AI-based trachinotus ovatus parent fish culture water body temperature determination method and system

Abstract

The invention discloses an AI-based trachinotus ovatus parent fish culture water body temperature determination method and system, and the method comprises the steps: collecting culture environment data and parent fish physiological data in real time, and generating a multi-modal time sequence database according to the collected data; based on the multi-modal time sequence database, establishing a target variable data set representing the reproductive performance of the parent fish; an AI prediction model is constructed and trained, a multi-modal time sequence database is used as an input feature, and a target variable data set is used as a prediction target; based on the trained AI prediction model, an optimization target is set, future reproductive performance under different water temperature strategies is simulated within a water temperature preset range, and a water temperature value capable of enabling the optimization target to be optimal is selected as a recommended set temperature according to a simulation result; and sending the recommended set temperature to a temperature control system of the culture pond to execute water temperature regulation and control. According to the method, the temperature of the trachinotus ovatus parent fish culture water body is accurately and intelligently regulated and controlled by fusing the multi-modal time sequence data and the AI prediction model.

Description

Technical Field

[0001] This invention relates to the field of water temperature monitoring for farmed fish, specifically to an AI-based method and system for determining the water temperature in the culture of oval pomfret broodstock. Background Technology

[0002] Currently, in the artificial breeding of broodstock of oval pomfret, water temperature is a key environmental factor affecting gonadal development, spawning behavior, and egg quality. Traditional breeding methods mainly rely on fixed temperature ranges or the experience of breeders for water temperature control, lacking a dynamic response to the real-time physiological state of the broodstock and the interaction of complex environmental factors. This extensive management approach struggles to accurately match the actual needs of the broodstock during their reproductive stages, often leading to unstable reproductive cycles, fluctuating spawning rates, and inconsistent fry quality. Although existing technologies can achieve automatic water temperature control, their setpoints are often based on static knowledge or historical averages, failing to effectively integrate multi-source real-time monitoring data, especially failing to incorporate changes in the physiological behavior of the broodstock into the decision-making loop, resulting in lagging regulation and insufficient adaptability. Furthermore, the relationship between reproductive performance and water temperature is complex and nonlinear, influenced by multiple factors such as dissolved oxygen and light, making it difficult for traditional methods or simple models to accurately capture its inherent patterns and make forward-looking optimizations. Therefore, the industry urgently needs a method that can integrate environmental and physiological time-series information, and achieve precise water temperature setting and dynamic optimization through intelligent learning and simulation, in order to improve the success rate and stability of parent fish breeding. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an AI-based method for determining the water temperature in the culture of broodstock of oval pomfret, comprising the following steps: Real-time data collection of aquaculture environment and parent fish physiological data, and generation of multimodal time-series database based on the collected data; Based on the aforementioned multimodal time-series database, a dataset of target variables characterizing the reproductive performance of parent fish is established. Construct and train an AI prediction model, using the multimodal time series database as input features and the target variable dataset as the prediction target; Based on the trained AI prediction model, the optimization target is set, and the future reproductive performance under different water temperature strategies is simulated within the preset water temperature range. Based on the simulation results, the water temperature value that makes the optimization target the best is selected as the recommended set temperature. The recommended set temperature is sent to the temperature control system of the aquaculture pond to perform water temperature regulation.

[0004] Preferably, the method for generating the multimodal time series database includes: Outlier detection and missing value processing were performed on the collected aquaculture environment data and parent fish physiological data to clean the data; Based on the cleaned data, spatiotemporal alignment and feature-level fusion are performed to generate a multimodal time series database.

[0005] Preferably, the method for establishing the target variable dataset includes: Dynamic change patterns are extracted from the multimodal time series database using the sliding window technique; Based on the extracted dynamic change patterns, time series decomposition is performed to obtain long-term trends, seasonal fluctuations, and residual components as features.

[0006] Preferably, the steps of building and training the AI ​​prediction model include: The multimodal time series database is used as input features and input into a time series prediction model based on an encoder-decoder architecture; The input feature sequence is processed by the bidirectional long short-term memory network at the encoder end to generate the hidden state sequence; The attention weights of the decoder on the hidden state sequence are dynamically calculated based on the attention mechanism, and a context vector is constructed based on the attention weights. Based on the context vector and the hidden state of the decoder, the prediction result is output through a fully connected layer.

[0007] Preferably, the steps for making the water temperature optimization decision include: The current aquaculture status is assessed based on the output of the AI ​​prediction model and the real-time multimodal time series database. Based on knowledge graphs, several candidate water temperature strategies are generated through policy reasoning. The simulator is used to simulate the impact of each candidate water temperature strategy on future reproductive performance, and the optimal water temperature value is selected based on the optimization objective.

[0008] Preferably, the step of sending the recommended set temperature to the temperature control system to perform water temperature regulation includes: The recommended temperature setting is parsed into device control commands and sent to the temperature control system. Based on equipment control instructions, a visual task list is generated for manual operation guidance; The system tracks command execution status in real time using a sensor network and optimizes the system based on feedback data.

[0009] Preferably, it also includes: after the transmission is completed, collecting actual effect data after the decision is executed; and triggering model retraining or rule revision based on the comparison results between the actual effect data and the expected effect.

[0010] The present invention also provides an AI-based system for determining the water temperature in the culture of oval pomfret broodstock. The system is used to implement the above method and includes: a data acquisition module, a characterization module, a construction module, an optimization module, and a transmission module.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves precise and intelligent temperature control in the culture water of oval pomfret broodstock by integrating multimodal time-series data with an AI prediction model. This method dynamically captures the complex nonlinear relationship between environmental parameters and the physiological state of the broodstock, proactively predicts reproductive performance, and simulates the potential effects of different water temperature strategies, thereby generating the optimal set temperature. This not only significantly improves the stability of spawning success rate and fry quality but also effectively reduces the human uncertainty inherent in traditional experience-based systems. The system possesses self-learning and adaptive capabilities, continuously optimizing decisions based on feedback, ultimately achieving the goals of improving aquaculture efficiency and saving energy and labor costs. Attached Figure Description

[0012] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0014] 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.

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] Example 1 like Figure 1 As shown, an AI-based method for determining the water temperature in the culture of oval pomfret broodstock includes: S1. Real-time collection of aquaculture environment data and parent fish physiological data to generate a multimodal time-series database; A sensor network deployed in aquaculture farms collects real-time data on the aquaculture environment, including parameters such as water temperature, dissolved oxygen, pH, and light intensity. Simultaneously, biomonitoring equipment records physiological data of the parent fish, such as weight, heart rate, and spawning behavior. This data is collected and stored synchronously in time-series format. Through data cleaning and fusion processing, environmental and physiological information is integrated according to a unified timestamp, ultimately constructing a multimodal time-series database. Each entry is associated with a specific point in time, ensuring data consistency and traceability, and providing a foundation for subsequent analysis and decision-making.

[0017] During the data cleaning phase, outlier detection and processing were first performed on the collected raw aquaculture environment data and parent fish physiological data. Statistical methods were used to identify and remove noise points that clearly exceeded the reasonable range of physiology or environment. At the same time, missing values ​​caused by sensor failure or transmission interruption were repaired using time series interpolation or correlation-based imputation methods to ensure the integrity and reliability of the data.

[0018] The data then undergoes fusion processing. This process uses high-precision timestamps as a benchmark to spatiotemporally align environmental sensor data and physiological monitoring data from different sampling frequencies. By establishing a unified timeline, the second-level collected water temperature data and minute-level recorded parent fish heart rate and other heterogeneous data are resampled and interpolated for synchronization. Subsequently, a feature-level fusion strategy is employed to extract biologically significant joint features from the aligned multimodal data. For example, a sliding window correlation analysis is performed between dissolved oxygen change curves and parent fish swimming behavior sequences to construct a fusion feature vector reflecting the environment-physiology interaction. Finally, a structured multimodal time-series data set is generated, in which each time node fully includes co-validated environmental parameters and physiological indicators, laying the foundation for establishing an accurate aquaculture status assessment model.

[0019] S2. Establish a target variable dataset, where the target variable is an indicator that characterizes the reproductive performance of the parent fish.

[0020] Using the aforementioned multimodal time-series database, dynamic change patterns are captured from the fused environmental-physiological data through the sliding window technique, and time series decomposition methods are introduced to extract the long-term trends, seasonal fluctuations, and residual components of different parameters.

[0021] In deep feature extraction, a sliding window technique is first applied, with a fixed-length window sliding along the time axis to extract continuous segments from the fused environmental-physiological data, thereby capturing dynamic change patterns such as the coupling relationship between sudden drops in dissolved oxygen and accelerated heart rate in parent fish. Based on this, a time series decomposition method is introduced to decompose each key parameter sequence (such as water temperature and activity level) within the window. Statistical or frequency domain-based algorithms are used to analyze these sequences into long-term trend terms representing slow evolution, seasonal fluctuation terms reflecting intraday or tidal periodic patterns, and random residual components containing information about sudden events. This deconstructs the original complex signal into components at different time scales, thus revealing more clearly the potential causal relationships and driving patterns between environmental factors and physiological responses, providing highly explanatory multi-scale feature inputs for subsequent state assessment models.

[0022] During the time series decomposition process, the system employs a hybrid algorithm based on statistics and frequency domain to perform multi-level analysis of multimodal parameters. For parameters with obvious periodicity, such as water temperature, a classic seasonal decomposition model is used, the additive form of which can be expressed as: in T t This represents the trend item; S t Seasonal component; R t The residuals after removing trend and seasonal factors.

[0023] For non-steady-state physiological signals such as heart rate variability, frequency domain analysis methods are introduced, using Fast Fourier Transform to transform the time-domain information. x ( n Convert to frequency domain representation: Where n represents the time index; N represents the total number of sampling points; and k represents the frequency index.

[0024] Based on the above steps, characteristic frequency band energy distributions corresponding to respiratory metabolism and stress response are identified in the spectrum. This deconstruction method, which combines statistics and the frequency domain, can capture the gradual changes in environmental parameters and the transient rhythmic characteristics of physiological signals, providing a multidimensional feature base that combines macroscopic trends and microscopic fluctuations for subsequent modeling.

[0025] S3. Build and train an AI prediction model, using a multimodal time series database as input features and a target variable set as the prediction target.

[0026] The multi-scale deep feature sequence generated in step S2 is input into a temporal prediction model based on an encoder-decoder architecture. The core of this model is a bidirectional long short-term memory network that incorporates an attention mechanism.

[0027] The specific processing steps are as follows: First, the bidirectional LSTM layer at the encoder reads the input feature sequence in parallel from both forward and backward directions. It processes the information cyclically through its internal gating mechanism to generate a hidden state sequence containing the context information of the entire sequence. Next, the attention layer dynamically calculates the attention weight distribution of the encoder's hidden state at each time step and constructs a context vector accordingly, enabling the model to focus on the historical period most relevant to the prediction target. Finally, the decoder uses this context vector and its own hidden state to output the prediction results and corresponding probabilities of key future aquaculture events (such as the start of spawning) through a fully connected layer. The principle is to enhance the model's ability to capture precursors of important physiological state changes through the attention mechanism, thereby achieving high-precision and interpretable temporal prediction of the parent fish's breeding cycle.

[0028] S4. Optimize water temperature based on the trained AI model: Set optimization goals, simulate future reproductive performance under different water temperature strategies within the preset water temperature range, and select the water temperature value that best achieves the optimization goals as the current recommended optimal setting temperature.

[0029] This step transforms the output of the deep prediction model into aquaculture management decisions that can directly guide production. Its core is the construction of a closed-loop decision support framework that integrates state assessment, strategy deduction, and feedback optimization. This process begins with a multi-dimensional analysis of the prediction results. It not only receives the predicted timing and probabilities of key physiological events in broodstock (such as spawning) from S3, but also simultaneously integrates the latest environmental indicators and broodstock behavioral performance from a real-time multimodal database, forming a comprehensive state matrix that covers future predictions, current conditions, and historical comparisons.

[0030] First, the matrix is ​​deconstructed and evaluated. For example, if the model predicts an spawning probability of over 85% within the next 72 hours, but real-time dissolved oxygen data shows a downward trend and is approaching the critical threshold of 5 mg / L, and computer vision analysis reveals that the activity level of the parent fish colony is significantly lower than expected, then the state will be judged as "high reproductive potential but with environmental inhibition risk," rather than simply adopting the high-probability prediction. Next, the decision engine is triggered. It reasons based on a knowledge graph containing hundreds of aquaculture procedures and coping strategies. This knowledge graph is organized in a semantic network structure, where nodes represent different aquaculture states or operational entities (such as "aerator turned on" or "feeding adjusted"), and edges represent the logical relationships and transition conditions between them. The engine uses a graph traversal algorithm to find all possible operational paths matching the current evaluation state and uses a case-based reasoning method to retrieve intervention measures that have proven effective in similar scenarios from the historical decision database.

[0031] The decision generation process is a multi-objective optimization problem, requiring the simultaneous balancing of several sometimes conflicting objectives, such as breeding success rate, parent fish welfare, energy costs, and management feasibility. The principle is to construct an objective function that aims to maximize expected benefits while minimizing potential risks and resource consumption. In the example above, multiple candidate strategies might be generated simultaneously: Strategy A is "immediately start the aerator and increase its power to 80%, while adding high-protein feed after 6 hours"; Strategy B is "phased aeration, first running at 50% power for 2 hours and monitoring the dissolved oxygen recovery curve; if it does not reach the expected level, activate the backup pure oxygen supply system." The system uses a simulator to simulate each strategy, predicting its cascading effects on environmental parameters and the physiological state of the parent fish over a future period, and scores each strategy based on the objective function. The final output is not a single instruction, but a priority-ordered decision sequence along with detailed execution plans, expected effects, and risk contingency plans.

[0032] To ensure the accuracy and self-evolution of decision-making, this step also incorporates a tight feedback optimization loop. All issued decision commands, their actual execution (recorded via sensors and actuator logs), and subsequent effects (such as final spawning time, spawn volume, and parent fish health) are collected and stored as new cases with complete contextual tags in the historical decision repository. Simultaneously, the expected results of decisions are periodically compared with the actual outcomes. If a significant deviation is found (e.g., an expected 20% increase in spawn volume but only a 5% increase in actual spawn volume), a root cause analysis process is initiated, tracing back to data quality in S1, feature effectiveness in S2, or model accuracy in S3. The analysis conclusions are then used to trigger model retraining or rule revisions in the knowledge graph.

[0033] This mechanism enables the entire method to continuously learn from practical experience and adapt to the dynamic changes in biological and abiotic factors in the aquaculture environment, thereby gradually evolving from a rule-based automated system into an intelligent aquaculture decision-making brain with cognitive and adaptive capabilities.

[0034] S5. Send the determined optimal set temperature to the temperature control system of the aquaculture pond for adjustment.

[0035] The serialized decision schemes generated by S4 are transformed into specific equipment control instructions and manual operation guidance. For example, the decision to "increase the power of the aerator to 80%" is parsed into a specific frequency signal of the frequency converter and sent to the field control.

[0036] Simultaneously, a visual task list is generated for aquaculture managers, including operational items, key indicator warning thresholds, and review time points, and pushed to them via mobile terminals. The execution status and effects of all instructions are tracked in real time through a sensor network, forming a closed loop of "decision-execution-feedback". The core of this stage is to seamlessly embed optimization strategies into the actual production process, and drive the entire intelligent aquaculture system to dynamically adjust and iteratively optimize by continuously comparing expected and actual environmental and physiological response data, ultimately achieving precise implementation of management decisions and a steady improvement in aquaculture efficiency.

[0037] Example 2 This embodiment also provides an AI-based system for determining the water temperature in the culture of oval pomfret broodstock, including: a data acquisition module, a characterization module, a construction module, an optimization module, and a transmission module.

[0038] The following will describe in detail, with reference to this embodiment, how the present invention solves the technical problems in practical work.

[0039] First, the data acquisition module collects aquaculture environment data and parent fish physiological data in real time to generate a multimodal time-series database; A sensor network deployed in aquaculture farms collects real-time data on the aquaculture environment, including parameters such as water temperature, dissolved oxygen, pH, and light intensity. Simultaneously, biomonitoring equipment records physiological data of the parent fish, such as weight, heart rate, and spawning behavior. This data is collected and stored synchronously in time-series format. Through data cleaning and fusion processing, environmental and physiological information is integrated according to a unified timestamp, ultimately constructing a multimodal time-series database. Each entry is associated with a specific point in time, ensuring data consistency and traceability, and providing a foundation for subsequent analysis and decision-making.

[0040] During the data cleaning phase, outlier detection and processing were first performed on the collected raw aquaculture environment data and parent fish physiological data. Statistical methods were used to identify and remove noise points that clearly exceeded the reasonable range of physiology or environment. At the same time, missing values ​​caused by sensor failure or transmission interruption were repaired using time series interpolation or correlation-based imputation methods to ensure the integrity and reliability of the data.

[0041] The data then undergoes fusion processing. This process uses high-precision timestamps as a benchmark to spatiotemporally align environmental sensor data and physiological monitoring data from different sampling frequencies. By establishing a unified timeline, the second-level collected water temperature data and minute-level recorded parent fish heart rate and other heterogeneous data are resampled and interpolated for synchronization. Subsequently, a feature-level fusion strategy is employed to extract biologically significant joint features from the aligned multimodal data. For example, a sliding window correlation analysis is performed between dissolved oxygen change curves and parent fish swimming behavior sequences to construct a fusion feature vector reflecting the environment-physiology interaction. Finally, a structured multimodal time-series data set is generated, in which each time node fully includes co-validated environmental parameters and physiological indicators, laying the foundation for establishing an accurate aquaculture status assessment model.

[0042] Then, the target variable dataset was built using the representation module, where the target variable is an indicator that represents the reproductive performance of the parent fish.

[0043] Using the aforementioned multimodal time-series database, dynamic change patterns are captured from the fused environmental-physiological data through the sliding window technique, and time series decomposition methods are introduced to extract the long-term trends, seasonal fluctuations, and residual components of different parameters.

[0044] In deep feature extraction, a sliding window technique is first applied, with a fixed-length window sliding along the time axis to extract continuous segments from the fused environmental-physiological data, thereby capturing dynamic change patterns such as the coupling relationship between sudden drops in dissolved oxygen and accelerated heart rate in parent fish. Based on this, a time series decomposition method is introduced to decompose each key parameter sequence (such as water temperature and activity level) within the window. Statistical or frequency domain-based algorithms are used to analyze these sequences into long-term trend terms representing slow evolution, seasonal fluctuation terms reflecting intraday or tidal periodic patterns, and random residual components containing information about sudden events. This deconstructs the original complex signal into components at different time scales, thus revealing more clearly the potential causal relationships and driving patterns between environmental factors and physiological responses, providing highly explanatory multi-scale feature inputs for subsequent state assessment models.

[0045] During the time series decomposition process, the system employs a hybrid algorithm based on statistics and frequency domain to perform multi-level analysis of multimodal parameters. For parameters with obvious periodicity, such as water temperature, a classic seasonal decomposition model is used, the additive form of which can be expressed as: in T t This represents the trend item; S t Seasonal component; R t The residuals after removing trend and seasonal factors.

[0046] For non-steady-state physiological signals such as heart rate variability, frequency domain analysis methods are introduced, using Fast Fourier Transform to transform the time-domain information. x ( n Convert to frequency domain representation: Where n represents the time index; N represents the total number of sampling points; and k represents the frequency index.

[0047] Based on the above steps, characteristic frequency band energy distributions corresponding to respiratory metabolism and stress response are identified in the spectrum. This deconstruction method, which combines statistics and the frequency domain, can capture the gradual changes in environmental parameters and the transient rhythmic characteristics of physiological signals, providing a multidimensional feature base that combines macroscopic trends and microscopic fluctuations for subsequent modeling.

[0048] An AI prediction model is built and trained using building blocks, with a multimodal time series database as input features and a target variable set as the prediction target.

[0049] The multi-scale deep feature sequence generated in step S2 is input into a temporal prediction model based on an encoder-decoder architecture. The core of this model is a bidirectional long short-term memory network that incorporates an attention mechanism.

[0050] The specific processing steps are as follows: First, the bidirectional LSTM layer at the encoder reads the input feature sequence in parallel from both forward and backward directions. It processes the information cyclically through its internal gating mechanism to generate a hidden state sequence containing the context information of the entire sequence. Next, the attention layer dynamically calculates the attention weight distribution of the encoder's hidden state at each time step and constructs a context vector accordingly, enabling the model to focus on the historical period most relevant to the prediction target. Finally, the decoder uses this context vector and its own hidden state to output the prediction results and corresponding probabilities of key future aquaculture events (such as the start of spawning) through a fully connected layer. The principle is to enhance the model's ability to capture precursors of important physiological state changes through the attention mechanism, thereby achieving high-precision and interpretable temporal prediction of the parent fish's breeding cycle.

[0051] The optimization module makes water temperature optimization decisions based on a trained AI model: it sets optimization goals, simulates future reproductive performance under different water temperature strategies within a preset water temperature range, and selects the water temperature value that best achieves the optimization goal as the current recommended optimal setting temperature.

[0052] This module transforms the output of deep prediction models into aquaculture management decisions that can directly guide production. Its core is the construction of a closed-loop decision support framework integrating state assessment, strategy deduction, and feedback optimization. This process begins with a multi-dimensional analysis of the prediction results. It not only receives predicted timing and probabilities of key physiological events in broodstock (such as spawning) from S3, but also simultaneously integrates the latest environmental indicators and broodstock behavioral performance from a real-time multimodal database, forming a comprehensive state matrix that predicts the future, describes the present, and compares with historical data.

[0053] First, the matrix is ​​deconstructed and evaluated. For example, if the model predicts an spawning probability of over 85% within the next 72 hours, but real-time dissolved oxygen data shows a downward trend and is approaching the critical threshold of 5 mg / L, and computer vision analysis reveals that the activity level of the parent fish colony is significantly lower than expected, then the state will be judged as "high reproductive potential but with environmental inhibition risk," rather than simply adopting the high-probability prediction. Next, the decision engine is triggered. It reasons based on a knowledge graph containing hundreds of aquaculture procedures and coping strategies. This knowledge graph is organized in a semantic network structure, where nodes represent different aquaculture states or operational entities (such as "aerator turned on" or "feeding adjusted"), and edges represent the logical relationships and transition conditions between them. The engine uses a graph traversal algorithm to find all possible operational paths matching the current evaluation state and uses a case-based reasoning method to retrieve intervention measures that have proven effective in similar scenarios from the historical decision database.

[0054] The decision generation process is a multi-objective optimization problem, requiring the simultaneous balancing of several sometimes conflicting objectives, such as breeding success rate, parent fish welfare, energy costs, and management feasibility. The principle is to construct an objective function that aims to maximize expected benefits while minimizing potential risks and resource consumption. In the example above, multiple candidate strategies might be generated simultaneously: Strategy A is "immediately start the aerator and increase its power to 80%, while adding high-protein feed after 6 hours"; Strategy B is "phased aeration, first running at 50% power for 2 hours and monitoring the dissolved oxygen recovery curve; if it does not reach the expected level, activate the backup pure oxygen supply system." The system uses a simulator to simulate each strategy, predicting its cascading effects on environmental parameters and the physiological state of the parent fish over a future period, and scores each strategy based on the objective function. The final output is not a single instruction, but a priority-ordered decision sequence along with detailed execution plans, expected effects, and risk contingency plans.

[0055] To ensure the accuracy and self-evolution of decision-making, this step also incorporates a tight feedback optimization loop. All issued decision commands, their actual execution (recorded via sensors and actuator logs), and subsequent effects (such as final spawning time, spawn volume, and parent fish health) are collected and stored as new cases with complete contextual tags in the historical decision repository. Simultaneously, the expected results of decisions are periodically compared with the actual outcomes. If a significant deviation is found (e.g., an expected 20% increase in spawn volume but only a 5% increase in actual spawn volume), a root cause analysis process is initiated, tracing back to data quality in S1, feature effectiveness in S2, or model accuracy in S3. The analysis conclusions are then used to trigger model retraining or rule revisions in the knowledge graph.

[0056] This mechanism enables the entire method to continuously learn from practical experience and adapt to the dynamic changes in biological and abiotic factors in the aquaculture environment, thereby gradually evolving from a rule-based automated system into an intelligent aquaculture decision-making brain with cognitive and adaptive capabilities.

[0057] Finally, the sending module sends the determined optimal set temperature to the temperature control system of the aquaculture pond for adjustment.

[0058] The serialized decision schemes generated by S4 are transformed into specific equipment control instructions and manual operation guidance. For example, the decision to "increase the power of the aerator to 80%" is parsed into a specific frequency signal of the frequency converter and sent to the field control.

[0059] Simultaneously, a visual task list is generated for aquaculture managers, including operational items, key indicator warning thresholds, and review time points, and pushed to them via mobile terminals. The execution status and effects of all instructions are tracked in real time through a sensor network, forming a closed loop of "decision-execution-feedback". The core of this stage is to seamlessly embed optimization strategies into the actual production process, and drive the entire intelligent aquaculture system to dynamically adjust and iteratively optimize by continuously comparing expected and actual environmental and physiological response data, ultimately achieving precise implementation of management decisions and a steady improvement in aquaculture efficiency.

[0060] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An AI-based method for determining the temperature of a Trachurus japonicus broodstock breeding water body, characterized by, The method comprises the following steps: Real-time collection of breeding environment data and parent fish physiological data, and generation of a multi-modal time series database based on the collected data; Based on the multi-modal time series database, a target variable dataset representing the breeding performance of parent fish is established; An AI prediction model is constructed and trained, with the multi-modal time series database as input features and the target variable dataset as prediction targets; Based on the trained AI prediction model, an optimization target is set, and the future breeding performance under different water temperature strategies within the preset water temperature range is simulated, and the water temperature value that can best achieve the optimization target is selected as the recommended setting temperature; The recommended setting temperature is sent to the temperature control system of the breeding pond for water temperature regulation.

2. The AI-based Trachurus japonicus broodstock aquaculture water temperature determination method according to claim 1, characterized in that, The method for generating the multi-modal time series database comprises: Anomaly detection and missing value processing are performed on the collected breeding environment data and parent fish physiological data to clean the data; Based on the cleaned data, spatio-temporal alignment and feature-level fusion are performed to generate a multi-modal time series database. 3.The AI-based Trachurus japonicus broodstock aquaculture water temperature determination method of claim 1, wherein, The method for establishing the target variable dataset comprises: Using a sliding window technique to extract dynamic change patterns from the multi-modal time series database; Based on the extracted dynamic change patterns, time series decomposition is performed to obtain long-term trends, seasonal fluctuations, and residual components as features.

4. The AI-based Trachurus japonicus broodstock aquaculture water temperature determination method according to claim 1, characterized in that, The steps for constructing and training the AI prediction model comprise: The multi-modal time series database is input into a time series prediction model based on an encoder-decoder architecture as input features; The input feature sequence is processed by a bidirectional long short-term memory network on the encoder side to generate a hidden state sequence; The attention mechanism is used to dynamically calculate the attention weight of the decoder on the hidden state sequence, and a context vector is constructed based on the attention weight; The context vector and the hidden state of the decoder are used to output the prediction result through a fully connected layer.

5. The AI-based Trachurus japonicus broodstock aquaculture water temperature determination method according to claim 1, characterized in that, The steps for making the water temperature optimization decision comprise: Based on the output of the AI prediction model and the real-time multi-modal time series database, the current breeding state is evaluated; Strategy reasoning is performed based on a knowledge graph to generate several candidate water temperature strategies; The impact of each candidate water temperature strategy on future breeding performance is deduced based on a simulator, and the optimal water temperature value is selected based on the optimization target.

6. The AI-based Trachurus japonicus broodstock aquaculture water temperature determination method according to claim 1, characterized in that, The steps for sending the recommended setting temperature to the temperature control system for water temperature regulation comprise: The recommended setting temperature is parsed into device control instructions and sent to the temperature control system; Based on the device control instructions, a visual task list is generated for manual operation guidance; The sensor network is used to track the instruction execution status in real time, and the system is optimized based on the feedback data.

7. The AI-based Trachurus japonicus broodstock aquaculture water temperature determination method according to claim 6, characterized in that, Further comprising: After sending, the actual effect data after decision execution is collected; Based on the comparison result of the actual effect data and the expected effect, model retraining or rule revision is triggered.

8. An AI-based Trachurus japonicus broodstock aquaculture water temperature determination system for implementing the method of any one of claims 1-7, characterized in that, Comprise: The collection module, the representation module, the construction module, the optimization module, and the sending module.

Images