machine collaboration.

[0079] S6: Receive operational feedback data from the aquaculture execution terminal and data on changes in aquaculture status after executing control operations, so as to continuously optimize the parameters of the state evolution model, risk thresholds, and aquaculture control strategies.

[0080] As one possible implementation, continuous optimization can be achieved through a dual-path feedback mechanism to adapt to scenarios with limited offshore communication and high real-time requirements in deep-sea areas. Path one is the edge-end rapid feedback path: data on changes in aquaculture status after executing control operations are first transmitted back to the edge acquisition terminal. This allows for rapid adjustment of the control parameters of the lightweight model built into the edge acquisition terminal to adapt to real-time environmental changes at the aquaculture site, ensuring basic control capabilities in scenarios with network outages or limited communication. Path two is the cloud-based global optimization path: after all operational feedback data and aquaculture status change data are synchronized to the cloud-based control platform, they are distributed to various functional modules. This enables iterative adjustments to multimodal fusion model parameters, dynamic adjustments to risk assessment thresholds, and adaptive updates to control strategy rules, thus forming a complete self-optimizing closed loop.

[0081] The technical effects of this invention will be verified through system simulation and comparative experiments.

[0082] 1. Overall system risk evolution and control scheduling simulation.

[0083] To further verify the effectiveness and superiority of the method in this embodiment, a comparative simulation analysis was conducted on complex sea conditions and sudden environmental disturbance scenarios simulated over a continuous 120-hour period. The simulation results are as follows: Figure 5 As shown. From Figure 5 It can be seen from this: (1) Comparison of risk control effectiveness: During the "complex sea state disturbance period" and the "communication restriction period", the traditional scheme, due to the use of single-point monitoring and fixed rules, has a delayed response, resulting in a lower comprehensive risk index. The risk level spiked to a high-risk zone; however, the method in this embodiment uses state evolution prediction to trigger early warnings, and combined with five-agent collaborative control, successfully manages the comprehensive risk indicators. Suppressing risks below the medium-to-low risk level significantly enhances the risk resistance of deep-sea aquaculture.

[0084] (2) Dynamic response of control scheduling: Figure 5 The lower part shows the dynamic control curves of the method in this embodiment at different time points for oxygenation intensity, feeding reduction, cage submersion, maintenance scheduling, and emergency response levels. The system can adaptively adjust the control intensity of each execution terminal when communication is restricted or sea conditions deteriorate, ensuring the stability of management and control and the safety of aquaculture in offshore scenarios.

[0085] 2. Comparison of prediction accuracy for state evolution across multiple time scales.

[0086] The system's built-in state evolution prediction unit was compared and evaluated with traditional time series prediction models (such as conventional LSTM and GRU) on three typical prediction time scales: 6-hour hypoxia risk, 12-hour disease development, and 24-hour feeding changes. The evaluation indicators were root mean square error (RMSE) and early warning accuracy (AUC / ACC). The results are shown in Table 1.

[0087] Table 1 - Performance Comparison of Multi-Timescale State Evolution Prediction Models

[0088] As can be seen from Table 1, after introducing the physical or physiological logical constraints of the aquaculture causal knowledge graph, the method of this embodiment can avoid unfounded numerical drift, and its accuracy and stability under long-span prediction are significantly better than the traditional pure data-driven model.

[0089] 3. Comparison of the comprehensive benefits of multi-agent collaborative control.

[0090] To verify the practical effect of multi-objective collaborative optimization of intelligent agents in five domains: "biology, environment, equipment, economy, and safety", the comprehensive operation and maintenance and economic indicators during the 120-hour simulation cycle were statistically analyzed and compared with the traditional single-objective / fixed rule control scheme. The results are shown in Table 2.

[0091] Table 2 - Simulation Comparison of Comprehensive Benefits of Multi-Agent Cooperative Control

[0092] As can be seen from Table 2, the multi-agent collaborative control adopted in this embodiment can not only significantly shorten the high-risk exposure time, but also effectively reduce oxygenation power consumption and feed waste through multi-objective conflict resolution, thereby improving the overall economic benefits of aquaculture.

[0093] 4. Multimodal causal fusion ablation experiment.

[0094] To verify the effects of the "aquaculture causal knowledge graph" and "causal attention mechanism" in this invention on improving multimodal feature fusion and system anti-interference capabilities, ablation comparison experiments were conducted under simulated sensor burst noise and underwater visual turbidity interference conditions of different levels. The experimental results are as follows: Figure 6 As shown. From Figure 6 It can be seen from this: (1) Single-modal and ordinary stitching schemes are susceptible to interference: The initial evaluation accuracy of single-modal data input (water quality only or video only) and ordinary multimodal feature stitching scheme without causal constraints is 82% and 88% respectively when there is no noise (interference intensity is 0); as the noise interference intensity increases (0 → 10), the state evaluation accuracy drops significantly. When the interference intensity rises to 10, the accuracy drops sharply to 48% and 62% respectively. This shows that the ordinary feature fusion model without domain propagation logic constraints is extremely sensitive to sudden noise and environmental disturbances.

[0095] (2) The causal fusion of the present invention has extremely high robustness: The present invention constructs a transmission constraint of "behavior → feeding → oxygen consumption → water quality → disease" based on the causal knowledge graph of aquaculture, and combines the causal attention mechanism to weight the features. The accuracy rate is as high as 93% when there is no noise; as the intensity of noise interference increases, the accuracy rate only slightly decreases to 90% due to the effective suppression of non-causal noise features by the causal attention mechanism, and remains at a high accuracy level of over 90% throughout the process.

[0096] In summary, the ablation experiment further confirms that the introduction of causal knowledge graphs and causal attention mechanisms not only improves the interpretability of multimodal fusion, but also significantly enhances the anti-interference ability and judgment robustness of the intelligent aquaculture system under complex and harsh sea conditions in the deep sea.

[0097] Compared with the prior art, the present invention has the following beneficial effects: (1) It has achieved an upgrade in management and control from passive response to proactive prediction. By constructing a digital state space for aquaculture and multi-timescale evolution prediction, it is possible to identify slowly developing aquaculture risks such as hypoxia and disease in advance, reserve sufficient time for treatment, and reduce aquaculture losses caused by sudden risks.

[0098] (2) Improved the rationality and accuracy of multimodal fusion. By introducing aquaculture causal knowledge graph as a prior constraint, the data-driven feature fusion is combined with the transmission logic in the aquaculture field. This not only retains the feature extraction capability of artificial intelligence, but also improves the interpretability of state judgment and has a stronger adaptability to environmental disturbances under complex sea conditions.

[0099] (3) Collaborative decision-making for multi-dimensional objectives has been achieved. A domain-specific intelligent agent architecture is adopted, taking into account the state of aquaculture organisms, environmental adaptation, equipment lifespan, input costs and production safety. Through multi-objective optimization, a control strategy that better meets the actual aquaculture needs is obtained, rather than a single-dimensional risk management.

[0100] (4) Adapt to real-world scenarios where offshore communication is limited in deep-sea areas. An edge-cloud collaborative deployment architecture is adopted, where the edge can autonomously complete basic management and control in the event of a network outage, while the cloud is responsible for global optimization and model iteration. This not only ensures the stability of the system's operation but also reduces the bandwidth pressure on offshore data transmission.

[0101] According to a further embodiment of the present invention, an electronic device is also provided for use as a cloud management platform. Figure 7 This diagram illustrates a schematic block diagram of an electronic device employing a processor-based hardware implementation according to an embodiment of the present invention. The hardware structure of the electronic device of the present invention can be implemented using a bus architecture. The bus architecture can include any number of interconnect buses and bridges, depending on the specific application and overall design constraints of the hardware. Bus 1100 connects various circuits including one or more processors 1200, memory 1300, and / or hardware modules. Bus 1100 can also connect various other circuits 1400 such as peripheral devices, voltage regulators, power management circuits, external antennas, etc. Bus 1100 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, only one connecting line is used in this figure, but this does not indicate that there is only one bus or one type of bus. The memory 1300 stores a computer program. When the processor 1200 executes the computer program, the processor 1200 can execute the deep-sea intelligent aquaculture control method of the above embodiments of the present invention.

[0102] The present invention also provides a readable storage medium storing a computer program, which, when executed by a processor, is used to implement the methods described above. A "readable storage medium" can be any means capable of containing, storing, communicating, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples of a readable storage medium include: an electrical connection (electronic device) having one or more wires, a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable read-only memory (CDROM), etc.

[0103] This invention also provides a computer program product. The method of this invention can be implemented wholly or partially through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented wholly or partially in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed, the processes or functions of this invention are performed wholly or partially.

[0104] Computer programs or instructions can be stored in a readable storage medium or transferred from one readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The readable storage medium can be any available medium capable of access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium can be a volatile or non-volatile storage medium, or it can include both volatile and non-volatile types of storage media.

[0105] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, or characteristic described in connection with that embodiment / mode or example is included in at least one embodiment / mode or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0110] Those skilled in the art should understand that the above embodiments are merely for illustrating the present invention and are not intended to limit the scope of the invention. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present invention.

Claims

1. A method for controlling intelligent aquaculture in deep-sea areas, characterized in that, include: The system receives a standardized multimodal data sequence uploaded by an edge acquisition terminal; the standardized multimodal data sequence is obtained by the edge acquisition terminal after local preprocessing of multi-source monitoring data collected at the aquaculture site. Using the causal knowledge graph of aquaculture as a prior constraint, multimodal fusion is performed on the standardized multimodal data sequence to generate a digital state vector for aquaculture. Based on the aforementioned aquaculture digital state vector and historical time-series data composed of historical aquaculture digital state vectors, state evolution is performed through a state evolution model to obtain predicted aquaculture state evolution results covering multiple time scales. Based on the marine environmental disturbance coefficient and the evolution results of the aquaculture state, multi-objective collaborative optimization is carried out through multiple groups of domain intelligent agents, and aquaculture control strategies and hierarchical early warning information are generated based on risk thresholds. The aquaculture control strategy is sent to the aquaculture execution terminal to drive the corresponding aquaculture equipment to complete the control operation and output the hierarchical early warning information; The system receives operational feedback data from the aquaculture execution terminal and data on changes in aquaculture status after executing control operations, in order to continuously optimize the parameters of the state evolution model, risk thresholds, and aquaculture control strategies. The process involves using a causal knowledge graph of aquaculture as a prior constraint to perform multimodal fusion on the standardized multimodal data sequence, generating a digital state vector for aquaculture. This includes: constructing a transmission chain based on the causal knowledge graph of aquaculture, incorporating changes in fish behavior, feeding status, oxygen consumption levels, water quality parameters, and disease risk, according to the physiological logic and environmental transmission patterns of aquaculture; extracting initial features from each data modality in the standardized multimodal data sequence; mapping each initial feature to the causal knowledge graph of aquaculture to obtain feature nodes corresponding to each initial feature; determining the causal relationships between feature nodes based on the transmission chain; and learning the feature weights of each feature node along the transmission chain using a causal attention mechanism, weakening the feature weights of noisy features without causal relationships, completing the cross-modal feature weighted fusion of each data modality, and outputting the fused features to generate the digital state vector for aquaculture.

2. The deep-sea intelligent aquaculture control method as described in claim 1, characterized in that, The local preprocessing includes: cleaning the multi-source monitoring data to obtain multimodal time-series data, wherein the data cleaning includes: timestamp alignment, data missing completion, removal of abnormal sampling points, unification of multi-source sampling frequencies, and removal of invalid sensing channels.

3. The deep-sea intelligent aquaculture control method as described in claim 2, characterized in that, The local preprocessing further includes temporal local normalization, wherein temporal local normalization includes: The multimodal time-series data is split into visual modal data, acoustic modal data, and environmental modal data according to data modality; The local mean and local variance of the visual modal data, the acoustic modal data, and the environmental modal data are calculated using an exponential moving average algorithm. Based on the local mean and the local variance, a time-by-time normalization operation is performed on the visual modal data, the acoustic modal data, and the environmental modal data; The visual modal data, acoustic modal data, and environmental modal data after performing the time-by-time normalization operation are combined to generate the standardized multimodal data sequence.

4. The deep-sea intelligent aquaculture control method as described in claim 1, characterized in that, The aquaculture digital state vector includes: fish health status, aquatic environment status, aquaculture equipment status, real-time risk status, and short-term trend representation.

5. The deep-sea intelligent aquaculture control method as described in claim 1, characterized in that, Based on the marine environmental disturbance coefficient and the evolution results of the aquaculture state, multi-objective collaborative optimization is performed through multiple built-in domain agents, and aquaculture control strategies and hierarchical early warning information are generated based on risk thresholds, including: Each agent in the multiple groups of domain agents generates a corresponding local control strategy. Using a multi-objective optimization algorithm, with the goal of maximizing the overall benefits of aquaculture, conflict resolution and weight balancing are performed on various local control strategies to obtain a preliminary control strategy that is globally optimal. The aquaculture digital state vector is analyzed to obtain the real-time risk probability of each dimension; the aquaculture state evolution result is analyzed to obtain the predicted risk probability of each dimension. The real-time risk probabilities of each dimension are weighted and summed to obtain the current risk weighted value; the predicted risk probabilities of each dimension are weighted and summed to obtain the predicted risk weighted value. The comprehensive risk value is calculated based on the marine environmental disturbance coefficient, the current risk weighting value, and the predicted risk weighting value. Based on the comprehensive risk value and risk threshold, the risk level is determined; Based on the risk level, corresponding graded early warning information is generated, and the preliminary control strategy is adaptively modified based on the risk level to generate an aquaculture control strategy.

6. The deep-sea intelligent aquaculture control method as described in claim 5, characterized in that, Based on the marine environmental disturbance coefficient, the current risk weighting value, and the predicted risk weighting value, a comprehensive risk value is calculated, including: The current risk weighted value is added to the predicted risk weighted value, and the sum is multiplied by the marine environmental disturbance coefficient to obtain the comprehensive risk value.

7. The deep-sea intelligent aquaculture control method as described in claim 5, characterized in that, The multiple domain intelligent agents include: biological intelligent agents, environmental intelligent agents, equipment intelligent agents, economic intelligent agents, and safety intelligent agents. Specifically, the biological intelligent agent is used to propose feeding and disease control strategies; the environmental intelligent agent is used to calculate the marine environmental disturbance coefficient and propose environmental adaptation and management suggestions; the equipment intelligent agent is used to propose equipment scheduling and maintenance suggestions; the economic intelligent agent is used to propose efficiency optimization suggestions; and the safety intelligent agent is used to propose emergency response and safety protection suggestions.

8. An electronic device, characterized in that, include: The memory stores execution instructions; as well as A processor that executes the execution instructions stored in the memory, causing the processor to perform the deep-sea intelligent aquaculture control method according to any one of claims 1 to 7.

9. A readable storage medium, characterized in that, The readable storage medium stores execution instructions, which, when executed by a processor, are used to implement the deep-sea intelligent aquaculture control method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep-sea intelligent aquaculture control method as described in any one of claims 1 to 7.

Images