AI-based intelligent control method and system for aquaculture environment

Abstract

This invention discloses an intelligent control method and system for aquaculture environments based on an AI model, belonging to the field of aquaculture management technology. The AI-based intelligent control method and system for aquaculture environments includes: S1, collecting water quality status data, feeding behavior data, and meteorological data from the aquaculture environment, preprocessing and storing them to construct an aquaculture control and management database; S2, constructing a time-series causal model of water body status based on multi-source time-series data and outputting a causal connection structure; S3, predicting and assessing the water quality evolution under different feeding schemes through multivariate time-series and causal coupling; S4, analyzing the feeding scheme objectives based on feeding decision quantities, water quality risks, and growth benefits; and S5, continuously correcting the causal relationship and feeding strategy through feeding execution feedback. This solves the problems of unclear bidirectional coupling between feeding behavior and water quality feedback, and the difficulty in coordinating the control of feeding decisions and water quality risks.

Description

Technical Field

[0001] This invention relates to the field of aquaculture management and control technology, specifically to an intelligent control method and system for aquaculture environment based on an AI model. Background Technology

[0002] Aquaculture, as an important form of agricultural production, relies heavily on the stability of the aquatic environment and the rationality of feeding management for its yield and quality. In recent years, monitoring and prediction schemes based on Internet of Things (IoT) sensor networks have emerged in the field of aquaculture environmental monitoring and control. These schemes utilize dissolved oxygen sensors, temperature sensors, pH sensors, ammonia nitrogen sensors, nitrite sensors, and meteorological data acquisition devices to construct online monitoring platforms. Furthermore, they explore the use of time series prediction models, Long Short-Term Memory (LSTM) networks, convolutional neural network structures, and reinforcement learning methods to predict changes in dissolved oxygen, water temperature, water quality levels, growth indicators, and to search for feeding strategies.

[0003] For example, the invention patent with announcement number CN118228908B discloses an Internet of Things (IoT)-based aquaculture environment monitoring system, belonging to the field of aquaculture management and control technology. Specifically, it is an IoT-based aquaculture environment monitoring system, including an IoT platform, a zone division module, a zone pollution level monitoring module, a zone comfort level monitoring module, a comprehensive aquaculture environment assessment module, and an aquaculture supervision terminal. This invention uses the zone pollution level monitoring module to monitor and analyze the water pollution status of each sub-region within the aquaculture area in real time and identify high-pollution and low-pollution zones. The zone comfort level monitoring module monitors and analyzes the aquaculture environment comfort status of low-pollution zones in real time and identifies high-comfort and low-comfort zones. The comprehensive aquaculture environment assessment module marks high-pollution and low-comfort zones as zones requiring adjustment and assesses the urgency of aquaculture environment regulation through an urgency analysis, ensuring the safe growth of aquatic products and reducing the management difficulty for supervisors. It has a high degree of intelligence.

[0004] For example, the invention patent with announcement number CN119130701B discloses a method and system for regionalized environmental monitoring of aquaculture based on data analysis. Specifically, the method includes: obtaining non-aquaculture periods in the aquaculture area; obtaining the aquatic plant growth rate and average growth duration based on these periods; selecting different levels of nutrient zones from the aquaculture area based on the aquatic plant growth rate and average growth duration; continuously monitoring and analyzing these different levels of nutrient zones; determining whether the aquaculture area is showing a trend of eutrophication based on the monitoring and analysis results; if a trend of eutrophication is found, analyzing the input of aquatic organisms into the aquaculture area; and regulating the environment of the aquaculture area based on the analysis results. This invention analyzes water quality trends in aquaculture areas, improves environmental monitoring efficiency, ensures that the environment of aquaculture areas meets actual aquaculture needs, increases aquaculture yield, and simultaneously controls aquaculture management costs.

[0005] However, most existing methods focus on correlation modeling, typically assuming that feeding behavior is an external input and water quality status is a passive response, failing to fully consider the two-way relationship of mutual influence and constraint between feeding decisions and water quality feedback. In actual aquaculture, feeding behavior not only affects subsequent water quality status, but changes in water quality, in turn, influence the feeding arrangements in the next stage, forming a dynamic closed-loop coupling. This coupling relationship has a time lag and cumulative nature, making it difficult for simple supervised learning or static prediction models to accurately characterize the real evolutionary process. Summary of the Invention

[0006] Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent control method and system for aquaculture environment based on an AI model, which solves the problems of unclear bidirectional coupling between feeding behavior and water quality feedback, and difficulty in coordinating the control of feeding decisions and water quality risks.

[0008] Technical solution

[0009] To achieve the above objectives, this invention provides the following technical solution: an intelligent regulation method and system for aquaculture environment based on an AI model, comprising: S1, collecting water quality status data, feeding behavior data, and meteorological data under the aquaculture environment, performing preprocessing operations, and constructing an aquaculture regulation and management database after storage; S2, constructing a time-series causal model of water body status based on multi-source time series data, evaluating the strength and stability of bidirectional coupling, and identifying, adjusting, and updating the causal connection structure based on the coupling evaluation results; S3, predicting and assessing the water quality evolution under different feeding schemes through multivariate time series and causal coupling information, and implementing feeding regulation and slow-release treatment based on the risk assessment results; S4, conducting target analysis of feeding schemes based on feeding decision quantity, water quality risk, and growth benefits, and implementing feeding scheme screening and feeding sequence adjustment based on the feeding analysis results; S5, continuously correcting the causal relationship and feeding strategy through feeding execution feedback.

[0010] Further, the specific measures for collecting water quality status data, feeding behavior data, and meteorological data under aquaculture conditions, performing preprocessing operations, and constructing an aquaculture regulation and management database after storage are as follows: Collecting water quality status data, feeding behavior data, and meteorological data under aquaculture conditions; collecting water quality status data: obtaining dissolved oxygen data through dissolved oxygen sensors, temperature data through temperature sensors, pH data through pH sensors, ammonia nitrogen data through ammonia nitrogen sensors, and nitrite data through nitrite sensors, and binding the sampled values ​​of each sensor with the sampling time to form a water quality status time series; collecting feeding behavior data: obtaining feeding amount through feeding equipment control records, feeding time periods through feeding execution time records, feed type through feed bin discharge identification records, feed unit price from feed batch ledgers, and unit selling price from aquaculture species transaction records, and matching the feeding behavior records with the water quality status time series using timestamps to form a feeding behavior time series; collecting meteorological data: obtaining meteorological data through meteorological acquisition devices. Temperature, air pressure, and rainfall data are collected and time-stamped with the water quality status time series to form a meteorological time series. Outlier removal and missing segment completion are performed on the collected water quality status data, feeding behavior data, and meteorological data. Missing segment completion uses linear interpolation between adjacent time points, and the interpolated segments are marked. Maximum and minimum value normalization methods are used to scale the data to a uniform range to standardize the quantitative values. Starting from the execution of a feeding decision, water quality data, feeding behavior records, meteorological change records, and water body regulation action records are continuously collected for the corresponding time period. After the feeding behavior is completed, water quality status data is continuously collected until the change amplitude of each key water quality indicator in multiple consecutive sampling periods is less than the corresponding water quality management threshold. The covered time interval is defined as a monitoring cycle. The standardized and normalized water quality status data, feeding behavior data, and meteorological data are stored and an aquaculture regulation and management database is constructed. A risk classification mapping table, a feeding execution dataset, and a causal relationship coupling library are built within the aquaculture regulation and management database.

[0011] Furthermore, based on multi-source time series data, a time-series causal model of water body state is constructed. Specific measures to evaluate the strength and stability of bidirectional coupling are as follows: Water quality data, feeding behavior data, and meteorological data form a multivariate time-series feature set, which is input into the causal relationship analysis unit. Combined with aquaculture mechanisms, the causal orientation between variables is clarified, forming an initial causal association set. The aquaculture mechanisms include the changing patterns of water body physicochemical indicators, the influence of feeding behavior on the nutrient load of the water body, and the constraint relationship of water quality state on feeding arrangements. The causal orientation is jointly determined by aquaculture standards and historical operational data statistics, serving as the directional constraint condition for the causal structure learning stage. A dynamic Bayesian network is used to conduct causal structure learning, characterizing the cross-time-step influence path of feeding behavior on the water body's physicochemical state, and the feedback relationship of water body state on subsequent feeding arrangements. Direct influence relationships, indirect transmission relationships, and time lag relationships are identified. A time-series causal model of water body state containing multi-time-step dependent features is constructed, and a causal connection structure graph is output. During the learning process, the causal connection strength and topological morphology are updated based on the Bayesian information content criterion, and a graph structure Laplace regularization is introduced. This approach weakens structural fluctuations caused by short-term disturbances. During the operational phase, newly collected time-series data is continuously input into the causal relationship analysis unit, and causal structure reassessment is conducted according to the monitoring cycle to adjust the causal connection strength and correlation pattern. Finally, connection strength sub-matrices pointing from feeding variables to water quality variables and from water quality variables to feeding variables are extracted from the water body state time-series causal model. Both connection strength sub-matrices are two-dimensional matrices, with their row and column dimensions corresponding to the set of feeding behavior variables and the set of water quality state variables, respectively. Matrix elements represent the causal interaction strength between corresponding variables. Before output, the connection strength is normalized based on historical value ranges to map each connection strength to a unified interval, ensuring the comparability of connection strengths under different variables and at different times. The squares of all elements in the connection strength submatrix from the feeding variable to the water quality variable are summed, and the square root is taken to obtain the Frobenius norm in the direction from feeding to water quality. Similarly, the connection strength submatrix from the water quality variable to the feeding variable is calculated in the same way to obtain the Frobenius norm in the direction from water quality to feeding. The two norms are multiplied and the square root is taken to obtain the bidirectional coupling strength value.

[0012] Furthermore, the specific measures for identifying, adjusting, and updating the causal connection structure based on the coupling assessment results are as follows: By comparing the bidirectional coupling strength value and the coupling strength threshold in real time, when the bidirectional coupling strength value is less than the coupling strength threshold, the existing feeding rhythm and water body regulation method are maintained, and only the causal connection changes between feeding behavior and water quality feedback are recorded, marked as stable operating segments, and archived in the causal relationship coupling library; when the bidirectional coupling strength value is greater than or equal to the coupling strength threshold, the relevant time segments are marked as severely coupled and enter the causal relationship stability analysis process: severely coupled time segments are grouped and compared with historical stable operating segments, and the feeding... For each candidate causal link between the feed variable and the water quality variable, the direction results are statistically analyzed according to the monitoring period to form a direction sequence. When the direction switching rate of the direction sequence is less than the stability determination threshold within n consecutive monitoring periods, the direction of the link is determined to be stable. If the direction is consistent with the original structure, the original link direction is maintained and its link strength is reassessed. If the direction is opposite to the original structure, the original link is removed and a new reverse link is established. When the direction determination result has a direction switching rate greater than or equal to the stability determination threshold within n consecutive monitoring periods, the link is determined to have no stable causal significance and is removed. After the direction processing is completed, a new causal link structure diagram is generated.

[0013] Furthermore, the specific measures for predicting and risk assessing water quality evolution under different feeding schemes by using multivariate time series and causal coupling information are as follows: Based on the multivariate time series feature set, a prediction time window is constructed with the feeding decision point as the benchmark, and feeding behavior is embedded as an exogenous driving force into the water body state evolution process to generate a time series inference sample sequence; a recurrent neural network structure is used to learn the time dependency relationship between water quality state, meteorological conditions and feeding behavior to construct a water quality time series prediction model; multiple sets of different hypothetical feeding schemes are constructed for the same decision point, and input into the water quality time series prediction model in sequence to predict the future... The water quality changes at various time points are extrapolated to form multiple water quality evolution paths. The water quality evolution results corresponding to different feeding schemes are compared and analyzed to output the predicted results and relative deviations of water quality indicators at multiple future time points. The bidirectional causal coupling strength value and the number of water quality indicators at time point t are obtained. The relative deviation is calculated, and the absolute value of the relative deviation is added together and then divided by two to obtain the effective risk deviation of a single water quality indicator. The effective risk deviations corresponding to all water quality indicators are summed and divided by the number of water quality indicators to obtain the comprehensive water quality deviation strength. The comprehensive water quality deviation strength is multiplied by the bidirectional causal coupling strength to obtain the water quality risk value.

[0014] Furthermore, the specific measures for implementing feeding regulation and slow-release treatment based on the risk assessment results are as follows: By comparing the water quality risk value and risk threshold in real time, when the water quality risk value is less than the risk threshold, the feeding plan is allowed to enter the subsequent evaluation process, and water quality change data is collected as a control sample; when the water quality risk value is greater than or equal to the risk threshold, risk mitigation measures are implemented: feeding behavior is constrained, the amount of feeding per unit time is reduced and the feeding interval is extended, and the original single feeding process is broken into multiple time-dispersed feeding sub-processes to weaken the instantaneous impact of nutrient input on the water state; water circulation and oxygenation regulation are activated in conjunction to accelerate the diffusion of metabolic products and dissolved oxygen replenishment, and to prolong the impact of feeding disturbance on the water in the time dimension; if the water quality risk value shows a monotonically decreasing trend and is stably less than the risk threshold within m consecutive sampling periods, it is determined that the risk has fallen back, and the normal feeding rhythm is restored; otherwise, the risk mitigation measures are maintained and the corresponding time segment is recorded as a high-risk operating sample, and the causal relationship stability analysis process is entered into the next monitoring period.

[0015] Furthermore, the specific measures for analyzing the feeding scheme objectives based on feeding decision quantity, water quality risk, and growth benefits are as follows: Obtain the feeding period and feeding amount; uniformly scale and encode the feeding amount and feeding period information corresponding to candidate feeding schemes to obtain candidate feeding decision quantities reflecting different schemes; estimate the target quantity based on aquaculture release records, historical survival rate statistics, and current cycle mortality records, and complete the scale conversion by combining the body length-to-width ratio of the cultured species; correct the weight estimation deviation through sampling weighing results, and simultaneously correct the survival quantity based on mortality records to obtain the estimated biomass values ​​at the start and end of the prediction time window, and calculate the difference to obtain the predicted growth benefit quantity; calculate the predicted growth benefit quantity under the action of candidate feeding schemes, add one to obtain the benefit adjustment term, multiply the feeding decision quantity by the water quality risk value, and then divide by the benefit adjustment term to obtain the target value of the feeding scheme.

[0016] Furthermore, the specific measures for screening feeding schemes and adjusting feeding sequences based on the feeding analysis results are as follows: By comparing the target value and target threshold of the feeding scheme in real time, when the target value of the feeding scheme is less than the target threshold, the corresponding feeding scheme is selected as the execution scheme, and the scheme parameters are simultaneously recorded and archived into the feeding execution dataset as a reference for subsequent feedback; when the target value of the feeding scheme is greater than or equal to the target threshold, the corresponding feeding scheme is removed, and it is determined that the scheme does not have an advantage in terms of benefit and environmental constraints, so it is removed from the set of executable schemes. At the same time, the remaining schemes are subjected to serialization reconstruction processing: the peak time of risk is located according to the prediction curve of water quality risk changing over time, and the midpoint time of adjacent risk peaks is used as the dividing boundary. The original single-cycle feeding period is divided into multiple risk segments. No sub-feeding is scheduled within the risk peak coverage window. The feeding amount is shifted to the feeding period when the water quality risk value is less than the risk threshold. The continuous feeding action is broken down into multiple time-dispersed sub-feeding segments. A single sub-feeding limit is set for each sub-feeding segment. The recommended upper limit coefficient is given by the risk classification mapping table and constrained by the total feeding amount of the original plan. At the same time, based on the oxygenation intensity and water circulation intensity given by the risk classification mapping table, matching oxygenation and circulation combinations are configured before and after each sub-feeding segment. The interval between adjacent sub-feeding segments is verified to be no less than the minimum interval threshold set by the risk classification mapping table. This ensures the continuity of feeding while reducing the environmental disturbance risk caused by a single decision.

[0017] Furthermore, the specific measures for continuously correcting causal relationships and feeding strategies through feeding execution feedback are as follows: After each round of feeding, the actual water quality sequence and the predicted environmental state results are compared point by point over time to extract the water quality deviation trajectory and the trajectory of changes in coupling strength, forming a sample set for causal structure correction and risk assessment. Incremental reconstruction is carried out on the connection strength and direction combination of feeding variables pointing to water quality variables and water quality variables pointing to feeding variables, identifying long-term stable transmission chains and eliminating false associations caused by random disturbances, so that the causal map continues to evolve with the evolution of the aquaculture scenario. By utilizing the deviation between the target value of the feeding plan and the actual growth benefits and water quality fluctuations, the execution records with excellent performance and controlled water quality risks are precipitated as positive templates, and the records with insufficient benefits and rising risks are marked as constraint templates and incorporated into the feeding strategy. This allows the predicted relationship to gradually adapt to changes in the aquaculture environment and feeding behavior, forming a closed-loop update mechanism that links feeding decisions, water quality responses, and risk feedback.

[0018] Furthermore, the second aspect of this invention provides an AI-based intelligent aquaculture environment control system, applied to an AI-based intelligent aquaculture environment control method, comprising: a data acquisition and preprocessing module for collecting water quality status data, feeding behavior data, and meteorological data from the aquaculture environment, performing preprocessing operations, and constructing an aquaculture control and management database after storage; a causal graph and coupling analysis module for constructing a water body state time-series causal model based on multi-source time series data, evaluating the strength and stability of bidirectional coupling, and identifying, adjusting, and updating the causal connection structure based on the coupling evaluation results; a state prediction and feeding simulation module for predicting and assessing the water quality evolution under different feeding schemes through multivariate time series and causal coupling information, and implementing feeding control and slow-release treatment based on the risk assessment results; a feeding decision optimization module for analyzing feeding scheme objectives based on feeding decision amount, water quality risk, and growth benefits, and implementing feeding scheme screening and feeding sequence adjustment based on the feeding analysis results; and a feedback and adaptive update module for continuously correcting causal relationships and feeding strategies through feeding execution feedback.

[0019] Beneficial effects

[0020] The present invention has the following beneficial effects:

[0021] (1) This invention constructs a time-series causal model of water state between water state and feeding behavior by using dynamic Bayesian network causal analysis method. It can accurately identify bidirectional influence paths and their coupling strength, thereby realizing dynamic quantification of the interaction between feeding and water quality. It effectively solves the problems of incomplete causal relationship identification and difficulty in grasping the nature of variable interaction in the prior art.

[0022] (2) By introducing a two-way coupling strength value evaluation mechanism, this invention monitors the degree of causal coupling between feeding behavior and water quality status in real time, realizes adaptive correction of causal structure, and thus realizes closed-loop linkage between control scheme and environmental feedback, effectively solving the problems of unclear feedback path and easy instability of control in the prior art.

[0023] (3) This invention constructs a multi-objective feeding scheme evaluation system, which integrates key indicators such as feeding amount, water quality risk and growth benefit into the decision-making objectives, realizes intelligent screening and sequential optimization of each candidate scheme, and thus achieves the effect of maximizing aquaculture benefits and synergistic control of environmental safety. It effectively solves the problems of single regulation objectives and separation of feeding decisions and ecological constraints in the existing technology.

[0024] (4) This invention uses a recurrent neural network for time-series prediction, combined with multiple hypothetical feeding schemes, to deduce the future water quality evolution path and calculate water quality risks, thereby realizing a forward-looking assessment of environmental risks for multiple feeding strategies, effectively solving the problems of weak linkage between feeding and environmental changes and untimely risk identification in the existing technology.

[0025] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0026] Figure 1 This is a flowchart of the intelligent control method for aquaculture environment based on an AI model, as described in this invention.

[0027] Figure 2 This is a structural diagram of the AI-based intelligent aquaculture environment control system of the present invention;

[0028] Figure 3 This is a two-way coupled causal network diagram of aquaculture feeding and water quality in this invention;

[0029] Figure 4 This is a graph showing the changing trend of the bidirectional coupling strength of aquaculture and water quality risk in this invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1-4 This invention provides a technical solution: an intelligent regulation method and system for aquaculture environment based on an AI model, comprising: S1, collecting water quality status data, feeding behavior data, and meteorological data under the aquaculture environment, performing preprocessing operations, and constructing an aquaculture regulation and management database after storage; S2, constructing a time-series causal model of water body status based on multi-source time series data, evaluating the strength and stability of bidirectional coupling, and judging, adjusting, and updating the causal connection structure according to the coupling evaluation results; S3, predicting and assessing the water quality evolution under different feeding schemes through multivariate time series and causal coupling information, and implementing feeding regulation and slow-release treatment based on the risk judgment results; S4, analyzing the feeding scheme objectives based on feeding decision quantity, water quality risk, and growth benefits, and implementing feeding scheme screening and feeding sequence adjustment based on the feeding analysis results; S5, continuously correcting the causal relationship and feeding strategy through feeding execution feedback.

[0032] Specifically, the measures for collecting water quality data, feeding behavior data, and meteorological data from the aquaculture environment, performing preprocessing operations, and constructing an aquaculture regulation and management database after storage are as follows: Collecting water quality data, feeding behavior data, and meteorological data from the aquaculture environment; collecting water quality data: obtaining dissolved oxygen data through a dissolved oxygen sensor, temperature data through a temperature sensor, pH data through a pH sensor, ammonia nitrogen data through a ammonia nitrogen sensor, and nitrite data through a nitrite sensor, and attaching a uniform time stamp to the sampled values ​​of each sensor, and comparing the sampling results with the corresponding sample values. Time is linked to form a time series of water quality status arranged chronologically; feeding behavior data is collected: feeding amount is obtained through feeding equipment control records, feeding time period is obtained through feeding execution time records, feed type is obtained through feed bin discharge identification records, feed unit price is obtained from feed batch ledgers, and unit selling price is obtained from aquaculture species transaction records. The feeding behavior records are then timestamped with the corresponding water quality status time series to form a feeding behavior time series reflecting the correspondence between feeding behavior and water quality status; meteorological data is collected: temperature data, air pressure data, and rainfall data are obtained through meteorological acquisition devices, and... Meteorological sampling results were time-aligned to align with the water quality time series, forming a meteorological time series. Outlier removal and missing segment completion were performed on the collected water quality data, feeding behavior data, and meteorological data. Missing segment completion was achieved using linear interpolation between adjacent time points, and the interpolated data segments were marked for subsequent analysis. After data integrity processing, maximum and minimum value normalization methods were used to scale all types of data to a uniform range to eliminate the influence of different units and value ranges on subsequent analysis. Starting from the execution of a feeding decision, water quality data was continuously collected for the corresponding time period. The system collects and tracks data on feeding behavior, meteorological changes, and water body regulation actions. After the feeding action is completed, water quality data continues to be collected until the variation of each key water quality indicator within multiple consecutive sampling periods is less than the corresponding water quality management threshold. The covered time interval is defined as a monitoring cycle. The standardized and normalized water quality data, feeding behavior data, and meteorological data are stored uniformly to construct an aquaculture regulation and management database. Within this database, a risk classification mapping table, a feeding execution dataset, and a causal relationship coupling library are constructed to support subsequent regulation analysis and decision-making processes.

[0033] In this implementation plan, water quality status data, feeding behavior data, and meteorological data are collected uniformly, time-aligned, quality-cleaned, and scale-standardized. Data organization and archiving are completed using the monitoring cycle as the basic unit. This achieves consistent expression of multi-source heterogeneous information in the aquaculture environment across time and numerical scales, ensuring a clear and traceable correspondence between feeding behavior, water quality changes, and the external environment. This provides a continuous, complete, and comparable data foundation for subsequent causal relationship analysis, risk assessment, and feeding decisions, effectively avoiding interference caused by data asynchrony, response truncation, and sample mismatch in the control and analysis results.

[0034] Specifically, the specific measures for constructing a time-series causal model of water body status based on multi-source time-series data and evaluating the strength and stability of bidirectional coupling are as follows: Water quality status data, feeding behavior data, and meteorological data are combined to form a multivariate time-series feature set. This set is then aligned with timestamps and segmented according to the monitoring cycle. After outlier removal and missing segment interpolation marking, the data is used as input to the causal relationship analysis unit. Combined with aquaculture mechanisms, the causal relationships between variables are clarified, forming an initial causal association set. The aquaculture mechanisms include the changing patterns of water body physicochemical indicators, the influence of feeding behavior on the nutrient load of the water body, and the constraint relationship between water quality status and feeding arrangements. The causal relationships are determined through aquaculture... The feeding standards and historical operational data statistics are jointly determined and used as directional constraints in the causal structure learning stage to limit the causal search space and improve the reliability of causal relationship identification. A dynamic Bayesian network is used to conduct causal structure learning, characterizing the cross-time-step influence path of feeding behavior on the physicochemical state of water bodies on samples from multiple monitoring periods, as well as the feedback relationship between water body state and subsequent feeding arrangements. Direct influence relationships, indirect transmission relationships, and time-lag relationships are identified. A time-series causal model of water body state containing multi-time-step dependency features is constructed, and a causal connection structure diagram is output, making the transmission chain between feeding behavior variables and water quality state variables visually presented, such as... Figure 3This is the bidirectional causal network diagram of feeding and water quality in aquaculture in this embodiment. During the learning process, the causal connection strength and topological form are updated based on the Bayesian information criterion, and a graph structure Laplace regularization term is introduced to weaken the structural fluctuations caused by short-term disturbances. Fine-grained corrections are made to the edge connections while maintaining the stability of the main structure. During the operation phase, newly collected time-series data are continuously input into the causal relationship analysis unit, and the causal structure is re-evaluated according to the monitoring cycle. The causal connection strength and correlation form are adjusted so that the water state time-series causal model can gradually evolve with changes in the aquaculture environment and feeding habits. Finally, the connection strength sub-matrix of feeding variables pointing to water quality variables and the connection strength sub-matrix of water quality variables pointing to feeding variables are extracted from the water state time-series causal model. The connection strength submatrices are all two-dimensional matrices, with their row and column dimensions corresponding to the set of feeding behavior variables and the set of water quality state variables, respectively. The matrix elements represent the causal interaction strength between the corresponding variables. Before output, the connection strengths are normalized based on historical value ranges to map each connection strength to a unified interval, ensuring the comparability of connection strengths for different variables and at different times. The squares of all elements in the connection strength submatrices from the feeding variable to the water quality variable are summed, and the square root is taken to obtain the Frobenius norm from feeding to water quality. Similarly, the same calculation is performed on the connection strength submatrices from water quality variable to feeding variable to obtain the Frobenius norm from water quality to feeding. The two norms are multiplied and the square root is taken to obtain the bidirectional coupling strength value.

[0035] The specific calculation method for the bidirectional coupling strength value is as follows:

[0036] ;

[0037] In the formula, Indicates time The strength of the two-way causal coupling between feeding behavior and water quality feedback is used to characterize the degree of interaction between feeding decisions and water conditions. Indicates time The connection strength submatrix between the feeding variable and the water quality variable. Indicates time The connection strength submatrix between water quality variables and feeding variables, This represents the Frobenius norm, used to map multidimensional connectivity strength to a single-scale quantity.

[0038] In this embodiment, the feeding amount in time step one is 0, the feeding period is no feeding, the dissolved oxygen is 6.8, the water temperature is 24.2, the ammonia nitrogen is 0.15, the nitrite is 0.02, the pH is 7.8, the air temperature is 23.5, the feeding-water quality norm is 1.28, the water quality-feeding norm is 1.15, and the calculated bidirectional coupling strength is 1.214; the feeding amount in time step two is 120, the feeding period is morning feeding, and the dissolved oxygen... The parameters were: pH 5.3, water temperature 25.8°C, ammonia nitrogen 0.38 g / L, nitrite 0.07 g / L, pH 7.6, air temperature 26.8°C, feeding-water quality norm 1.68, water quality-feeding norm 1.48, and the calculated bidirectional coupling strength was 1.578. In time step three, the feeding amount was 150 g / L, the feeding time was morning, dissolved oxygen was 4.9 g / L, water temperature 27.5°C, ammonia nitrogen 0.65 g / L, and nitrite 0.1 g / L. 4. With a pH of 7.4, an air temperature of 30.2°C, a feeding-to-water quality norm of 2.055, and a water quality-to-feeding norm of 2.164, the calculated bidirectional coupling strength is 2.109. In time step four, the feeding amount is 0, the feeding period is no feeding, dissolved oxygen is 5.8, water temperature is 26.8°C, ammonia nitrogen is 0.55, nitrite is 0.12, pH is 7.6, air temperature is 28.5°C, and the feeding-to-water quality norm is 1.82. The water quality → feeding norm is 1.76, and the calculated bidirectional coupling strength is 1.789. The feeding amount in time step five is 80, the feeding time is evening, the dissolved oxygen is 6.2, the water temperature is 25.5, the ammonia nitrogen is 0.35, the nitrite is 0.08, the pH is 7.7, the air temperature is 26.0, the feeding → water quality norm is 1.45, the water quality → feeding norm is 1.38, and the calculated bidirectional coupling strength is 1.415.

[0039] Table 1. Time-series monitoring data of feeding-water quality coupling in aquaculture

[0040] Time step Feeding amount Feeding time Dissolved oxygen water temperature ammonia nitrogen nitrite pH Temperature Feeding → Water Quality Norm Water quality → Feeding norm bidirectional coupling strength value 1 0 Unfed 6.8 24.2 0.15 0.02 7.8 23.5 1.28 1.15 1.214 2 120 Feeding in the morning 5.3 25.8 0.38 0.07 7.6 26.8 1.68 1.48 1.578 3 150 Feeding in the morning 4.9 27.5 0.65 0.14 7.4 30.2 2.055 2.164 2.109 4 0 Unfed 5.8 26.8 0.55 0.12 7.6 28.5 1.82 1.76 1.789 5 80 Feeding in the evening 6.2 25.5 0.35 0.08 7.7 26.0 1.45 1.38 1.415

[0041] like Figure 4 The figure shown is a trend diagram of the relationship between the bidirectional coupling strength of aquaculture and water quality risk provided in the embodiments of this application. Table 1 and Figure 4Data shows that when feeding behavior changes, an increase in feeding amount directly enhances the bidirectional coupling effect between feeding and water quality. At time step 3, the highest bidirectional coupling strength value of 2.109 was generated due to the combined effect of increasing the feeding amount to 150 kg and the ammonia nitrogen concentration rising to 0.65. When the bidirectional coupling strength value exceeds the coupling threshold, it enters a strong coupling region, at which point water quality deteriorates, with ammonia nitrogen at 0.68. By activating aeration and adjusting the feeding strategy, feeding 80 kg in the evening, the bidirectional coupling strength value dropped to 1.415, and water quality recovered. This demonstrates that the intelligent aquaculture environment control method and system provided in this application, through real-time calculation and monitoring of the bidirectional coupling strength, can effectively provide early warning of water quality deterioration and dynamically adjust the feeding strategy, thereby ensuring the stability and reliability of aquaculture.

[0042] In this implementation scheme, water quality status data, feeding behavior data, and meteorological data are uniformly constructed into a multivariate time series feature set. In the causal relationship analysis unit, causal orientation constraints and dynamic Bayesian network learning are completed in combination with aquaculture mechanisms. This yields a water body state time-series causal model that can simultaneously characterize the influence path of feeding behavior on water quality status and the feedback path of water quality status on subsequent feeding arrangements. Based on this, connection strength sub-matrices from feeding to water quality and from water quality to feeding are extracted. After normalization and Frobenius norm operations, the high-dimensional connection relationships are compressed into bidirectional coupling strength values. This achieves a quantitative expression and stable characterization of the bidirectional causal coupling relationship between feeding behavior and water quality feedback. It provides a comparable coupling degree input for subsequent water quality risk assessment and feeding scheme target analysis, effectively solving the problem in existing technologies where the dependence relationship between feeding behavior and water quality feedback is ambiguous and difficult to reflect the degree of bidirectional coupling with a single indicator.

[0043] Specifically, the measures for identifying, adjusting, and updating the causal connection structure based on the coupling assessment results are as follows: By comparing the bidirectional coupling strength value and the coupling threshold in real time, when the bidirectional coupling strength value is less than the coupling threshold, the existing feeding rhythm and water body regulation method are maintained. Simultaneously, the causal connection changes between feeding behavior and water quality feedback within each monitoring cycle are continuously archived, and these time periods are uniformly marked as stable operating segments and archived in the causal relationship coupling library. When the bidirectional coupling strength value is greater than or equal to the coupling threshold, the relevant time segments are marked as severely coupled, and the process automatically enters the causal relationship stability analysis workflow: severely coupled time segments are grouped and compared with historical stable operating segments. For each candidate causal connection between a feeding variable and a water quality variable, the pointing results are statistically analyzed according to the monitoring cycle, forming a detailed analysis. The analysis process involves a fine directional sequence. If the directional switching rate of the directional sequence is less than the stability threshold within n consecutive monitoring periods, the connection direction is considered stable. If the direction is consistent with the original structure, the original connection direction is maintained, and its connection strength is reassessed based on the latest data. Here, n is the number of consecutive monitoring periods, which is the statistical interval used to determine the stability of the causal connection direction, ranging from 5 to 20. If the direction is opposite to the original structure, the original connection is removed in a timely manner, and a new reverse connection is established to reflect the real change. When the directional switching rate of the direction determination result is greater than or equal to the stability threshold within n consecutive monitoring periods, the connection is deemed to lack stable causal significance and is removed. Finally, after all directional processing is completed, a new causal connection structure diagram is generated, providing basic support for subsequent intelligent control and strategy optimization.

[0044] In this implementation plan, this step, by dynamically monitoring and comparing the bidirectional coupling strength value and the coupling threshold, can accurately identify the changing characteristics of the causal relationship between feeding behavior and water quality feedback, realize the automatic classification and archiving of stable operation segments and severely coupled segments, and further combine directional sequence analysis to effectively judge the stability and directional evolution of each causal connection, ensuring that the causal structure diagram always reflects the real linkage mechanism between the current aquaculture environment and feeding regulation, thereby laying a solid data and structural foundation for the dynamic optimization and long-term adaptive evolution of intelligent regulation strategies.

[0045] Specifically, the measures for predicting and risk assessment of water quality evolution under different feeding schemes by using multivariate time series and causal coupling information are as follows: Based on the multivariate time series feature set, a prediction time window is constructed with the feeding decision point as the benchmark, and feeding behavior is embedded as an exogenous driving force into the water body state evolution process. First, the collected water quality state, meteorological conditions, and feeding behavior variables are time-series aligned and normalized to generate a time-series inference sample sequence with continuous dependency characteristics. A recurrent neural network structure is used to train the above samples to fully learn the temporal correlation and nonlinear dynamic relationship between the variables, forming a water quality time-series prediction model. For the same decision point, multiple sets of different hypothetical feeding schemes are constructed, and they are input into the water quality time-series prediction model in conjunction with environmental state variables to extrapolate the changes in water quality indicators at multiple future times, generating multiple water quality evolution paths. Subsequently, for each set of feeding schemes, a comparative analysis is conducted on the corresponding water quality evolution results, and the predicted values ​​of water quality indicators at multiple future times and their relative deviations from water quality management thresholds are output one by one. The system obtains the bidirectional causal coupling strength value and the number of water quality indicators at each time point. The relative deviation of each water quality indicator is added to its absolute value and then divided by two to calculate the effective risk deviation of a single water quality indicator. Then, the effective risk deviations corresponding to all water quality indicators are summed and divided by two to obtain the comprehensive water quality deviation strength. Finally, the comprehensive water quality deviation strength is multiplied by the bidirectional causal coupling strength value to output the water quality risk value under the current decision scheme, which is used for subsequent risk assessment and strategy optimization.

[0046] The specific calculation method for water quality risk value is as follows:

[0047] ;

[0048] In the formula, This represents the comprehensive water quality risk value at time t, used to quantify the degree to which the future water body state deviates from the safe range under the current coupling strength conditions; This represents the strength of the two-way causal coupling. denoted by , m represents the relative deviation of the m-th water quality indicator at the prediction time, used to characterize the proportion of difference between the predicted value and the safety threshold; M represents the number of water quality indicators involved in the risk assessment.

[0049] In this implementation plan, this step, by predicting and quantifying the future water quality status under different hypothetical feeding schemes over time and combining it with the real-time acquired bidirectional causal coupling strength value, can comprehensively assess the potential risks of each feeding decision to the aquatic environment, achieve a fine characterization of multivariate dynamic coupling and environmental evolution trends, and thus provide a scientific basis for intelligent screening and optimization of feeding schemes, thereby significantly improving the foresight, safety and decision-making accuracy of aquaculture environment regulation.

[0050] Specifically, the specific measures for implementing feeding regulation and slow-release treatment based on the risk assessment results are as follows: By comparing the water quality risk value and risk threshold in real time, when the water quality risk value is less than the risk threshold, the feeding plan is allowed to enter the subsequent evaluation process. Changes in dissolved oxygen, temperature, pH, ammonia nitrogen, and nitrite are continuously monitored before and after feeding, and the corresponding water quality change data, along with the feeding amount, feeding time, and water body regulation records, are included in the control sample as a reference baseline for subsequent water quality time-series prediction model calibration and feeding plan effectiveness evaluation; when the water quality risk value is high... When the risk threshold is met or equal to the risk level, risk mitigation measures are implemented. First, feeding behavior is restricted by reducing the amount of feed given per unit time and extending the feeding interval. The original single feeding process is broken down into multiple staggered feeding sub-processes on the time axis. While ensuring that the total daily feed amount and energy supply remain essentially unchanged, nutrient input is distributed across multiple time points to weaken the instantaneous impact of nutrient input on the water body. Simultaneously, water quality monitoring points are set up before and after each feeding sub-process to track the disturbance response. After the feeding sub-processes are arranged, water circulation is initiated in a coordinated manner. In conjunction with oxygenation regulation, the combination of circulation flow rate and oxygenation duration is selected based on the current water quality risk value to facilitate faster outward diffusion of metabolic products and improve dissolved oxygen replenishment efficiency. This extends the impact of feeding disturbances on the water body over time, slowing down the rate of risk accumulation. If the water quality risk value shows a monotonically decreasing trend and remains consistently below the risk threshold for m consecutive sampling periods, the risk is considered to have decreased. In subsequent feeding periods, the regular feeding rhythm is gradually restored, and additional water body regulation actions are reduced. These periods are marked as successful risk mitigation segments for subsequent strategy evaluation, where m represents the number of consecutive sampling periods used for risk trend determination, ranging from 3 to 5. Conversely, if the water quality risk value remains near the risk threshold or even continues to rise for multiple consecutive sampling periods, the risk mitigation measures are maintained, and the corresponding time segments are recorded as severe risk operation samples. The current feeding plan, water quality evolution trajectory, and water body regulation records are also stored, and the process proceeds to the next monitoring period's causal stability analysis flow to reassess the bidirectional coupling strength and causal connection structure.

[0051] In this implementation plan, by comparing the water quality risk value with the risk threshold in real time, the feeding scheme is allowed to proceed to the subsequent evaluation and be settled as a control sample when the water quality risk value is within the safe range. When the water quality risk value exceeds the risk threshold, risk mitigation measures such as reduced feeding in multiple sessions and linkage between water circulation and aeration are triggered. The risk change trend of multiple consecutive sampling cycles is used to determine whether the risk has decreased or continued to increase, and is marked as a successful mitigation segment and a severe risk operation sample, respectively. This achieves the effect of adjusting the feeding intensity, triggering water environment intervention, and marking and archiving operation segments based on the water quality risk value before and after the feeding is implemented, providing a reliable basis for subsequent causal stability analysis and optimization of control strategies.

[0052] Specifically, the specific measures for analyzing feeding program objectives based on feeding decision quantities, water quality risks, and growth benefits are as follows: Obtain the feeding time period and feeding amount; perform unified-scale coding on the corresponding feeding amount and feeding time period information in the candidate feeding programs, mapping feeding programs at different time locations and with different feeding intensities to comparable candidate feeding decision quantities, reflecting the comprehensive characteristics of each program in terms of feeding intensity and timing; estimate the target number based on aquaculture release records, historical survival rate statistics, and mortality records within the current monitoring period, and combine this with the existing body length-to-width ratio of the aquaculture species to complete the calculation from body size to body weight. The scaling process involves correcting for weight estimation biases using sampling weighing results and dynamically adjusting the number of survivors based on mortality records. This yields estimated biomass values ​​at the start and end of the prediction time window, respectively. The difference between these two values ​​is then used to obtain the predicted growth benefit under the corresponding feeding scheme. Based on this, a benefit adjustment term is formed by adding one to the predicted growth benefit under the candidate feeding scheme. The aforementioned candidate feeding decision value is then multiplied by the water quality risk value and normalized using the benefit adjustment term as a constraint. Finally, the target value of the feeding scheme, which comprehensively measures the benefit potential and environmental risk, is obtained.

[0053] The specific calculation method for the target value of the feeding plan is as follows:

[0054] ;

[0055] In the formula, Indicates time The target value of the feeding scheme is used to sort and filter different feeding schemes; Indicates time The candidate feeding decision quantity includes feeding quantity and feeding time information and is obtained through a unified scale coding. This represents the water quality risk value under the corresponding feeding scheme, calculated using the aforementioned formula; This represents the predicted growth return under the influence of the candidate feeding scheme.

[0056] In this implementation plan, by expressing the feeding time and amount in the feeding scheme at a unified scale, and by combining the target number of aquaculture, changes in biomass and water quality risk factors to quantitatively evaluate different feeding schemes, a comprehensive characterization of the profit potential and environmental constraints of each candidate feeding scheme is achieved. This transforms the feeding decision-making process from a single experience judgment into a comparable and ranked target analysis process, providing a clear and consistent evaluation basis for subsequent feeding scheme screening and feeding sequence optimization.

[0057] Specifically, the specific measures for screening feeding schemes and adjusting feeding sequences based on feeding analysis results are as follows: By comparing the target value and target threshold of the feeding scheme in real time, when the target value of the feeding scheme is less than the target threshold, the corresponding feeding scheme is selected as the execution scheme, and the scheme parameters are simultaneously recorded and archived into the feeding execution dataset as a subsequent feedback comparison benchmark; when the target value of the feeding scheme is greater than or equal to the target threshold, the corresponding feeding scheme is removed, and it is determined that the scheme does not have an advantage in terms of benefit and environmental constraints, so it is removed from the set of executable schemes. At the same time, the remaining schemes are subjected to serialization reconstruction processing: based on the prediction curve of water quality risk changing over time, the peak time of risk is located, and the midpoint time of adjacent risk peaks is used as the dividing boundary to divide the original single-cycle feeding period into multiple continuous risk segments. The number of segments is limited by the number of risk peaks and the preset maximum segmentation constraint to avoid excessive dispersion of feeding behavior; No sub-feeding is scheduled within the risk peak coverage window. The originally concentrated feeding amount is shifted to feeding periods when the water quality risk value is less than the risk threshold, thus breaking down the continuous feeding action into multiple time-dispersed sub-feeding segments. A single sub-feeding limit is set for each sub-feeding segment. The limit is determined by the recommended limit coefficient given by the corresponding risk level in the risk classification mapping table, and the distribution among the sub-feeding segments is completed under the condition of meeting the total feeding amount constraint. At the same time, according to the configuration rules of oxygenation intensity and water circulation intensity corresponding to different risk levels in the risk classification mapping table, the matching oxygenation and circulation combination is configured in conjunction before and after the start and end of each sub-feeding segment. The time interval between adjacent sub-feeding segments is checked to ensure that it is not less than the minimum interval threshold set in the risk classification mapping table. In this way, while ensuring the continuity of feeding, the risk of instantaneous disturbance to the aquatic environment caused by a single feeding decision is effectively reduced.

[0058] In this implementation plan, by serializing and reconstructing high-risk feeding schemes and linking them with water body regulation, the feeding behavior that was originally concentrated and prone to water quality fluctuations is broken down into a segmented feeding process subject to risk constraints. This achieves coordinated control between feeding timing, single feeding intensity and supporting aeration and circulation actions, so that the feeding process can effectively avoid peak risk periods while meeting growth needs, reduce the instantaneous impact of feeding on the aquatic environment, and improve the safety and stability of overall feeding regulation.

[0059] Specifically, the measures for continuously correcting causal relationships and feeding strategies through feeding execution feedback are as follows: After each round of feeding, the actual water quality sequence and the predicted environmental state are compared point by point over time to extract the water quality deviation trajectory and the coupling strength change trajectory, forming a sample set for causal structure correction and risk assessment. Incremental reconstruction is then implemented on the combination of connection strength and direction between feeding variables and water quality variables. This incremental reconstruction involves statistically analyzing the changing trend of connection strength within a sliding window of the monitoring period, iteratively updating the values ​​in the connection strength submatrix with small steps. When the update result continuously deviates from the historical stable range, topological adjustment is triggered, weakening or deleting unstable connections and strengthening and retaining long-term stable transmission chains, allowing the causal graph to continuously evolve with the aquaculture scenario. The deviation between the target value of the feeding scheme and the actual growth benefits and water quality fluctuations is used to adjust the target value of the feeding scheme. Execution records where the water quality risk value is consistently below the historical median and remains below the risk threshold for multiple monitoring periods are precipitated as positive templates. Records where the target value of the feeding scheme is greater than or equal to the target threshold and the water quality risk value frequently equals the risk threshold are marked as constraint templates. The template's credibility weight and trigger priority are calculated based on the frequency of its occurrence in historical samples and its performance. The template library is input into the feeding strategy learning unit. In the new round of feeding scheme search, positive templates are prioritized for retrieval, while scheme combinations that are close to constraint templates are suppressed. Simultaneously, in the causal relationship analysis unit and the training process of the water quality time series prediction model, the connection strength interval and feeding rhythm characteristics obtained from the template statistics are embedded as prior constraints into the loss function. These constraints participate in parameter updates as a penalty term for deviations from the template statistical characteristics, allowing the prediction relationship to gradually adapt to changes in the aquaculture environment and feeding behavior, forming a closed-loop update mechanism that links feeding decisions, water quality responses, and risk feedback.

[0060] In this implementation scheme, a causal structure correction and risk assessment sample set is constructed by utilizing the deviation between the actual water quality trajectory and the prediction results after each round of feeding. Based on this, the connection strength between feeding and water quality and feeding is updated incrementally with small steps and the topology is optimized to retain long-term stable transmission chains and eliminate spurious associations caused by random disturbances. At the same time, positive and constraint templates are precipitated based on the differences between the target value of the feeding scheme and the actual growth benefits and water quality fluctuations, and trigger priorities are set. The statistical characteristics of the templates are injected into the training process of causal modeling and time series prediction, guiding subsequent feeding scheme searches to be closer to strategy combinations that have performed well in the past. This achieves the effect of synergistic adaptive evolution of the causal map, water quality time series prediction model and feeding strategy as the aquaculture scenario evolves, effectively solving the problem of rigid control rules caused by insufficient utilization of feedback information in the existing technology.

[0061] Specifically, this embodiment provides an AI-based intelligent aquaculture environment control system, applied to an AI-based intelligent aquaculture environment control method. It includes: a data acquisition and preprocessing module, responsible for collecting water quality status data, feeding behavior data, and meteorological data from the aquaculture environment. First, it performs outlier removal, missing segment completion, and normalization on all raw data, aligning data from different sources and collection frequencies. After processing, the data is stored in an aquaculture control and management database, achieving full-process data archiving and management. A causal mapping and coupling analysis module, based on the processed multi-source time-series data, constructs a time-series causal model of water body status. This model dynamically assesses the bidirectional coupling strength and stability between feeding behavior and water quality feedback. Based on the assessment results, it performs real-time identification, adjustment, and updating of the causal connection structure, ensuring that the causal structure always reflects the actual linkage relationships at the aquaculture site. The state prediction and feeding simulation module combines multivariate time series and causal coupling information, using a water quality time series prediction model to extrapolate and assess the future water quality status under different feeding schemes. Based on this, it automatically triggers corresponding feeding controls and risk mitigation measures, enhancing environmental safety capabilities. The feeding decision optimization module comprehensively analyzes multiple indicators such as current feeding decision amount, water quality risk, and growth benefits, intelligently selecting the optimal feeding scheme and adaptively adjusting the feeding sequence to maximize aquaculture benefits and balance ecological safety. The feedback and adaptive update module collects real-time feedback data after each feeding cycle, continuously comparing it with the water quality time series prediction model and causal structure. This drives dynamic correction and adaptive optimization of causal relationships and feeding strategies, forming a closed-loop update mechanism to ensure the system maintains efficient and accurate control capabilities at all times.

[0062] In this implementation plan, this step establishes a complete system covering data acquisition and preprocessing, causal graph analysis, state prediction and feeding simulation, feeding decision optimization, and feedback and adaptive updates. This enables multi-source data fusion, dynamic analysis of causal relationships, intelligent prediction of feeding and water quality evolution, selection of optimal control schemes, and continuous adaptive feedback correction throughout the entire aquaculture process. This significantly improves the scientific, intelligent, and safe nature of aquaculture environment management, providing a solid system foundation for efficient and stable aquaculture control.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0064] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for intelligent control of aquaculture environment based on AI model, characterized in that, Includes the following steps: S1: Collect water quality data, feeding behavior data and meteorological data in the aquaculture environment, perform preprocessing operations, and construct an aquaculture regulation and management database after storage. S2, a time-series causal model of water state is constructed based on multi-source time series data, the strength and stability of bidirectional coupling are evaluated, and the causal connection structure is judged, adjusted and updated according to the coupling evaluation results; The specific measures for constructing a time-series causal model of water state based on multi-source time series data and evaluating the strength and stability of bidirectional coupling are as follows: Water quality data, feeding behavior data, and meteorological data are combined to form a multivariate time series feature set, which is then input into the causal relationship analysis unit. Combined with the aquaculture mechanism, the causal relationships between the variables are clarified, forming an initial causal association set. The aquaculture mechanism includes the changing patterns of water physicochemical indicators, the influence of feeding behavior on the nutrient load of the water body, and the constraint relationship between water quality and feeding arrangements. The causal orientation is determined jointly by aquaculture standards and historical operational data statistics, serving as directional constraints in the causal structure learning stage. A dynamic Bayesian network is used to conduct causal structure learning, characterizing the cross-time-step influence path of feeding behavior on the water physicochemical state, and the feedback relationship of water state on subsequent feeding arrangements. Direct influence relationships, indirect transmission relationships, and time lag relationships are identified, constructing a time-series causal model of water state containing multi-time-step dependency features and outputting a causal connection structure diagram. During the learning process, the causal connection strength and topological form are updated based on the Bayesian information content criterion, and a graph structure Laplace regularization term is introduced to weaken the structural fluctuations caused by short-term disturbances. During the operation phase, newly collected time series data are continuously input into the causal relationship analysis unit, and causal structure reassessment is carried out according to the monitoring cycle to adjust the causal connection strength and correlation form. Finally, the connection strength sub-matrix of feeding variable pointing to water quality variable and the connection strength sub-matrix of water quality variable pointing to feeding variable are extracted from the water body state time series causal model. The connection strength sub-matrix is ​​a two-dimensional matrix, and its row and column dimensions correspond to the set of feeding behavior variables and the set of water quality state variables, respectively. The matrix elements represent the causal effect strength between the corresponding variables. The connection strength is normalized based on the historical value range before output, and each connection strength is mapped to a unified interval to ensure the comparability of connection strengths under different variables and different times. The Frobenius norm of the connection strength submatrix from the feeding variable to the water quality variable is obtained by squaring all elements and summing them, and then taking the square root of the summation. Similarly, the Frobenius norm of the connection strength submatrix from the water quality variable to the feeding variable is obtained by performing the same calculation. The two norms are multiplied and then the square root is taken to obtain the bidirectional coupling strength value. S3 uses multivariate time series and causal coupling information to predict and assess the water quality evolution under different feeding schemes, and implements feeding regulation and slow-release treatment based on the risk assessment results. S4. Based on the feeding decision amount, water quality risk and growth benefit, conduct feeding program target analysis, and implement feeding program screening and feeding sequence adjustment according to the feeding analysis results; S5 continuously corrects the causal relationship and feeding strategy through feeding execution feedback.

2. The intelligent control method for aquaculture environment based on an AI model according to claim 1, characterized in that: The specific measures for collecting water quality data, feeding behavior data, and meteorological data under aquaculture conditions, performing preprocessing operations, and constructing an aquaculture regulation and management database after storage are as follows: Collect water quality data, feeding behavior data, and meteorological data in aquaculture environments; Collect water quality status data: Obtain dissolved oxygen data through a dissolved oxygen sensor, temperature data through a temperature sensor, pH data through a pH sensor, ammonia nitrogen data through ammonia nitrogen sensor, and nitrite data through a nitrite sensor. Then, bind the sampled values ​​of each sensor with the sampling time to form a water quality status time series. Data collection on feeding behavior: The amount of feed is obtained through the control record of the feeding equipment, the feeding period is obtained through the feeding execution time record, the feed type is obtained through the feed bin discharge identification record, the feed unit price is obtained from the feed batch ledger, and the unit selling price is obtained from the aquaculture species transaction record. The feeding behavior record is then matched with the water quality status time series to form a feeding behavior time series. Meteorological data collection: Temperature data, air pressure data, and rainfall data are acquired through meteorological data acquisition devices and then timestamped with the water quality status time series to form a meteorological time series. Outlier removal and missing segment completion were performed on the collected water quality status data, feeding behavior data, and meteorological data. Missing segment completion was achieved using linear interpolation between adjacent time points, and the interpolated segments were marked. The maximum and minimum value normalization method was used to scale the data to a uniform range to unify the quantitative values. Starting from the execution of a feeding decision, water quality data, feeding behavior records, meteorological change records, and water body regulation action records were continuously completed within the corresponding time period. After the feeding behavior was completed, water quality status data was continuously collected until the change range of each key water quality indicator in multiple consecutive sampling periods was less than the corresponding water quality management threshold. The time interval covered was defined as a monitoring period. The water quality status data, feeding behavior data, and meteorological data after standardization and normalization were stored and an aquaculture regulation and management database was constructed. A risk classification mapping table, a feeding execution dataset, and a causal relationship coupling library were constructed in the aquaculture regulation and management database.

3. The intelligent control method for aquaculture environment based on an AI model according to claim 1, characterized in that: The specific measures for identifying, adjusting, and updating the causal connection structure based on the coupling evaluation results are as follows: By comparing the bidirectional coupling strength value and the coupling strength threshold in real time, when the bidirectional coupling strength value is less than the coupling strength threshold, the existing feeding rhythm and water body regulation mode are maintained. Only the causal connection changes between feeding behavior and water quality feedback are recorded, marked as stable operation segments, and archived into the causal relationship coupling library. When the bidirectional coupling strength value is greater than or equal to the coupling strength threshold, the relevant time segments are marked as severely coupled, and the causal relationship stability analysis process is initiated: severely coupled time segments are grouped and compared with historical stable operation segments. For each candidate causal connection between the feeding variable and the water quality variable, the pointing results are statistically analyzed according to the monitoring cycle to form a direction sequence. When the direction switching rate of the direction sequence is less than the stability judgment threshold within n consecutive monitoring cycles, the connection direction is determined to be stable. If the direction is consistent with the original structure, the original connection direction is maintained and its connection strength is re-evaluated. If the direction is opposite to the original structure, the original connection is removed and a new reverse connection is established. When the direction judgment result has a direction switching rate greater than or equal to the stability judgment threshold within n consecutive monitoring cycles, the connection is determined to have no stable causal significance and is removed. After completing the direction processing, a new causal connection structure diagram is generated.

4. The intelligent control method for aquaculture environment based on an AI model according to claim 1, characterized in that: The specific measures for predicting and assessing the water quality evolution under different feeding schemes by using multivariate time series and causal coupling information are as follows: Based on a multivariate time series feature set, a prediction time window is constructed with the feeding decision time point as the benchmark, and the feeding behavior is embedded as an exogenous driving force into the water body state evolution process to generate a time series inference sample sequence. A recurrent neural network structure is used to learn the time dependency relationship between water quality status, meteorological conditions and feeding behavior, and a water quality time series prediction model is constructed. Multiple sets of different hypothetical feeding schemes are constructed for the same decision point and input into the water quality time series prediction model in sequence to extrapolate water quality changes at multiple future times and form multiple water quality evolution paths. The water quality evolution results corresponding to different feeding schemes are compared and analyzed, and the predicted results of water quality indicators and relative deviations at multiple future times are output. Obtain the bidirectional causal coupling strength value and the number of water quality indicators at time point t; The effective risk deviation of a single water quality indicator is obtained by adding the relative deviation amount to the absolute value of the relative deviation amount and then dividing by two. The effective risk deviation amounts corresponding to all water quality indicators are summed and divided by the number of water quality indicators to obtain the comprehensive water quality deviation intensity. The water quality risk value is obtained by multiplying the overall water quality deviation intensity by the two-way causal coupling intensity.

5. The intelligent control method for aquaculture environment based on an AI model according to claim 1, characterized in that: The specific measures for implementing feeding regulation and slow-release treatment based on the risk assessment results are as follows: By comparing water quality risk values ​​and risk thresholds in real time, when the water quality risk value is less than the risk threshold, the feeding plan is allowed to enter the subsequent evaluation process, and water quality change data is collected as a control sample. When the water quality risk value is greater than or equal to the risk threshold, risk mitigation measures are implemented: feeding behavior is restricted, the amount of feed per unit time is reduced and the feeding interval is extended, and the original single feeding process is broken into multiple time-dispersed feeding sub-processes to weaken the instantaneous impact of nutrient input on the water state; water circulation and oxygenation regulation are activated in conjunction to accelerate the diffusion of metabolic products and the replenishment of dissolved oxygen, thus extending the impact of feeding disturbances on the water body in the time dimension; if the water quality risk value shows a monotonically decreasing trend and is stably less than the risk threshold within m consecutive sampling periods, it is determined that the risk has fallen back, and the normal feeding rhythm is restored; Conversely, if risk mitigation measures are maintained and the corresponding time segments are recorded as high-risk operational samples, the causal stability analysis process will proceed to the next monitoring cycle.

6. The intelligent control method for aquaculture environment based on an AI model according to claim 1, characterized in that: The specific measures for analyzing the feeding scheme objectives based on feeding decision-making quantity, water quality risk, and growth benefits are as follows: Obtain feeding time periods and feeding amounts; perform unified scale coding on the feeding amounts and feeding time periods corresponding to candidate feeding schemes to obtain candidate feeding decision quantities reflecting different schemes; estimate the target quantity based on breeding and feeding records, historical survival rate statistics and current cycle mortality records, and complete scale conversion by combining the body length and body width ratio of the breeding species; correct the weight estimation deviation through sampling and weighing results, and correct the survival quantity based on mortality records to obtain the estimated biomass values ​​at the beginning and end of the prediction time window, and calculate the difference to obtain the predicted growth benefit quantity; The growth benefit prediction under the candidate feeding scheme is calculated by adding one to obtain the benefit adjustment term. The feeding decision amount is multiplied by the water quality risk value and then divided by the benefit adjustment term to obtain the target value of the feeding scheme.

7. The intelligent control method for aquaculture environment based on an AI model according to claim 1, characterized in that: The specific measures for screening feeding schemes and adjusting feeding sequences based on feeding analysis results are as follows: By comparing the target value and the target threshold of the feeding plan in real time, when the target value of the feeding plan is less than the target threshold, the corresponding feeding plan is selected as the execution plan, and the plan parameters are recorded and archived to the feeding execution dataset as a reference for subsequent feedback. When the target value of a feeding scheme is greater than or equal to the target threshold, the corresponding feeding scheme is eliminated. If the scheme is deemed not advantageous in terms of benefit and environmental constraints, it is removed from the set of executable schemes. Simultaneously, the remaining schemes undergo serialization and reconstruction: the peak risk time is located based on the predicted curve of water quality risk over time, and the original single-cycle feeding period is divided into multiple risk segments using the midpoint time of adjacent risk peaks as the dividing boundary. No sub-feeding is arranged within the risk peak coverage window, and the feeding amount is migrated to feeding periods where the water quality risk value is less than the risk threshold. Continuous feeding actions are broken down into multiple temporally dispersed sub-feeding segments. A single sub-feeding upper limit is set for each sub-feeding segment, with a recommended upper limit coefficient provided by the risk grading mapping table and constrained allocation with the original scheme's total feeding amount. Simultaneously, based on the oxygenation intensity and water circulation intensity given by the risk grading mapping table, matching oxygenation and circulation combinations are configured before and after each sub-feeding segment, and the interval between adjacent sub-feeding segments is verified to be no less than the minimum interval threshold set by the risk grading mapping table, ensuring feeding continuity while reducing the environmental disturbance risk caused by a single decision.

8. The intelligent control method for aquaculture environment based on an AI model according to claim 1, characterized in that: The specific measures for continuously correcting the causal relationship and feeding strategy through feeding execution feedback are as follows: After each feeding cycle, the actual water quality sequence and the predicted environmental status are compared point-by-point over time. The deviation trajectory of water quality and the trajectory of changes in coupling strength are extracted to form a sample set for causal structure correction and risk assessment. Incremental reconstruction is performed on the combination of connection strength and direction between feeding variables and water quality variables, and between feeding variables and water quality variables. Long-term stable transmission chains are identified and spurious associations caused by random disturbances are eliminated, so that the causal map continues to evolve with the aquaculture scenario. By using the deviation between the target value of the feeding plan and the actual growth benefits and water quality fluctuations, the execution records with excellent performance and controlled water quality risks are precipitated as positive templates, while the records with insufficient benefits and rising risks are marked as constraint templates and incorporated into the feeding strategy. The predicted relationship is gradually adapted to changes in the aquaculture environment and feeding behavior, forming a closed-loop update mechanism that links feeding decisions, water quality responses, and risk feedback.

9. An AI-based intelligent aquaculture environment control system, employing the AI-based intelligent aquaculture environment control method according to any one of claims 1-8, comprising: The data acquisition and preprocessing module is used to collect water quality data, feeding behavior data and meteorological data in the aquaculture environment, perform preprocessing operations, and build an aquaculture regulation and management database after storage. The causal mapping and coupling analysis module is used to construct a time-series causal model of water state based on multi-source time series data, evaluate the strength and stability of bidirectional coupling, and identify, adjust and update the causal connection structure based on the coupling evaluation results. The state prediction and feeding simulation module is used to predict and assess the water quality evolution under different feeding schemes through multivariate time series and causal coupling information, and to implement feeding regulation and slow release treatment based on the risk assessment results. The feeding decision optimization module is used to analyze the feeding scheme objectives based on the feeding decision amount, water quality risk and growth benefits, and to implement feeding scheme screening and feeding sequence adjustment based on the feeding analysis results; The feedback and adaptive update module is used to continuously correct causal relationships and feeding strategies through feedback during feeding.

Images