Intensive breeding precise feeding control method based on fuzzy-MPC control

Abstract

The invention discloses an intensive culture precise feeding control method based on fuzzy-MPC control, and relates to the technical field of aquaculture intelligent control, and the method comprises the steps: employing a sensing system, collecting multi-source sensing data including fish activity images, healthy growth states and water quality environment data in real time, and forming a comprehensive state data set; and based on the acquired multi-source sensing data, constructing an MTL-LSTM-SAT multi-task learning model. According to the method, the multi-task learning model fusing the LSTM and the soft attention mechanism is constructed, the long-term dependency relationship and dynamic association among the water quality, the feeding and the fish state can be captured, the attention mechanism enables the model to be automatically focused on a key time step, the perception ability of the water quality fluctuation and the feeding influence stage is enhanced, and the accuracy of the water quality fluctuation and feeding influence stage is improved. Meanwhile, double-task collaborative prediction of the growth speed and the disease proportion is achieved through the shared feature layer, a prospective state sequence of the future 7 days is output, and a quantitative and reliable prediction basis is provided for feeding decision making.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for aquaculture, specifically a precise feeding control method for intensive aquaculture based on fuzzy-MPC control. Background Technology

[0002] With the improvement of living standards, the demand for fish products is increasing, and traditional river, lake and pond aquaculture can hardly meet the demand. Intensive high-density aquaculture has become the mainstream. However, in this model, feed feeding depends on the experience and labor of farmers. Automatic feeding equipment has problems such as delayed response to fish diseases and feed waste. Among them, the inaccuracy of feeding amount and timing is an important reason for fish diseases and feed waste. Although some studies have integrated multiple technologies to achieve precise feeding, they lack in-depth analysis of fish health and growth status, and the data interaction between water quality sensors and vision systems is insufficient. Therefore, the development of a feeding system that can combine multi-source data and dynamically adjust feeding strategies has become an inevitable requirement for precise feeding control in intensive aquaculture.

[0003] Existing intensive aquaculture feeding control methods mainly rely on timed and quantitative strategies or feedback based on single apparent indicators such as water quality and uneaten feed. They lack direct integrated analysis of fish health and growth status, leading to a disconnect between feeding decisions and the actual physiological needs of the fish population. Furthermore, these methods struggle to effectively handle the significant time lag between feeding behavior and fish growth and health status, failing to achieve multi-objective collaborative optimization based on long-term prediction. In addition, existing systems often operate vision, sensing, and control modules in isolation, lacking multi-source information fusion and collaborative decision-making capabilities. Moreover, single control strategies cannot simultaneously address real-time rapid response and long-term economic benefit optimization, thus restricting the accuracy, adaptability, and overall aquaculture efficiency of the feeding process. Therefore, this paper proposes a precision feeding control method for intensive aquaculture based on fuzzy-MPC control to solve the aforementioned problems. Summary of the Invention

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for precise feeding control in intensive aquaculture based on fuzzy-MPC control, comprising the following steps: S1. Employ a sensing system that includes underwater cameras and water quality sensors to collect multi-source sensing data in real time, including images of fish activity, health status, and water quality environment data, forming a comprehensive status dataset. S2. Based on the collected multi-source sensing data, construct an MTL-LSTM-SAT multi-task learning model, integrate soft attention mechanism, predict the growth rate and disease rate of fish in the next 7 days, and guide the formulation of feeding strategies. S3. A fuzzy control layer is designed using a fuzzy control algorithm, and a fuzzy control rule table is formulated with morbidity and growth rate as inputs to achieve fast and robust feeding suggestions based on the current state. S4. Using a hierarchical control strategy, the lower-level fuzzy controller and the upper-level MPC controller are divided into a lower-level fuzzy controller and an upper-level MPC controller, which respectively receive the short-term state output by the model and the long-term prediction value of the MTL-LSTM-SAT multi-task learning model. S5. The fuzzy controller outputs a primary recommended feeding level based on real-time predicted status through fuzzy inference to cope with sudden situations such as water quality changes and ensure system response speed and stability. S6. The upper-level MPC controller uses the primary recommended feeding level output by the fuzzy controller as the reference trajectory, and combines the long-term predicted value output by the MTL-LSTM-SAT multi-task learning model. With the goal of maximizing growth and minimizing disease, it performs a constrained 7-day rolling optimization, balances growth and health goals, fine-tunes the feeding amount, and calculates a feeding strategy sequence that is optimal in both the immediate and long-term. S7. Convert the control strategy in the feeding strategy sequence into a feeding instruction, send it to the feeding actuator to complete precise feeding, and collect new multi-source sensing data feedback to realize closed-loop adaptive control based on multi-cycle prediction and optimization.

[0005] Preferably, S1 specifically includes: A sensing system, including an underwater camera and a water quality sensor, is deployed in the rearing tank. The underwater camera takes side shots at different depths along the tank wall, while the water quality sensor is inserted into the water at a fixed depth along the tank wall. Simultaneously, images of fish activity and water quality parameters are acquired, which effectively solves the problem of limited sensing range of traditional feeding systems and realizes all-round real-time monitoring of the multi-dimensional and multi-level aquaculture environment. The underwater camera collected images of fish underwater activities from 6:00 to 18:00 at a sampling frequency of 0.1 frames per second. The water quality sensor collected data on water temperature, pH value, dissolved oxygen and ammonia nitrogen every 10 minutes and recorded the daily feeding amount. Through time-series high-frequency data collection, the continuity and integrity of the aquaculture status data were ensured, providing a reliable data foundation for subsequent accurate analysis and prediction. The collected fish activity image data is input into a visual recognition algorithm to identify the fish's health status and key growth points. It is then fused with water quality data and feeding data to construct a comprehensive status dataset containing time-series features for subsequent prediction modeling. This achieves deep fusion of multi-source heterogeneous data and constructs a comprehensive dataset covering fish status, water quality, and feeding behavior, thereby improving the comprehensiveness of status representation and the accuracy of prediction modeling.

[0006] Preferably, S2 specifically includes: By combining a comprehensive state dataset, an LSTM network framework based on multi-task learning is constructed. The input is three-dimensional time series data containing water quality, feeding amount and fish status. Temporal features are extracted through LSTM layers and hidden states are output, realizing effective modeling of long-term temporal dependencies and providing a feature foundation for multi-task prediction. A soft attention mechanism is introduced to weight and focus the hidden state of the LSTM output, calculate the attention weight at each time step, enhance the perception of key water quality fluctuations and feeding impact stages, improve the model's sensitivity to key time series information, and avoid interference from invalid historical information. By generating a context vector through weighted summation and inputting it into the shared feature layer, the system outputs the predicted growth rate and disease rate of fish for the next 7 days, thus achieving synergistic prediction of growth and health and serving as a forward-looking basis for feeding decisions.

[0007] Preferably, S2 further includes: The model was trained using water quality, feeding, and fish growth and health data from historical aquaculture cycles. Mean squared error and cross-entropy were used as multi-task loss functions. The network parameters were optimized through backpropagation to ensure that the prediction model had good generalization and fitting capabilities on historical data. The model performance was evaluated on the validation set. The number of LSTM units, the number of attention heads, and the fully connected layer structure were adjusted to achieve the optimal balance between growth rate prediction and disease proportion prediction tasks. The model structure was optimized to improve the accuracy and robustness of prediction. The trained MTL-LSTM-SAT multi-task learning model is embedded into the feeding control system, which receives multi-source sensing data in real time and outputs the growth and disease prediction sequence for the next 7 days in a rolling manner. This achieves seamless integration between the prediction model and the control system, supporting real-time decision-making.

[0008] Preferably, S3 specifically includes: A fuzzy control layer was designed using a fuzzy control algorithm, with morbidity and growth rate as input variables and feeding level as output variables. The input variables were divided into low / medium / high and slow / normal / fast fuzzy sets, and the output variables were divided into five levels: ultra-low / low / medium / high / ultra-high. Through multi-level fuzzy division, the precision and adaptability of feeding control were improved. Based on the knowledge of aquaculture experts, a fuzzy control rule table containing nine rules was developed. Triangular, trapezoidal, and Gaussian membership functions were used to describe the fuzziness of the input and output variables, respectively. The rule table based on expert experience ensures that the control strategy conforms to aquaculture practice and improves the credibility of decision-making.

[0009] Preferably, S4 specifically includes: The feeding control system is divided into a lower-level fuzzy controller and an upper-level MPC controller. The lower level responds to real-time prediction of the state in a high-frequency manner, while the upper level performs long-term rolling optimization in a low-frequency manner, so as to achieve the synergy between the system's fast response and long-term optimization, and to take into account both real-time performance and global optimality. The fuzzy controller receives the short-term state output from the MTL-LSTM-SAT multi-task learning model, which covers the current prevalence and growth rate. The MPC controller receives the output from the fuzzy controller and the 7-day prediction sequence output from the MTL-LSTM-SAT multi-task learning model. The hierarchical structure effectively integrates the short-term state and long-term prediction, supporting dynamic and forward-looking decision optimization.

[0010] Preferably, S5 specifically includes: Based on the predicted prevalence and growth rate values ​​at the current moment output by the MTL-LSTM-SAT multi-task learning model, the activation intensity of each fuzzy rule is calculated through the membership function to effectively quantify the degree of rule matching and ensure the objectivity of the decision-making basis. Based on the fuzzy control rule table, reasoning is performed to obtain the feeding level output value corresponding to each rule. The weighted average method is used for defuzzification calculation to smooth the output of feeding suggestions and improve the continuity and stability of control decisions. The system outputs a primary recommended feeding level, which is used as a reference input for the MPC controller. In the event of sudden changes in water quality, the system can be used to quickly adjust the feeding behavior, enabling real-time emergency response and ensuring the safe operation of the system under emergencies. The formula for calculating the primary recommended feeding level is as follows: ; In the formula: This is the initial recommended feeding level; For the nonlinear mapping function of the fuzzy controller; , These represent the disease rate and average growth rate of the fish population, respectively. Total number of rules (9 rules); For the first Fuzzy rules in input The activation intensity below; For the first The output feeding level corresponding to each fuzzy rule (e.g., the defuzzified value of "low", "medium", "high" etc.).

[0011] Preferably, S6 specifically includes: Using the initial recommended feeding level output by the fuzzy controller as a reference trajectory, a multi-objective optimization function is constructed, which includes a growth deviation objective function, a disease deviation objective function, a control quantity deviation penalty and a smoothing control term. This achieves multi-objective collaborative optimization and balances growth promotion, disease inhibition and operational stability. The MTL-LSTM-SAT multi-task learning model is used as an internal prediction model to predict changes in growth rate and morbidity under different feeding strategies in the next 7 days. The physical constraint that the feeding amount is non-negative and does not exceed 5% of the fish body weight is embedded to ensure the feasibility and safety of control commands and avoid the generation of invalid or harmful commands. At each decision moment, a finite-time domain optimization problem is solved to calculate a series of future feeding strategies. Only the optimal feeding amount at the current moment is output as the final control command, realizing rolling optimization and real-time feedback, and ensuring that the control decision dynamically adapts to environmental changes.

[0012] Preferably, S6 further includes: The initial values ​​of the growth deviation weight coefficient, disease deviation weight coefficient, fuzzy output reference value weight coefficient, and smoothing control term weight coefficient are set. The weight coefficients are dynamically adjusted according to the real-time prediction status to achieve multi-objective adaptive optimization and improve the dynamic response and long-term stability of the system. When the growth rate decreases, the growth deviation weight coefficient is increased; when the morbidity rate increases, the morbidity deviation weight is increased; when the equipment control quantity fluctuates too much, the smoothing control weight is increased to ensure that the control strategy focuses on the main contradictions and maintains a dynamic balance between growth and health. By using a weighted adaptive mechanism to dynamically balance control strategies, the feeding control system can achieve long-term optimal decision-making at different breeding stages and fish population conditions, thereby enhancing the system's environmental adaptability and realizing optimal feeding management throughout the entire cycle.

[0013] Preferably, S7 specifically includes: The optimal feeding amount in the feeding strategy sequence output by the MPC controller is converted into a pulse width modulation signal, which drives the servo motor of the feeder and the screw feeder to perform precise feeding. The optimization decision is accurately mapped into physical control action, realizing the quantitative and timed control of feed feeding. After feeding, fish images and water quality data are re-acquired, the input of the MTL-LSTM-SAT multi-task learning model is updated, and a new round of state prediction and control decision-making is initiated. Real-time data updates ensure that state prediction is synchronized with the aquaculture environment, maintaining the timeliness and accuracy of decision-making. Through continuous multi-cycle data acquisition, prediction optimization, and execution feedback, a closed-loop adaptive feeding control system is formed, realizing dynamic adjustment of feeding strategies and long-term performance optimization, establishing a self-adjusting cycle mechanism, and enabling the system to have continuous adaptability and long-term benefit optimization potential.

[0014] This invention provides a method for precise feeding control in intensive aquaculture based on fuzzy-MPC control. It has the following beneficial effects: (I) This intensive aquaculture precision feeding control method based on fuzzy-MPC control constructs a multi-task learning model that integrates LSTM and soft attention mechanisms. It can capture the long-term dependence and dynamic correlation between water quality, feeding and fish status. The attention mechanism enables the model to automatically focus on key time steps, enhancing the perception of water quality fluctuations and feeding impact stages. At the same time, by sharing feature layers, it achieves dual-task collaborative prediction of growth rate and disease rate, outputting a prospective state sequence for the next 7 days, providing quantitative and reliable prediction basis for feeding decisions.

[0015] (II) This intensive aquaculture precision feeding control method based on fuzzy-MPC control constructs a comprehensive status dataset covering fish activity images, key growth points, water quality parameters, and feeding amount through the collaborative deployment of underwater binocular cameras and water quality sensors. Combined with visual recognition algorithms, it can extract fish health characteristics and growth indicators in real time, perform time alignment and feature fusion with water quality data, and form a structured multi-dimensional time series data input, providing a comprehensive and accurate data foundation for predictive modeling, effectively overcoming the limitations of traditional methods that rely on a single appearance indicator.

[0016] (III) This intensive aquaculture precision feeding control method based on fuzzy-MPC control adopts a fuzzy control algorithm, which encodes the experience of aquaculture experts into a rule base with disease rate and growth rate as input. The input and output are fuzzy mapped by triangular, trapezoidal and Gaussian membership functions. It can complete state assessment and reasoning in milliseconds and output primary feeding suggestions. It can not only serve as a reference trajectory for upper-level optimization, but also directly drive the actuator in case of sudden changes in water quality, so as to achieve rapid response and robust control of the system.

[0017] (iv) This intensive aquaculture precision feeding control method based on fuzzy-MPC control, through the hierarchical architecture of dividing the lower-level fuzzy controller and the upper-level MPC controller, the system has both real-time response capability and long-term rolling optimization capability. The MPC uses fuzzy suggestions as reference trajectories and combines the output of the multi-task prediction model to construct a multi-objective optimization function including growth, disease, tracking and smoothing terms. The weight of each objective is dynamically adjusted through a weight adaptive mechanism to achieve intelligent optimization and multi-objective balance of feeding strategies in different aquaculture stages and fish population conditions. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the workflow of a precision feeding control method for intensive aquaculture based on fuzzy-MPC control according to the present invention. Figure 2This is a layout diagram of the sensor and camera of the present invention, wherein A is a physical diagram of the sensor and camera arrangement, and B is a schematic diagram of the sensor and camera arrangement positions; Figure 3 This is a schematic diagram of the multi-task learning model (MTL-LSTM-SAT) that combines LSTM and soft attention mechanisms according to the present invention. Figure 4 This is a schematic diagram of the membership functions of the input and output variables of the present invention, wherein (a) is the membership function of disease rate, (b) is the membership function of growth rate, and (c) is the membership function of feeding level; Figure 5 This is a flowchart of the rolling optimization mechanism of the present invention; Figure 6 This is a graph showing the feeding level curves under the five feeding strategies of this invention; Figure 7 This is a growth rate curve diagram under the five feeding strategies of this invention; Figure 8 This is a disease incidence curve under the five feeding strategies of this invention; Figure 9 This is a comparison chart of the breeding effects under different feeding strategies according to the present invention; Figure 10 This is a comparison chart of dissolved oxygen levels in two groups of feeding trials according to the present invention (7:00 AM on September 11th to 7:00 PM on September 12th). Figure 11 This is a comparison chart of unit breeding benefits under the three feeding modes of the present invention, with the unit being yuan; Figure 12 This is a hardware architecture diagram of the feeder motor control system of the present invention.

[0019] In the diagram: 1. Power supply module; 2. STM32 main controller; 3. Drive module; 4. Stepper motor; 5. Feeder housing; 6. Encoder. Detailed Implementation

[0020] 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.

[0021] Example 1, please refer to Figures 1 to 5 This invention provides a technical solution: a method for precise feeding control in intensive aquaculture based on fuzzy-MPC control, comprising the following steps: S1. A sensing system encompassing underwater cameras and water quality sensors is employed to collect multi-source sensing data in real time, including images of fish activity, health status, and water quality data, forming a comprehensive status dataset. The sensing system, including underwater cameras and water quality sensors, is deployed in the rearing tank. The underwater cameras take side-view images at different depths along the tank wall, while the water quality sensors are inserted into the water at a fixed depth along the tank wall, simultaneously acquiring images of fish activity and water quality parameters. This effectively solves the problem of limited sensing range in traditional feeding systems, achieving comprehensive real-time monitoring of the multi-dimensional and multi-layered aquaculture environment. The underwater cameras collect underwater fish activity images from 6:00 to 18:00 at a sampling frequency of 0.1 frames / second. For example, water quality sensors collect data on water temperature, pH, dissolved oxygen, and ammonia nitrogen every 10 minutes and record the daily feed amount. Through time-series high-frequency data collection, the continuity and integrity of aquaculture status data are ensured, providing a reliable data foundation for subsequent accurate analysis and prediction. The collected fish activity image data is input into a visual recognition algorithm to identify the fish's health status and key growth points. This data is then fused with water quality data and feed amount data to construct a comprehensive status dataset containing time-series features for subsequent predictive modeling. This achieves deep fusion of multi-source heterogeneous data and constructs a comprehensive dataset covering fish status, water quality, and feeding behavior, improving the comprehensiveness of status representation and the accuracy of predictive modeling. It should be noted that the data was collected from the Modern Aquaculture Base in the Agricultural High-tech Zone of Xian'an District, Xianning City, Hubei Province. The aquaculture mode was intensive tub-type culture, with tubs having a diameter of 4 meters and a water depth of 2 meters. The stocking density was 2500 fish per tub, and the species cultured was red tilapia (rainbow sea bream). Fish were raised from fry with an average length of 10 cm, and the culture period was from March 31st to December 31st each year. Figure 2 As shown, the sensor was inserted 1m deep into the water along the side of the barrel, and the binocular camera also took pictures along the side of the barrel, collecting images of fish activity at depths of 0.6m, 1.2m, and 1.8m. The positions are as follows. Figure 2 As shown, the water quality sensor is a customized REMOND online water quality parameter monitoring instrument, and the underwater camera is an OAK-D-Pro wide-angle intelligent depth camera. The system technical parameters are detailed in Table 1. Table 1. Data Technical Parameters: .

[0022] The specific work involves physically deploying a sensing system inside the rearing tanks. This includes installing underwater cameras (underwater binocular cameras) at multiple fixed points along the inner wall of the tank, selecting different water depths (0.6 meters, 1.2 meters, and 1.8 meters) to cover the main activity areas of the fish and achieve continuous side-view imaging. A water quality sensor is vertically inserted into the water to a fixed depth (1 meter) along the tank wall using a dedicated bracket, ensuring full contact between the probe and the water being measured. Both types of equipment are synchronized via a unified timing module to ensure data acquisition time alignment. In actual operation, the underwater cameras continuously acquire underwater video streams at a preset sampling frequency, while the water quality sensor automatically measures multiple water quality parameters at a set cycle. The system automatically records the daily feed input, forming a structured multi-source data acquisition system. After data acquisition, the system enters the preprocessing and feature extraction stage. The video stream collected by the underwater camera is processed by frame extraction and then input into a deep learning-based visual recognition algorithm to automatically detect individual fish, identify abnormal behaviors and pathological features on the body surface, and extract the pixel coordinates of key growth points such as body length and width. The raw data collected by the water quality sensor is filtered, outlier removed, and standardized by unit. It is then aligned and fused with the visual feature data according to the timestamp, and the daily feed input data is integrated as a control variable. Finally, a comprehensive status dataset is formed in time order, which includes fish status indicators, water quality parameters, and control inputs. S2. Based on the collected multi-source sensing data, an MTL-LSTM-SAT multi-task learning model is constructed, integrating a soft attention mechanism to predict the growth rate and disease rate of fish over the next 7 days, guiding feeding strategy formulation. Combining a comprehensive state dataset, an LSTM network framework based on multi-task learning is constructed. The input is three-dimensional time-series data including water quality, feed quantity, and fish status. Temporal features are extracted through LSTM layers, and hidden states are output, achieving effective modeling of long-term temporal dependencies and providing a feature foundation for multi-task prediction. A soft attention mechanism is introduced to weight and focus the hidden states output by the LSTM, calculating the attention weights at each time step to enhance the perception of key water quality fluctuations and feeding impact stages, improving the model's sensitivity to key temporal information, and avoiding interference from invalid historical information. A context vector is generated through weighted summation and input into the shared feature layer, simultaneously outputting the predicted growth rate and disease rate of fish over the next 7 days, achieving collaborative prediction of growth and health, serving as a forward-looking basis for feeding decisions. The workflow of this model is as follows: Figure 3 As shown; The working principle of each layer in the diagram is as follows: Input layer: Transforms the preprocessed data into a three-dimensional matrix adapted for time series feature extraction, with type dimension S, time dimension T, and feature input dimension X; LSTM layer: Input data step by step, LSTM units process time dependencies and output hidden states. ; Soft attention layer: based on the hidden state output by the LSTM layer Calculate attention weights This allows us to focus on key time steps in long-term series to avoid forgetting important information. Weighted summation layer: Performs weighted summation to generate context vector Z; Shared feature layer: Shared features are passed to the output layer or decoder branch of different tasks to achieve different objectives; Output layer: Outputs the predicted results. In this study, the output results are mainly growth rate and disease rate. The detailed construction process, training method, parameter optimization, and performance evaluation of this model have been compared and tested. The results show that the model has a mean absolute error (MAE) of 0.423 g / day and a root mean square error (RMSE) of 0.598 g / day in predicting fish growth rate. In the disease incidence prediction task, the MAE is 0.042 and the RMSE is 0.061, which are significantly improved compared with traditional RNN, GRU, BiLSTM, LSTM and single-task models. Based on this validated model, we focus on its downstream application in feeding control strategies. The specific work involves constructing a Long Short-Term Memory (LSTM) network framework based on multi-task learning to process time-series data from aquaculture environments. The network input is three-dimensional time-series data, with the three dimensions corresponding to data type, time step size, and feature dimension, respectively. The data type covers water quality parameters, feed amount, and extracted fish state indicators. The core of the framework is the LSTM layer, which receives the preprocessed and standardized comprehensive state dataset and processes the input sequentially according to time steps. The LSTM layer, through its internal gating mechanism, can effectively capture long-term dependencies and dynamic change patterns in the data, outputting the corresponding hidden state vectors for each time step. The hidden state vectors encode the comprehensive system state information up to the current time step. A soft attention mechanism is introduced after the LSTM layer. The hidden state sequence of all time steps output by the previous LSTM stage is used as input to calculate a learnable attention weight distribution. This distribution is generated by a small feedforward neural network combined with a nonlinear activation function to evaluate the importance of different historical time points to the current prediction task. The calculation of attention weights enables the model to automatically cluster. Instead of treating all historical information equally, this approach focuses on periods of drastic water quality fluctuations or specific feeding phases that are most relevant to changes in fish growth rate and health status. It selectively strengthens or suppresses historical information by weighting the hidden states at each time step according to their attention weights. The weighted hidden states at each time step are then summed to generate a fixed-length context vector, which integrates the most informative state features within the entire attention window. This vector serves as a condensed input for subsequent prediction tasks and is then fed into a shared feature layer for further feature transformation and abstraction. Common high-order features valuable for multiple prediction tasks are extracted. The network then branches into two independent output branches, corresponding to the growth rate prediction task and the disease rate prediction task, respectively. Each branch consists of a fully connected layer that maps the shared features to a specific prediction target space. Finally, the model simultaneously outputs the predicted average growth rate of fish over the next seven consecutive days and the predicted disease rate of the fish population. These two outputs together constitute a direct, quantitative, and forward-looking decision-making basis for formulating feeding control strategies. S2 also includes: using water quality, feeding, and fish growth and health data from historical aquaculture cycles for model training; employing mean squared error and cross-entropy as multi-task loss functions; optimizing network parameters through backpropagation to ensure the prediction model has good generalization and fitting capabilities on historical data; evaluating model performance on the validation set; adjusting the number of LSTM units, attention heads, and fully connected layer structure to achieve an optimal balance between growth rate prediction and disease rate prediction tasks; optimizing the model structure to improve prediction accuracy and robustness; embedding the trained MTL-LSTM-SAT multi-task learning model into the feeding control system; receiving multi-source sensing data in real time; and continuously outputting the growth and disease prediction sequences for the next 7 days, thus achieving seamless integration of the prediction model and the control system and supporting real-time decision-making. The specific work involves: constructing a dataset for model training based on water quality parameters, feeding records, and corresponding fish growth and health monitoring data collected during historical aquaculture cycles. This dataset covers the entire aquaculture period, ensuring the continuity of the time series and the authenticity of state evolution. Mean squared error is used as the loss function for growth rate prediction, and cross-entropy is used as the loss function for disease proportion classification prediction. The weighted combination of these two functions constitutes a multi-task learning objective. The network weights are iteratively optimized using the backpropagation algorithm, and the learning rate is adjusted using an adaptive moment estimation optimizer to ensure stable convergence during training. An early stopping strategy is employed during training to prevent overfitting, and the best-performing model parameters on the validation set are retained. The model's predictive performance is systematically evaluated on an independent validation set. The root mean squared error of growth rate prediction and the accuracy and recall of disease proportion prediction are analyzed. Based on the validation results, targeted adjustments are made to the network hyperparameters, including adding or removing layers in the LSTM layer. The number of memory units is adjusted to balance long-term reliance on capture capability and model complexity; the number of heads in the multi-head attention mechanism is adjusted to enhance the model's ability to focus on multi-dimensional temporal features; the hierarchical structure and number of neurons in the fully connected layer are optimized to improve feature mapping and task adaptability; through multiple cross-validation and grid search, the final network structure that achieves the optimal balance between growth prediction and health prediction tasks is determined; the trained and validated MTL-LSTM-SAT multi-task learning model is integrated into the software framework of the feeding control system to realize an automated process from data reception, preprocessing to real-time inference. After the model is deployed, it continuously receives multi-source temporal data uploaded in real time by the sensing system, automatically performs standardization and alignment processing according to time windows, and inputs it into the network for forward computation. The system uses a rolling prediction method to output the fish-specific growth rate curve and disease proportion trend sequence for the next 7 consecutive days based on the latest data sequence at each decision moment; S3. A fuzzy control layer is designed using a fuzzy control algorithm. A fuzzy control rule table is formulated with disease rate and growth rate as inputs to achieve rapid and robust feeding suggestions based on the current state. The fuzzy control layer is designed with disease rate and growth rate as input variables and feeding level as output variables. The input variables are divided into low / medium / high and slow / normal / fast fuzzy sets, and the output variables are divided into five levels: ultra-low / low / medium / high / ultra-high. Through multi-level fuzzy division, the precision and adaptability of feeding control are improved. A fuzzy control rule table containing 9 rules is formulated based on the knowledge of aquaculture experts. Triangular, trapezoidal and Gaussian membership functions are used to describe the fuzziness of input and output variables, respectively. The rule table based on expert experience ensures that the control strategy conforms to aquaculture practice and improves the credibility of decision-making. The specific work involves designing and deploying a control layer based on fuzzy control algorithms within an intensive aquaculture precision feeding control system. This layer enables rapid decision-making and response based on the real-time status of fish. The control layer selects fish disease rate and average specific growth rate as input variables, and the percentage of daily feed intake relative to the average fish mass as the output variable, i.e., the feeding level. The fuzzy classification of the input variables follows empirical knowledge in aquaculture management: disease rate is divided into three fuzzy sets—low, medium, and high—corresponding to different disease risk levels; growth rate is divided into three fuzzy sets—slow, normal, and fast—reflecting the physiological metabolism and nutritional needs of the fish population. The output variable, feeding level, is finely divided into five levels: ultra-low, low, medium, high, and ultra-high, to achieve gradient adjustment of feed intake. By converting precise sensor data into fuzzy language variables that conform to human expert experience, a basis for simulating human decision-making is provided. A hierarchical logic framework is constructed, defining corresponding membership functions for each variable to describe its fuzziness and achieving fuzzy mapping between input and output variables. The input variable, disease rate, is characterized by a triangular membership function, with the low, medium, and high fuzzy sets overlapping reasonably in the domain to ensure smooth state transitions. The input variable, growth rate, is described by a trapezoidal membership function, allowing it to accommodate speed fluctuations within a certain range and clearly define the boundaries between slow, normal, and fast. The output variable, feeding level, is defined by a Gaussian membership function, utilizing its continuous differentiability and well-defined center to generate smooth and precise control outputs. Based on long-term accumulated knowledge from aquaculture experts, a fuzzy control rule base containing nine IF-THEN rules is constructed. Each rule clarifies the appropriate feeding level under different combinations of disease rate and growth rate, encoding expert experience into automatically executable reasoning logic. It should be noted that the fuzzy control module selects morbidity rate (range 0%–100%) and growth rate (range 0–5 g / day) as input variables, and feed amount (5% of average body weight per tail) as output variable (range 0%–5%). Based on relevant literature and breeding experience, a fuzzy control (IF-THEN) rule base (Table 2) is established for fuzzy decision output. The membership degrees of input and output variables are as follows: Figure 5 To describe the fuzziness of the variables, triangular (morbidity), trapezoidal (growth rate), and Gaussian (feeding level) membership functions were used respectively. Morbidity was divided into low, medium, and high; growth rate was divided into slow, normal, and fast; and feed level was divided into very low, low, medium, high, and very high. Table 2. Fuzzy Control Rule Table: .

[0023] S4. Using a hierarchical control strategy, the system is divided into a lower-level fuzzy controller and an upper-level MPC controller. The lower-level controller receives the short-term state output by the model and the long-term prediction value of the MTL-LSTM-SAT multi-task learning model, respectively. The feeding control system is divided into a lower-level fuzzy controller and an upper-level MPC controller. The lower-level controller responds to the real-time prediction state in a high-frequency manner, while the upper-level controller performs long-term rolling optimization in a low-frequency manner. This achieves the synergy between the system's rapid response and long-term optimization, taking into account both real-time performance and global optimality. The fuzzy controller receives the short-term state output by the MTL-LSTM-SAT multi-task learning model, which covers the current morbidity rate and growth rate. The MPC controller receives the output of the fuzzy controller and the 7-day prediction sequence output by the MTL-LSTM-SAT multi-task learning model. The hierarchical structure effectively integrates the short-term state and long-term prediction, supporting dynamic and forward-looking decision optimization. The specific work involves the following steps: In the actual operation of the feeding control system, the configuration and deployment of the lower-level fuzzy controller are first completed. This fuzzy controller is embedded in the real-time control module in a high-frequency operating mode, continuously receiving data on the current fish disease rate and average specific growth rate from the output of the MTL-LSTM-SAT multi-task learning model. The input values ​​are converted into corresponding fuzzy linguistic variables through a fuzzification interface, and real-time inference is performed based on preset membership functions and a fuzzy rule base. The inference process simulates human expert experience, completing the assessment and decision-making of the current aquaculture status within milliseconds, and outputting a preliminary feeding level recommendation. This recommendation serves as a rapid response command to sudden situations (such as drastic changes in water quality parameters) and can be directly used to drive the actuators to adjust the feeding, ensuring the system... The system demonstrates agility and robustness at the real-time control level. An upper-level MPC controller is deployed independently of the fuzzy controller, performing optimization calculations periodically at a low frequency. At each decision point (e.g., daily scheduled startup), the MPC controller synchronously receives primary feeding suggestions from the lower-level fuzzy controller, as well as the predicted sequence of fish growth rate and disease prevalence for the next 7 consecutive days, continuously output by the MTL-LSTM-SAT multi-task learning model. The MPC utilizes its built-in prediction model (i.e., the MTL-LSTM-SAT multi-task learning model) to construct a multi-objective optimization problem within a finite time domain based on the current system state and the future prediction sequence. This problem aims to promote growth, suppress disease, track fuzzy suggestions, and smooth control inputs, while strictly embedding physical constraints on feeding amounts. S5. The fuzzy controller outputs a primary recommended feeding level based on real-time predicted status through fuzzy inference to cope with sudden situations such as water quality changes and ensure system response speed and stability. S6. The upper-level MPC controller uses the primary recommended feeding level output by the fuzzy controller as the reference trajectory, and combines the long-term predicted value output by the MTL-LSTM-SAT multi-task learning model. With the goal of maximizing growth and minimizing disease, it performs a constrained 7-day rolling optimization, balances growth and health goals, fine-tunes the feeding amount, and calculates a feeding strategy sequence that is optimal in both the immediate and long-term. S7. Convert the control strategy in the feeding strategy sequence into a feeding instruction, send it to the feeding actuator to complete precise feeding, and collect new multi-source sensing data feedback to realize closed-loop adaptive control based on multi-cycle prediction and optimization.

[0024] Example 2, as Figures 1 to 5As shown, based on Embodiment 1, the present invention provides a technical solution: S5 specifically includes: based on the predicted values ​​of morbidity and growth rate at the current moment output by the MTL-LSTM-SAT multi-task learning model, calculating the activation intensity of each fuzzy rule through the membership function, effectively quantifying the rule matching degree, ensuring the objectivity of the decision basis, reasoning based on the fuzzy control rule table to obtain the feeding level output value corresponding to each rule, using the weighted average method for defuzzification calculation, smoothing the output of feeding suggestions, improving the continuity and stability of control decisions, outputting the primary recommended feeding level, using the primary recommended feeding level as the reference input of the MPC controller, and directly using it to quickly adjust the feeding behavior in the event of a sudden change in water quality, realizing real-time emergency response, and ensuring the safe operation of the system in the event of an emergency; The formula for calculating the primary recommended feeding level is as follows: ; In the formula: This is the initial recommended feeding level; For the nonlinear mapping function of the fuzzy controller; , These represent the disease rate and average growth rate of the fish population, respectively. Total number of rules (9 rules); For the first Fuzzy rules in input The activation intensity below; For the first The output feeding level corresponding to each fuzzy rule (e.g., the defuzzified value of "low", "medium", "high" etc.); The specific tasks are as follows: The fuzzy controller continuously receives the predicted values ​​of fish morbidity and average specific growth rate from the real-time output of the MTL-LSTM-SAT multi-task learning model. It then performs fuzzification processing on the two precise input values ​​using preset membership functions. Specifically, for the morbidity input, the system calls its triangular membership function to calculate the membership degree of the value to the three fuzzy linguistic variables: low, medium, and high. For the growth rate input, the system calls its trapezoidal membership function to calculate the membership degree of the value to the three fuzzy linguistic variables: slow, normal, and fast. Subsequently, based on the IF- The system calculates the activation strength of each rule in the fuzzy rule base based on its preconditions (i.e., the fuzzy set combination of input variables), quantifying the degree of matching between the current input state and each expert rule. Based on the calculated activation strengths of each rule, the system enters the fuzzy inference stage. The consequent (i.e., the conclusion) of each rule corresponds to a specific feeding level fuzzy set, including ultra-low, low, medium, high, and ultra-high. The system prunes or scales the consequent fuzzy set according to the activation strength of each rule, obtaining a series of activated output fuzzy sets. Then, the system uses a weighted average method to evaluate all activated output fuzzy sets. The process involves aggregating and defuzzifying fuzzy sets. Specifically, the explicit value (the core value of its membership function) corresponding to each output fuzzy set is multiplied by the activation strength of the rule. All products are then summed, and finally divided by the sum of the activation strengths of all rules to obtain a precise, quantified primary recommendation feeding level. This primary recommendation feeding level represents a rapid decision based on the current instantaneous state and the expert rule base. The generated primary recommendation feeding level serves a dual function in the system: firstly, it acts as a reference input signal for the upper-level Model Predictive Control (MPC) optimizer, which then performs subsequent weekly... In the periodic optimization calculation, this suggested amount is included as part of the expected trajectory, and a corresponding tracking term is set in its multi-objective optimization function. This ensures that the final optimized control amount respects the rapid judgment based on expert experience. Secondly, this suggested amount is directly connected to the system's fast execution channel. When the system detects sudden and drastic changes in water quality parameters such as a sudden drop in dissolved oxygen, it can bypass the upper-level periodic optimization process and directly convert this primary recommended feeding level into a control command to drive the feeding execution mechanism to make adjustments. This ensures that the system has the ability to intervene directly at the millisecond level when facing sudden situations, effectively maintaining the steady state of the fish population during stress. like Figure 4As shown, in (a), the morbidity rate, the input variable of the membership function of the disease rate, is divided into three fuzzy sets: "low," "medium," and "high" using a triangular membership function. The "low" morbidity rate covers the interval from 0 to 30%, intersects with the "medium" morbidity rate at 20%, the center of the "medium" morbidity rate is located at 40%, the "medium" morbidity rate intersects with the "high" morbidity rate at 65%, and the "high" morbidity rate covers the interval from 50% to 100%. In (b), the growth rate, the input variable of the membership function of the growth rate, is divided into three fuzzy sets: "slow," "normal," and "fast" using a trapezoidal membership function. The "slow" growth rate covers the interval from 0 to 10%. The range is 0.0 g / day, "normal" growth covers 5.0 to 20.0 g / day, and "fast" growth covers 15.0 to 25.0 g / day. This design ensures that the controller can effectively respond to the feeding demand under different growth rates, and the overlap between the sets ensures the smoothness of the control output. In (c), the feeding level membership function represents the percentage of daily feed amount to the average weight of a single fish. To obtain a smoother and more accurate control output, five fuzzy sets were designed for the output variables—"ultra-low", "low", "medium", "high", and "ultra-high"—and Gaussian membership functions were adopted. Their center points are located at 0.5%, 1.5%, 2.5%, 3.5%, and 4.5%, respectively. Gaussian functions can generate continuous and differentiable control surfaces, which is beneficial for optimizing control performance and achieving fine adjustment of feed amount. S6 specifically includes: using the initial recommended feeding level output by the fuzzy controller as a reference trajectory, constructing a multi-objective optimization function that includes a growth deviation objective function, a disease deviation objective function, a control quantity deviation penalty, and a smoothing control term, to achieve multi-objective collaborative optimization, balance growth promotion, disease inhibition, and operational stability, using the MTL-LSTM-SAT multi-task learning model as an internal prediction model to predict changes in growth rate and disease rate under different feeding strategies within the next 7 days, and embedding a physical constraint that the feeding amount is non-negative and does not exceed 5% of the fish's body weight to ensure the feasibility and safety of control commands and avoid the generation of invalid or harmful commands, solving a finite-time domain optimization problem at each decision moment, calculating a series of future feeding strategy sequences, and only using the optimal feeding amount at the current moment as the final control command output to achieve rolling optimization and real-time feedback, ensuring that the control decision dynamically adapts to environmental changes; The specific work involves the following steps: In the optimization phase of the MPC controller, the initial recommended feeding level output by the fuzzy controller is used as a reference trajectory. This is combined with expert experience based on the current state to provide guidance for higher-level optimization that aligns with aquaculture management intuition. Based on this, the MPC controller constructs a multi-objective optimization function that comprehensively considers multiple aquaculture objectives. This function comprises four core parts: a growth deviation objective function, used to quantify the gap between the predicted growth rate and the expected growth target; a disease deviation objective function, aimed at minimizing the predicted disease rate and reducing disease risk; a control deviation penalty term, used to measure the degree of deviation between the actual feeding level and the fuzzy recommendation; and a smoothing control term, which ensures stable equipment operation and reduces fish stress by limiting drastic fluctuations in feeding amounts between adjacent time points. Each objective is dynamically balanced through weight coefficients, forming an optimization objective system that considers both short-term response and long-term benefits. The trained MTL-LSTM-SAT multi-task learning model is embedded into the MPC framework as an internal prediction model, based on current water quality, fish status, and historical feeding data. Data is used to dynamically predict changes in fish growth rate and disease trends under different feeding strategies over the next 7 days. During the optimization process, physical constraints on feeding amount are embedded, explicitly limiting the feeding level to non-negative values ​​and ensuring that its upper limit does not exceed 5% of the average fish mass. This constraint ensures the feasibility and safety of control commands from both mechanical execution capability and fish physiological load perspectives, avoiding the generation of invalid or harmful control commands, thereby ensuring the reliable operation of the system. At each predetermined decision moment, the MPC controller solves the constrained finite-domain optimization problem based on the latest system state and prediction information. Through numerical optimization algorithms, a series of optimal feeding strategy sequences for a future period are calculated. Only the optimal feeding amount corresponding to the current moment in the sequence is used as the final control command and immediately output to the execution layer. The remaining future strategies are only used for internal rolling optimization calculations and are not directly executed. This rolling optimization and feedback correction mechanism enables the system to continuously update decisions based on the latest perception and prediction information, achieving dynamic and precise closed-loop control of the feeding process. It should be noted that the MPC controller receives the initial recommended feeding level from the lower-level fuzzy controller at each decision time. Based on this, a constrained finite-time optimization problem is solved to calculate a feeding strategy that combines immediate benefits with long-term optimality; the initial recommended feeding level is... As a reference target, one of the optimization objectives of MPC is to make the final control sequence as close as possible to this reference value. If the fuzzy suggestion is reasonable in the long run, MPC will adopt it; if the suggestion is based on the possibility of long-term negative effects from short-term states (such as disease induced by overfeeding), MPC will make forward-looking corrections and fine adjustments based on model predictions, according to the control deviation penalty introduced in the multi-objective optimization function. This quantifies how far the current decision deviates from the ideal target; a larger value indicates a worse predictive control effect. The calculation process is as follows: ; in: Penalty for deviation in control quantity; The sum of growth deviation penalty, disease deviation penalty, and smoothness control; Weighting coefficients for fuzzy output reference values; For the current moment, To predict the step size (day); The total number of future steps (the total number of future steps is set to 7). To plan future feeding levels based on the current day's feeding levels; The solution sought by the rolling optimization mechanism, the process of which is as follows: Figure 5 As shown; S6 also includes: setting initial values ​​for growth deviation weight coefficient, disease deviation weight coefficient, fuzzy output reference value weight coefficient, and smoothing control term weight coefficient; dynamically adjusting each weight coefficient according to real-time predicted status to achieve multi-objective adaptive optimization, improve system dynamic response and long-term stability; increasing the growth deviation weight coefficient when growth rate decreases, increasing the disease deviation weight when disease rate increases, and increasing the smoothing control weight when equipment control quantity fluctuates too much; ensuring that the control strategy focuses on the main contradictions, maintaining a dynamic balance between growth and health; and dynamically balancing the control strategy through the weight adaptive mechanism, enabling the feeding control system to achieve long-term optimal decision-making in different breeding stages and fish population states, enhancing the system's environmental adaptability, and achieving optimal feeding management throughout the entire cycle. The specific work involves: During the system initialization phase, setting the initial values ​​of the growth deviation weight coefficient, disease deviation weight coefficient, fuzzy output reference value weight coefficient, and smoothing control term weight coefficient to a unified benchmark value to establish a balanced multi-objective optimization foundation. After system startup, based on real-time acquired fish status prediction data and equipment operating parameters, dynamically adjusting each weight coefficient. When the predicted growth rate output by the MTL-LSTM-SAT multi-task learning model is lower than a set threshold, automatically increasing the growth deviation weight coefficient to strengthen the control strategy's inclination towards the growth-promoting objective; when the predicted disease rate exceeds the warning range, correspondingly increasing the disease deviation weight coefficient to enhance the system's response to suppressing health risks; if the fluctuation range of feeding instructions in adjacent cycles is detected to exceed the allowable range, increasing the smoothing control term weight coefficient to suppress drastic changes in feeding amount and ensure the feeding control system's sensitivity to each objective under different operating conditions. Adjustable properties provide flexible and targeted adjustment methods for rolling optimization; the weight adaptive adjustment mechanism continuously monitors and predicts growth rate, disease rate, and control quantity fluctuations, and real-time corrects the relative importance of each objective to achieve dynamic trade-offs in control strategies. In the stage of slow fish growth, the growth deviation weight coefficient is increased to make the optimization process focus more on improving feeding levels to promote growth; in the stage of increased disease risk, the disease deviation weight coefficient is increased to guide the MPC controller to adopt a conservative feeding strategy to reduce disease pressure; when the feeding equipment is unstable or water quality parameters fluctuate frequently, the smoothing control term weight coefficient is increased to force the optimization results to maintain the continuity of feeding instructions, avoiding fish stress or equipment damage caused by sudden changes in control quantities. This allows the MPC controller to autonomously adjust the competitive relationship between each objective in different breeding stages and fish conditions, achieving a balance between breeding benefits and risk control in long-term optimization. It should be noted that the growth deviation objective function To predict the difference between the target weight and the ideal weight, the intervention of this parameter requires predicting the weight. As close as possible to the target weight Its formula is as follows: ; In the formula: This is the growth deviation weighting coefficient; Disease bias objective function The goal is to have a prevalence rate approaching 0. The higher the prevalence rate D, the greater the bias, as shown in the following formula: ; In the formula, This is the disease bias weighting coefficient; To avoid frequent and significant starts and stops of the feeder motor, extend the equipment's lifespan, and reduce fish stress caused by drastic fluctuations in water quality (such as pH and ammonia nitrogen) due to large changes in feed volume, a smoothing control term is introduced into the control algorithm. Its expression is as follows: ; In the formula: For smoothing control term weighting coefficients; MPC control performance depends on four weighting coefficients. , , , The initial value setting is 1. When the input data becomes unstable, the corresponding weight coefficient will increase, for example, when the growth rate slows down. The scale will increase, and the entire system will focus on fish growth, increasing the amount of feed; when the proportion of disease increases, As the morbidity rate increases, the entire system should focus on reducing the incidence rate and appropriately reduce the amount of feed. When the operating time of the feeding equipment differs significantly between adjacent periods... Increasing the size of the controller will result in very smooth and gentle control commands, but may sacrifice some dynamic response speed. This is for manual adjustment. Increasing the feed amount respects the fuzzy controller's decision-making, resulting in an output very close to the expert rule base. MPC primarily plays a "fine-tuning" and "smoothing" role. Conversely, decreasing the feed amount yields a significantly different result from the fuzzy suggestion, where MPC dominates the decision-making process, pursuing long-term optimality. This ensures that any control quantity obtained through optimization is physically executable (the feed amount is non-negative and does not exceed 5% of the average fish body weight), fundamentally eliminating invalid control commands at the algorithmic level, thus guaranteeing the system's safety and reliability. The control method incorporates constraint handling, as shown in the following formula: ; S7 specifically includes: converting the optimal feeding amount in the feeding strategy sequence output by the MPC controller into a pulse width modulation signal, driving the servo motor of the feeder and the screw feeder to perform precise feeding, accurately mapping the optimization decision into physical control actions, realizing the quantitative and timed control of feed delivery, after feeding is completed, re-collecting fish images and water quality data, updating the input of the MTL-LSTM-SAT multi-task learning model, starting a new round of state prediction and control decision, ensuring that the state prediction is synchronized with the aquaculture environment through real-time data updates, maintaining the timeliness and accuracy of decision-making, and forming a closed-loop adaptive feeding control system through continuous multi-cycle data acquisition, prediction optimization and execution feedback, realizing the dynamic adjustment and long-term performance optimization of the feeding strategy, establishing a self-adjusting cycle mechanism, and enabling the system to have continuous adaptability and long-term benefit optimization potential; The specific work involves the following: In the feeding control system, after the Model Predictive Control (MPC) controller periodically calculates the optimal feeding amount, it converts the control command into specific physical actions. The system quantifies the optimal feeding amount and maps it to corresponding pulse width modulation signal parameters based on the mechanical characteristics of the feeding equipment and the control interface specifications. This signal is output through an industrial-grade I / O module to precisely control the speed of the feeder's servo motor and the operating time of the screw feeder, achieving quantitative and timed feed delivery. The entire command issuance process employs strict timing logic and a safety verification mechanism to ensure that the control command is executed within the specified time frame. The reliability and real-time performance of the transmission and execution links are ensured through high-precision signal conversion and device driving methods, which translate optimized decisions into specific feeding actions without distortion, achieving precise control of the feeding process at the physical level. After the feeding operation is completed, a new round of multi-source data acquisition is immediately initiated. The underwater binocular camera continues to capture images of fish activity at a preset frequency, and the water quality sensor continuously monitors key parameters such as water temperature, pH value, dissolved oxygen, and ammonia nitrogen. The newly acquired sensing data is preprocessed and then time-aligned and feature-fused with the current feeding record. The updated dataset is then input into MTL-LSTM-SAT in real time. In the multi-task learning model, the latest historical sequence is used as input. The model performs forward calculations based on the new time-series data and continuously outputs updated predicted values ​​of fish growth rate and disease rate for the next 7 days, providing a basis for state prediction for the next control cycle. Through real-time data acquisition and model input update mechanisms, the dynamic changes in the aquaculture environment are continuously tracked to ensure that state predictions are always based on the latest observation information. By continuously looping the workflow of data acquisition-state prediction-decision optimization-instruction execution-feedback update, a complete closed-loop adaptive feeding control system is formed. Each control cycle takes the execution results of the previous stage and the latest environmental state as input. The hierarchical control architecture achieves synergy between short-term response and long-term optimization. The fuzzy controller ensures the system's real-time intervention capability for emergencies. The MPC controller dynamically balances multiple objectives such as growth promotion, disease suppression, and operation smoothing in rolling optimization. During continuous operation, the feeding strategy is continuously adjusted through weight adaptive mechanism and model prediction update to gradually approach the optimal long-term comprehensive benefits. Based on the closed-loop control structure of multi-cycle iterative feedback, the system has the ability to continuously learn and adaptively optimize, realizing intelligent management and long-term performance optimization of the entire feeding process.

[0025] Example 3, as Figures 1 to 11As shown, based on Examples 1-2, this invention, in order to verify the superiority of the proposed algorithm from a theoretical perspective and the role of related improvement measures in performance enhancement, constructs a fish growth-disease simulation system based on nonlinear dynamic equations to simulate key processes in a real aquaculture environment. This simulation model comprehensively considers the diminishing returns of feed input, the coupling relationship between disease and growth, and random environmental disturbances, allowing all common control algorithms to run in a virtual "fishpond" with identical indicators. Simulation environment settings: Using red tilapia as a template, the stocking density was 100 fish / m³, with a total number of 1000 fish. The initial weight of each fish was 500g. A basic suspended feed (32% soybean meal, 33% wheat bran, 16% corn, 4.7% fish meal, 1% bone meal, and 0.3% salt) was used. The simulation period was 60 days, from September 1st to October 30th (water quality data during this period was the average daily change). The feeding effect followed the growth and disease characteristics of red tilapia. Initially, increasing the feeding level had a significant effect, but beyond a certain level, excessive feeding had a diminishing effect on fish growth. Insufficient or excessive feeding would lead to decreased fish health and exacerbated disease. Reasonable feeding could inhibit the rapid growth of disease. The growth rate was constrained to 0-20g / day, and the disease rate was constrained to 0.5%-80%. This setting conforms to the classic biological model. The model selection included five feeding strategies: PID (Proportional-Integral-Derivative Control), LQR (Linear Quadratic Regulator), Fuzzy Control, MPC, and the fuzzy-MPC precise control of this invention. The control rules and membership degrees of the fuzzy control algorithm were the same as those of the fuzzy-MPC precise control algorithm of this invention, and the weight coefficients of the MPC algorithm were the same as those of the fuzzy-MPC precise control algorithm of this study. The evaluation indicators were average growth rate (g / day), disease rate, and feeding level (average single tail body mass percentage). The final evaluation indicator was feed conversion efficiency (average growth rate / feeding level). This indicator is different from the feed utilization rate (FCR) in actual breeding. This concept is a relative indicator defined for horizontal comparison of algorithm performance in simulation, that is, how many grams of growth can be generated by feeding 1% of body mass of feed per day. The higher the value, the higher the efficiency. Simulation results and analysis: Depend on Figure 6It can be seen that the PID control feeding strategy is relatively aggressive and coarse when the fish are at a normal growth rate and a low disease rate. The simulation used the highest feeding level (5% of the average single-tail weight) for 60 days, as this level allows the fish to maintain a good growth rate and a low disease rate. Conversely, the LQR control feeding strategy is more conservative. Since the initial disease rate in the simulation was 15%, this strategy uses a low feeding level (0.5% of the average single-tail weight) to control the disease rate. This low feeding level leads to increased disease due to insufficient food, and the situation escalates. The feeding strategy employed a low feeding level (0.5% of average single-tail weight) for 60 consecutive days; the fuzzy control feeding strategy selected different feeding levels (2% or 3% of average single-tail weight) based on fish growth and disease status; the MPC control feeding strategy made fine adjustments based on daily water quality, fish growth, and disease status to maintain a reasonable range; while the fuzzy-MPC precision control feeding strategy was more stable, and the overall feeding level was slightly lower than that of the single MPC control feeding strategy, proving that this feeding strategy can reduce a certain amount of feed waste. As shown in Figure 7, the PID control feeding strategy has the fastest growth rate of fish due to its aggressive feeding level; the fuzzy-MPC precise control feeding strategy has the second fastest growth rate after the PID control feeding strategy, slightly higher than the MPC control feeding strategy, and significantly higher than the fuzzy control feeding strategy; the LQR control feeding strategy has a significantly slower growth rate of fish than other feeding strategies due to its conservative feeding level. Depend on Figure 8 It can be seen that the LQR-controlled feeding strategy, due to insufficient feed to meet the normal growth needs of the fish, led to increasingly severe disease, to the point that the disease rate reached a high level after 60 days of simulation, and in actual production, it had entered a period of collective outbreak of fish diseases. The fuzzy control feeding strategy had the second highest disease rate. As can be seen from the curve, the fuzzy control feeding strategy could not control the disease rate, but there was no trend of an outbreak. The PID-controlled feeding strategy showed a significant decrease in disease rate on day 18 and continued until day 50, after which the disease rate rose slightly again, indicating that its feeding strategy was ineffective. The feeding strategy with MPC control consistently controlled the morbidity rate. Among the five strategies, the morbidity rate was at a moderate level, and it was able to contain the morbidity after day 35, indicating that it could basically control the morbidity rate at a low feeding level. The feeding strategy with fuzzy-MPC precision control did not have a stronger ability to contain the morbidity rate in the early stage of the simulation than the feeding strategy with PID control, but the curve showed an overall downward trend. Especially after day 35, its morbidity rate was the lowest among the five feeding strategies, with a final morbidity rate as low as 5.16%, lower than PID's 6.25%. comprehensive Figure 6 , 78. The feeding strategy controlled by LQR has no feasibility in selecting the feeding level, and this control method is not suitable for intelligent intensive fish farming; the feeding strategy controlled by PID uses the maximum amount of feed within a reasonable range to ensure the normal growth of farmed fish; the feeding strategy controlled by fuzzy logic can select different feeding levels more reasonably, but it cannot precisely control the amount of feed each time; the feeding strategy controlled by MPC can take into account the amount of feed, the growth and health of the fish; while the feeding strategy controlled by fuzzy-MPC precision can not only select the feeding level reasonably, but also combine the feeding level with the growth and health of the fish, and has long-term planning, and can maintain the best growth and health of the fish with a lower amount of feed. Table 3. Average performance results of the five control methods: .

[0026] To comprehensively evaluate the performance of the fuzzy-MPC control-based precision feeding system proposed in this invention, a 60-day simulation comparison test was conducted. The daily average values ​​of various performance indicators are shown in Table 3. The fuzzy-MPC precision control feeding strategy exhibited a good growth-promoting effect, with an average growth rate of 12.05 g / day, second only to PID control (13.77 g / day). (g / day); However, the high growth rate of PID control is achieved at the cost of extremely high feed intake (5% of average single tail mass), while the intelligent control system achieves similar growth results using only 3.08% of the feed intake. In terms of disease control, the fuzzy-MPC precision control feeding strategy performs best, with an average disease rate of only 8.29%, significantly lower than other control methods. Fuzzy control and LQR control have poor disease control effects, with average disease rates of 19.12% and 31.52%, respectively. In summary, the fuzzy-MPC precision feeding strategy achieves the best balance between growth rate, disease control, and feed conversion efficiency, making it the optimal choice among the five control methods.

[0027] Example 4, as Figures 1 to 11 As shown, based on Examples 1-3, the present invention also conducted performance tests. These performance tests were conducted to verify the effect of deploying the fuzzy-MPC precise control feeding strategy to the fish production end, and to comprehensively analyze its advantages in increasing yield, saving feed, reducing disease and stabilizing water quality. Experimental Design: The experiment was conducted from August 1st to September 29th, 2025 at the Modern Aquaculture Base in Xian'an District Agricultural High-tech Zone, Xianning City, Hubei Province. The aquaculture mode was intensive enclosed tank culture. The tanks were 4m in diameter and 2m deep, corresponding to an effective water volume of approximately 25m³. The stocking density was 2500 fish / tank. The cultured species was red tilapia (rainbow seabream), and the basic suspended feed was selected (32% soybean meal, 33% wheat bran, 16% corn, 4.7% fish meal, 1% bone meal, and 0.3% salt). The average body weight of the fish was (421±50) g, and the average body length was 25-30cm (from the snout to the end of the caudal fin). The fish were randomly divided into three groups. Each tank was equipped with an independent aeration (bottom aeration disc), water circulation, and filtration system. Throughout the experiment, the initial culture water was freshwater that had undergone sedimentation and filtration treatment. Natural light cycle was used, and water quality data was collected 6 times / hour, i.e., once every 10 minutes. The experimental control group was divided into 3 groups, and the 3 groups were configured as follows: (1) Artificial feeding group (Group A): Feeding decisions are made by aquaculture technicians with 20 years of experience. The feeding time and amount are determined based on the weather conditions, water temperature changes and observation of the fish feeding behavior, and the actual amount of feed is recorded in detail every day. (2) Automatic feeding group (Group B): A conventional automatic feeder was used, with fixed feeding times (06:00, 10:00, 14:00, 18:00) and fixed feeding amounts (based on computer images to assess the body weight of a single fish, with a feeding rate of 3.5% of the body weight). (3) Intelligent feeding group (Group C): Deployed and developed intelligent feeding system. The system collects fish images in real time through underwater cameras, assesses their growth and disease, and combines the data collected by multi-parameter water quality sensors (monitoring dissolved oxygen, temperature, pH and ammonia nitrogen) with the data input into the MTL-LSTM-SAT model to predict the fish's growth and health status in the next 7 days. Finally, the fuzzy + MPC controller decides the optimal feeding strategy for each day. Apart from the different feeding strategies, the three groups of test tanks were completely identical in terms of aeration equipment start-up and shutdown, water quality control, disease prevention and control, and feed type usage to ensure the reliability of the test results. Evaluation indicators: Feed conversion ratio (FCR): FCR = total feed intake / total fish weight gain; this indicator is the core indicator for evaluating feed utilization efficiency. The lower the FCR, the higher the feed conversion efficiency, which means that the same amount of feed can be converted into more fish weight.

[0028] Specific growth rate (SGR): SGR = InW2 - InW1 / t × 100%, where W1 and W2 are the average body mass (g) of a single fish at the beginning and end of the experiment, respectively, and t is the number of days in the experiment. This index reflects the growth rate of the fish. Survival rate (SR): SR = (final number of fish / initial number of fish) × 100%. Dead fish are removed to calculate the number of survivors, which directly reflects the impact of feeding strategies on the health of the fish population. Water quality stability: Dissolved oxygen (DO), ammonia nitrogen (TAN), and pH value were evaluated. The percentage of time during the experiment with dissolved oxygen below 4 mg / L, ammonia nitrogen above 0.5 mg / L, and pH below 6.5 and above 8.5 were calculated. The impact of these three evaluation indicators on the aquatic environment was assessed. The selection of these three indicators was based on the following reasons: 1) Oxygen is consumed by the decomposition of feed residue and excrement in the water, leading to a decrease in dissolved oxygen content in the aquaculture water; 2) Excessive protein feed leads to an increase in ammonia nitrogen concentration due to increased fish excrement and decomposition products of leftover feed; 3) Excessive organic matter decomposition produces acidic substances, or phytoplankton proliferates, causing a rise in pH during daytime due to photosynthesis and a drop in pH at night due to respiration, resulting in increased pH fluctuations in the aquaculture water. In summary, these three points indicate that abnormal changes in water quality can negatively impact fish growth and health. Experimental data and analysis: like Figure 9 As shown, under the three feeding strategies, the feed conversion ratio (FCR), specific growth rate (SGR), and survival rate (SR) were all optimal. Regarding water quality stability, the following three factors were lowest: ① the percentage of time spent with dissolved oxygen below 4 mg / L; ② the percentage of time spent with ammonia nitrogen above 0.5 mg / L; and ③ the percentage of time spent with pH below 6.5 and above 8.5. This demonstrates that the feeding strategy under the fuzzy + MPC controller has a stronger ability to stabilize aquaculture water compared to traditional manual experience and timed automatic feeding, which is also reflected in its higher survival rate; the FCR of group C was as low as 1.2. 3. Compared with groups A and B, the feed waste was reduced by 18.5% and 30.1% respectively, indicating that the method in this paper can significantly reduce feed waste. Group C had the highest SGR (1.11% / d), and the specific growth rate of the fish was increased by 7.77% and 15.63% compared with groups A and B respectively. The survival rate of group C reached 94.21%, which was also significantly higher than the other two groups. After the end of the breeding experiment, the average body weight of the fish in group A reached 815±50g, which was the highest among the three groups. This shows that the feeding strategy under the fuzzy-MPC controller decision can enable farmed fish (red tilapia) to reach market size faster and effectively shorten the entire breeding cycle. To further elucidate the decision-making mechanism of the fuzzy-MPC control feeding strategy, such as Figure 10The example shown illustrates a specific decision-making case from 7:00 AM on September 11th to 7:00 PM on September 12th (early autumn with large diurnal temperature variations and short-term showers). From 8:00 AM to 9:00 AM on the 11th, with rising water temperature and showers, the system detected a decreasing trend in dissolved oxygen (from 5.2 mg / L to 4.6 mg / L). The MTL-LSTM-SAT model predicted that further decreases in dissolved oxygen would negatively impact fish growth and health. The fuzzy controller initially output a "moderate" feeding adjustment factor based on the current growth rate being "normal" and the disease rate being "moderate." Based on this, the MPC optimizer decided to delay the feeding time and reduce the amount of feed by 20% to avoid the risk of low dissolved oxygen in the future. After that, the dissolved oxygen level increased significantly. This decision effectively avoided the waste of feed under low dissolved oxygen conditions, and at the same time reduced the oxygen consumption of feed residue, thus avoiding low dissolved oxygen conditions. On the other hand, in the artificial experience group (Group A), the technicians still fed according to experience and did not realize that the fish's appetite would decrease under low dissolved oxygen conditions, resulting in a lot of feed left over on the day. On the afternoon of the 11th, the water temperature rose sharply, the oxygen consumption of feed residue increased, and the dissolved oxygen was too low (3.4 mg / L). Economic benefit analysis: The costs of intensive aquaculture mainly consist of seedlings, feed, medicine, labor, water and electricity, and equipment depreciation, with feed costs accounting for the highest proportion, typically reaching 60%–70%. This study statistically analyzed the aquaculture benefits under three feeding strategies using a single tank of 2500 fish fry over a 60-day experimental period. The costs and benefits are as follows: Figure 11 As shown in the figure; the results show that group C, which adopts the fuzzy-MPC precision feeding strategy, has significantly lower feed costs than group A (human experience group) and group B (timed feeding group), and the net profit per unit area is increased by about 28.5% and 41.6% respectively, demonstrating a clear economic advantage; in conclusion: Based on previous research (fish disease, growth detection, and MTL-LSTM-SAT prediction model), this invention uses the MTL-LSTM-SAT model to predict fish growth and disease status. It also combines fuzzy control and MPC to formulate a precise feeding strategy based on fuzzy-MPC control. In this strategy, fuzzy control is used as the lower-level control and MPC control is used as the upper-level control. A fuzzy control rule base is established, and an initial membership degree is set. The fuzzy control outputs an initial suggested value, which is then processed by the MPC algorithm for multi-objective trade-offs and constraint calculations to obtain a precise feed amount output. Simulation data over 90 days showed that the feeding strategy constructed in this invention had the highest feed conversion rate of 3.91 compared to common control methods (PID, fuzzy control, MPC, LQR). At the end of the simulation, the final morbidity rate was 5.16%, and the morbidity rate continued to decrease, which was the lowest among the five feeding strategies. It achieved the optimal balance in terms of improving growth rate, controlling disease, and saving feed. The results of a 60-day group culture trial showed that Group C, which adopted the fuzzy-MPC precision feeding strategy, significantly outperformed the artificially fed group (Group A) and the automatic feeding group (Group B) in all aspects. Specifically, Group C's feed conversion ratio (FCR) was as low as 1.23, a reduction of 18.5% and 30.1% compared to Groups A and B, respectively, indicating that this method can significantly reduce feed waste. Group C also had the highest specific growth rate (SGR) at 1.11% / day, which was 7.77% and 15.63% higher than that of Groups A and B, respectively. The survival rate of Group C reached 94.21%, also significantly higher than the other two groups.

[0029] Example 5, as Figure 12 As shown, based on Examples 1-4, this invention also provides a design and implementation scheme for a feeding execution system. In order to achieve accurate execution of the feeding instructions output by the fuzzy-MPC controller, a feeding machine motor control system based on an STM32 microcontroller is designed. This system converts the target feeding amount generated by the upper-level control strategy into a motor drive signal, and realizes quantitative and timed feeding through closed-loop control.

[0030] System Hardware Architecture: The feeding execution system consists of a host decision-making unit, a motor controller, a drive module, and an actuator. Its hardware architecture is as follows: Figure 12 As shown, the AC-DC conversion module provides power, and the fuzzy-MPC controller outputs one feed amount data to the STM32F103ZET6 main controller. The main controller uses the calculation input drive module to control the stepper motor speed. The stepper motor replaces the original brushless DC motor in the feeder, driving the screw feeder to convert the rotational motion into feed conveying. An incremental encoder (1000 lines) is installed on the motor shaft to monitor the speed in real time and provide speed feedback to the main controller. The communication module uses an RS485 bus to realize reliable data transmission between the main controller and the host computer (running the fuzzy-MPC algorithm).

[0031] The system uses an STM32F103ZET6 as the main controller, whose 72 MHz main frequency and rich peripheral resources meet the real-time control requirements; the motor is a 42-step motor (compatible with DC geared motor), which converts the rotational motion into feed conveying through a screw feeder; an incremental encoder (1000 lines) is installed on the motor shaft end to monitor the speed in real time and provide speed feedback; the communication module uses an RS485 bus to realize reliable data transmission with the host computer (running the fuzzy-MPC algorithm); the key hardware parameters are shown in Table 4.

[0032] Table 4 Hardware parameters of the feeder control system: .

[0033] Feeding rate - motor speed conversion model: conveying capacity of screw feeder (g) and motor speed (rmp) and time The relationship between (s) is described as follows: ; in, The displacement per revolution of the screw feeder (calibrated according to the feeder's instruction manual and on-site testing). This refers to the conveying efficiency coefficient. Feeding level based on the output of the fuzzy-MPC controller It can calculate the target feeding The calculation method is as follows: ; In the formula, The number of fish raised in the bucket; This represents the average body weight of a single fish. Based on the output target feeding amount The required average motor speed can be calculated. ,as follows: ; In the formula, The preset feeding time is 45 seconds.

[0034] Communication Protocol and System Integration: The host computer (fuzzy-MPC controller) and the motor controller use a custom RS485 communication protocol. The STM32 receives the feeding instruction frame via RS485, extracts the target feeding amount after verification, determines the speed according to the conversion model, and then outputs a PWM signal to the motor driver through timer TIM1. The encoder collects the speed in real time and feeds back the current feeding amount and motor status to the host computer via RS485 to ensure that the feeder can operate normally. The microcontroller data frame format is shown in Table 5.

[0035] Table 5 Feeding instruction frame format: .

[0036] Feeding performance testing and verification: The accuracy of the feeder control system was tested under ideal conditions, and the results are shown in Table 6. Each feeding rate was tested 10 times. The feed type was standard 3-8mm diameter semi-suspended extruded feed, which was dry, hard, and not easily broken.

[0037] Table 6. Feeding accuracy test results: .

[0038] Test data shows that the improved feeder has a relative error of less than 0.3% within the feeding range of 5000-50000 g, meeting the requirements for precise feeding.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for precise feeding control in intensive aquaculture based on fuzzy-MPC control, characterized in that, Includes the following steps: S1. Employ a sensing system to collect multi-source sensing data in real time, including images of fish activity, health status, and water quality data, to form a comprehensive status dataset. S2. Based on the collected multi-source sensing data, construct an MTL-LSTM-SAT multi-task learning model, integrate soft attention mechanism, and predict the growth rate and disease rate of fish in the next 7 days. S3. Design a fuzzy control layer using a fuzzy control algorithm and formulate a fuzzy control rule table with morbidity and growth rate as inputs. S4. Using a hierarchical control strategy, the lower-level fuzzy controller and the upper-level MPC controller are divided into a lower-level fuzzy controller and an upper-level MPC controller, which respectively receive the short-term state output by the model and the long-term prediction value of the MTL-LSTM-SAT multi-task learning model. S5. The fuzzy controller outputs a preliminary recommended feeding level based on real-time predicted status through fuzzy inference. The S6 and MPC controllers use the initial recommended feeding level output by the fuzzy controller as the reference trajectory, and combine the long-term predicted value output by the MTL-LSTM-SAT multi-task learning model to perform constrained 7-day rolling optimization with the goals of maximizing growth and minimizing disease, and calculate the feeding strategy sequence that is optimal in both the immediate and long-term. S7. Convert the control strategy in the feeding strategy sequence into a feeding instruction and send it to the feeding execution mechanism to complete the precise feeding.

2. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 1, characterized in that: S1 specifically includes: A sensing system, including an underwater camera and a water quality sensor, is deployed in the captive tank. The underwater camera takes side shots at different depths along the tank wall, while the water quality sensor is inserted into the water at a fixed depth along the tank wall, simultaneously acquiring images of fish activity and water quality parameters. The underwater camera collects images of fish underwater activity from 6:00 to 18:00 at a sampling frequency of 0.1 frames per second. The water quality sensor collects data on water temperature, pH value, dissolved oxygen and ammonia nitrogen every 10 minutes and records the daily feeding amount. The collected fish activity image data is input into a visual recognition algorithm to identify the fish's health status and key growth points. The data is then fused with water quality data and feeding data to construct a comprehensive status dataset containing time-series features.

3. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 1, characterized in that: S2 specifically includes: By combining a comprehensive state dataset, an LSTM network framework based on multi-task learning is constructed. The input is three-dimensional time series data containing water quality, feeding amount and fish status. The LSTM layer extracts temporal features and outputs hidden states. A soft attention mechanism is introduced to weight and focus the hidden state of the LSTM output, and the attention weights at each time step are calculated. The context vector is generated by weighted summation and then input into the shared feature layer. At the same time, the predicted results of fish growth rate and disease rate for the next 7 days are output.

4. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 3, characterized in that: S2 further includes: The model was trained using water quality, feeding, and fish growth and health data from historical aquaculture cycles. Mean squared error and cross-entropy were used as multi-task loss functions, and the network parameters were optimized through backpropagation. The model performance was evaluated on the validation set, and the number of LSTM units, attention heads, and fully connected layer structure were adjusted to achieve the optimal balance between growth rate prediction and disease proportion prediction tasks. The trained MTL-LSTM-SAT multi-task learning model is embedded into the feeding control system to receive multi-source sensing data in real time and output the growth and disease prediction sequence for the next 7 days in a rolling manner.

5. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 1, characterized in that: S3 specifically includes: A fuzzy control layer was designed using a fuzzy control algorithm, with morbidity and growth rate as input variables and feeding level as output variables. The input variables were divided into low / medium / high and slow / normal / fast fuzzy sets, and the output variables were divided into five levels: ultra-low / low / medium / high / ultra-high. Based on the knowledge of aquaculture experts, a fuzzy control rule table containing nine rules was formulated, and triangular, trapezoidal and Gaussian membership functions were used to describe the fuzziness of the input and output variables respectively.

6. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 5, characterized in that: S4 specifically includes: The feeding control system is divided into a lower-level fuzzy controller and an upper-level MPC controller. The lower level responds to real-time prediction of the state in a high-frequency manner, while the upper level performs long-term rolling optimization in a low-frequency manner. The fuzzy controller receives the short-term state output by the MTL-LSTM-SAT multi-task learning model, which covers the current prevalence and growth rate. The MPC controller receives the output of the fuzzy controller and the 7-day prediction sequence output by the MTL-LSTM-SAT multi-task learning model.

7. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 6, characterized in that: S5 specifically includes: Based on the predicted prevalence and growth rate values ​​at the current moment output by the MTL-LSTM-SAT multi-task learning model, the activation intensity of each fuzzy rule is calculated through the membership function. Based on the fuzzy control rule table, reasoning is performed to obtain the feed level output value corresponding to each rule, and the weighted average method is used for defuzzification calculation. Output a primary recommended feeding level, which is used as a reference input for the MPC controller and can be directly used to quickly adjust feeding behavior in the event of sudden changes in water quality. The formula for calculating the primary recommended feeding level is as follows: ； In the formula: This is the initial recommended feeding level; For the nonlinear mapping function of the fuzzy controller; , These represent the disease rate and average growth rate of the fish population, respectively. The total number of rules; For the first Fuzzy rules in input The activation intensity below; For the first The output feeding level corresponding to each fuzzy rule.

8. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 7, characterized in that: S6 specifically includes: Using the initial recommended feeding level output by the fuzzy controller as the reference trajectory, a multi-objective optimization function is constructed, which includes a growth deviation objective function, a disease deviation objective function, a control quantity deviation penalty, and a smoothing control term. The MTL-LSTM-SAT multi-task learning model was used as an internal prediction model to predict changes in growth rate and morbidity under different feeding strategies in the next 7 days, and the physical constraint that the feeding amount is non-negative and does not exceed 5% of the fish body weight was embedded. At each decision point, a finite-time optimization problem is solved to calculate a series of future feeding strategies. Only the optimal feeding amount at the current moment is output as the final control command.

9. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 8, characterized in that: S6 further includes: Set initial values ​​for the growth deviation weight coefficient, disease deviation weight coefficient, fuzzy output reference value weight coefficient, and smoothing control term weight coefficient, and dynamically adjust each weight coefficient according to the real-time prediction status. Increase the growth deviation weight coefficient when the growth rate decreases, increase the disease deviation weight when the morbidity rate increases, and increase the smoothing control weight when the equipment control quantity fluctuates too much. By using a weighted adaptive mechanism to dynamically balance control strategies, the feeding control system can achieve long-term optimal decision-making under different breeding stages and fish population conditions.

10. The method for precise feeding control in intensive aquaculture based on fuzzy-MPC control according to claim 1, characterized in that: Specifically, S7 includes: The optimal feeding amount in the feeding strategy sequence output by the MPC controller is converted into a pulse width modulation signal to drive the servo motor of the feeder and the screw feeder to perform precise feeding. After feeding, fish images and water quality data are re-acquired, the input of the MTL-LSTM-SAT multi-task learning model is updated, and a new round of state prediction and control decision-making is initiated. Through continuous multi-cycle data acquisition, prediction optimization, and execution feedback, a closed-loop adaptive feeding control system is formed.

Images