A duck house intelligent ventilation control method and system based on dynamic regulation effect graph

Abstract

The application discloses a kind of duck house intelligent ventilation control method and system based on dynamic regulation effect diagram, belong to wisdom breeding and livestock and poultry house environment regulation technology field.The application obtains breeding knowledge unit and historical operation data, constructs the dynamic regulation effect diagram including scene node, state node, operation node and response node, the regulation effect of single operation is represented by directed weighted edge, the synergies, cancellation and timing effect between multiple operations are represented by high-order interaction relationship;Baseline environmental state under no new operation is predicted in real-time operation, candidate combination control strategy is generated, the incremental effect of each strategy is predicted using effect diagram and the target strategy is determined by rolling optimization under safety constraints;After execution, update effect diagram parameters according to the deviation between actual response and predicted response.The application realizes the combination effect prediction and adaptive optimization of ventilation operation, improves the adaptability, synergy and safety of duck house environment regulation.

Description

Technical Field

[0001] This invention belongs to the fields of smart farming, intelligent regulation of livestock and poultry house environment, artificial intelligence and automatic control technology, and specifically discloses a method and system for intelligent ventilation control of duck houses based on dynamic regulation effect diagram. Background Technology

[0002] The environmental conditions of large-scale duck houses are influenced by a variety of factors, including indoor and outdoor temperature and humidity, duck age and weight, stocking density, heat and moisture production by the ducks, bedding moisture content, ventilation equipment performance, and airflow organization within the house. The ventilation system of a duck house not only needs to remove ammonia, carbon dioxide, water vapor, and waste heat, but also needs to prevent problems such as excessively rapid temperature drops, excessive wind speeds in the duck activity areas, and direct drafts of cold air. Different control objectives within the duck house are often coupled. For example, increasing fan speed can increase ventilation volume and reduce ammonia concentration, but may simultaneously cause a drop in house temperature, increased energy consumption, and cold stress in the ducks; increasing the opening of air inlets can reduce airflow resistance, but may alter the negative pressure and airflow point within the house; intermittent ventilation can balance air exchange and heat preservation, but improper ventilation cycle settings may cause periodic fluctuations in the concentration of harmful gases.

[0003] Currently, various technical solutions have been proposed in the field of duck house or poultry house environmental control. For example, Chinese patent application CN116029604A discloses a method for environmental control of caged meat ducks based on health and comfort. This method obtains environmental monitoring information inside and outside the duck house and health characteristics of the meat ducks, uses correlation analysis to determine key health and comfort indicators and constructs a comfort index, then uses a neural network to predict future environmental conditions and the comfort index, and finally conducts environmental control based on the prediction results. This method incorporates duck flock health characteristics into the control basis, which improves the scientific nature of the control to a certain extent. However, the control decision of this method still uses the deviation between the predicted value and the preset threshold as the trigger condition, which is essentially a "deviation-driven" threshold-triggered control and fails to achieve multi-objective collaborative optimization. Its control output is a single fuzzy control quantity, and does not involve the generation and evaluation of combined strategies of various ventilation operations such as fan speed adjustment, air inlet window adjustment, intermittent ventilation, and zoned ventilation. In addition, this method lacks an online learning and closed-loop update mechanism based on actual control response feedback, making it difficult to adapt to the dynamic changes in duck house structure, equipment performance, and breeding batches.

[0004] For example, Chinese patent application CN119620808A discloses a method for predicting and controlling the winter breeding environment in poultry houses. It predicts temperature and carbon dioxide levels through time series decomposition and causal feature screening, uses PID controllers to regulate temperature and carbon dioxide separately, and optimizes PID parameters using a genetic algorithm. This method achieves coordinated control of temperature and carbon dioxide to a certain extent. However, its control execution layer is still traditional PID feedback control with a fixed control law, unable to dynamically adjust the structure and parameters of the control strategy according to the age of the duck flock, stocking density, season, and other scenario conditions. Its modeling of ventilation operations is limited to a single device (fan), failing to address the combined effects of multiple types of operations such as air inlet window opening adjustment, air inlet direction adjustment, intermittent ventilation, and zoned ventilation. It also fails to describe the synergistic effects (e.g., fan speed adjustment combined with air inlet window opening can improve airflow distribution while reducing energy consumption) and offsetting effects (e.g., increasing the conflict between ventilation and heat preservation requirements). Its optimization objectives are limited to only two parameters: temperature and carbon dioxide, without incorporating multi-dimensional control objectives such as equipment load, energy consumption, and duck flock behavior responses. In addition, although Chinese patent application CN121165860A involves multi-objective optimization and self-learning mechanisms, its control object is a general farm environment and it does not specifically model the special characteristics of duck house ventilation operation (combination of fans, air inlets, intermittent ventilation, and zoned ventilation); Chinese patent application CN121613990A mainly focuses on the linkage control of environment and feeding, and does not involve the modeling and rolling optimization of the combined effect of ventilation operation.

[0005] In summary, the existing technologies have the following shortcomings: (1) They often solidify breeding experience into fixed thresholds or condition-action rules, making it difficult to describe the differentiated effects of the same control operation in different seasons, breeding batches, duck flock status and outdoor environments; (2) They usually analyze single equipment operations separately or use fixed combinations for multiple operations, which cannot effectively characterize the synergistic, offsetting and sequential effects between fan speed regulation, air inlet window adjustment, air inlet direction adjustment, intermittent ventilation and zoned ventilation; (3) They use general machine learning models to directly output control quantities, lacking the constraints of breeding knowledge and safety boundaries, which may generate control strategies that do not meet the minimum safe ventilation volume, the allowable wind speed of the duck flock or the equipment operating capacity; (4) They lack a closed-loop learning mechanism based on actual control response, which cannot continuously optimize the control effect model.

[0006] To address the aforementioned problems, this invention provides a method and system for intelligent ventilation control in duck houses based on dynamic regulation effect graph learning and rolling optimization of combined control strategies enhanced by aquaculture knowledge. Summary of the Invention

[0007] In view of the problems mentioned in the background art, such as the existing duck house ventilation control methods having fixed thresholds based on breeding experience which are difficult to adapt to different scenarios, the inability to characterize the synergistic and offsetting effects between multiple ventilation operations, and the lack of a closed-loop learning mechanism based on actual response feedback, this invention provides a duck house intelligent ventilation control method and system based on dynamic regulation effect diagram.

[0008] This invention includes the following technical solutions: This invention provides a method for intelligent ventilation control in duck houses based on a dynamic regulation effect diagram, comprising the following steps: S1: Data acquisition and knowledge acquisition.

[0009] Acquire aquaculture knowledge units, historical environmental data of duck houses, outdoor meteorological data, duck flock status data, historical operation data of equipment, and response data before and after operation execution; the aquaculture knowledge units shall include at least the applicable scenarios, control operations, expected effects, safety constraints, equipment interlock relationships, and operation sequence relationships.

[0010] The historical data of the duck house includes one or more of the following: temperature, relative humidity, ammonia concentration, carbon dioxide concentration, wind speed, pressure difference, and bedding moisture content; the historical operating data of the equipment includes one or more of the following: fan start / stop status, fan speed, fan current, fan power, air inlet window opening, and air inlet direction; the response data includes one or more of the following: changes in environmental conditions, changes in duck flock behavior, changes in equipment load, and energy consumption.

[0011] The aquaculture knowledge unit also includes initial environmental state, operation execution intensity, operation duration, potential side effects, and knowledge confidence. The aquaculture knowledge unit is divided into two categories: hard constraint knowledge and soft prior knowledge. Hard constraint knowledge is used to determine the aquaculture safety boundary, equipment capability boundary, equipment interlocking relationships, and directional locking effect edges that cannot be modified online. Soft prior knowledge is used to initialize the learnable parameters of the directed control effect edges and higher-order interaction relationships, and allows for limited updates based on operational data. The aquaculture knowledge unit is not directly output as a fixed ventilation control scheme, but is used to determine the initial nodes, connectivity relationships, direction of action, parameter range, and safety boundary of the dynamic control effect diagram.

[0012] S2: Construction of effect map.

[0013] Based on the aforementioned aquaculture knowledge units and historical operational data, a dynamic regulation effect diagram is constructed. The effect diagram is expressed as equation (1): in, This is a dynamic regulation effect diagram. For a set of nodes, For the set of directed control effect edges, It is a set of higher-order interaction relationships. Let be the parameter set of edges and higher-order interaction relationships. The set of nodes is represented by equation (2): in, For scene nodes, For state nodes, For operation nodes, For the response node.

[0014] The scenario nodes include at least one of the following: duck house season, breeding batch, duck age, weight, and stocking density; the status nodes include at least one of the following: duck house temperature, relative humidity, ammonia concentration, carbon dioxide concentration, wind speed, pressure difference, bedding moisture content, and duck behavior status; the operation nodes include at least one of the following: fan start / stop, fan speed adjustment, air inlet window opening adjustment, air inlet direction adjustment, intermittent ventilation, and zoned ventilation; the response nodes include at least one of the following: environmental status change, duck behavior change, equipment load change, and energy consumption.

[0015] The directed control effect edge connects the operation node and the state node or response node, and is used to characterize the change of a single control operation relative to the baseline without new control intervention. The directed control effect edge includes at least the starting node, ending node, effect direction, effect strength, response delay, effect duration, effect decay parameter, and applicable scenario. The edge weight at the current moment is determined according to equation (3): in, For the current operation node To the status node or response node The right of the border; As a scene state, the scene nodes include at least one of the following: duck house season, breeding batch, duck age, weight, and breeding density; The overall status of the duck house includes at least one of the following: duck house temperature, relative humidity, ammonia concentration, carbon dioxide concentration, wind speed, pressure difference, bedding moisture content, and duck flock behavior. For the first The control operation includes at least one of the following: fan start / stop, fan speed adjustment, air inlet window opening adjustment, air inlet direction adjustment, intermittent ventilation, and zoned ventilation.

[0016] The higher-order interaction relationship connects at least two operation nodes with at least one state node or response node, and is used to characterize the synergistic effect, offsetting effect, or temporal effect among multiple control operations. The higher-order interaction relationship includes at least the set of participating control operations, the order of operation execution, the time interval between operations, the direction of the interaction effect, the intensity of the interaction effect, and the confidence level of the interaction effect.

[0017] Operation set The interaction effect is determined according to equation (4): in, For operation set The interaction effect vector, For operation set The actual state change vector after execution relative to the baseline without new control intervention. For the first in the operation set The vector of the single-item regulatory effect of the operation. When When the direction of the effect is consistent with that of a single regulatory effect and its absolute value increases, it is determined to be a synergistic effect; when When the overall regulatory effect is weakened or its direction is changed, it is identified as a counterbalancing effect; when the order or time interval of operation results in different response outcomes, it is identified as a timing effect.

[0018] An initial feature vector is constructed for each type of node, and scene nodes, state nodes, operation nodes, and response nodes are uniformly mapped to a linear mapping. The dimensional node features are shown in equation (5): in, For nodes The original characteristics, For node type, and These are the mapping parameters for the corresponding node type.

[0019] In one implementation, a directed graph neural network, graph attention network, or hypergraph neural network is used to update node features based on neighboring node information and higher-order interaction relationships, as shown in equation (6): in, For nodes The set of adjacent nodes, For containing nodes A set of higher-order interaction relationships and These are the edge attention weights and higher-order interaction weights, respectively. It is a non-linear activation function.

[0020] S3: Baseline prediction.

[0021] The system acquires the current environmental status of the duck house, the status of the duck flock, outdoor weather conditions, and equipment operating status. Under the condition of maintaining the current equipment operating status and preset safe ventilation control unchanged, and without superimposing candidate incremental operations, it predicts the baseline environmental status within a preset time domain. It integrates the current environmental data from multiple areas of the duck house, the duck flock status, scene information, and equipment status to form a comprehensive duck house status. The duck house is divided into multiple environmental monitoring areas, with each area serving as a spatial node. Area connection edges are established based on the location relationships and airflow connectivity between areas. Based on the environmental data, duck flock status, and equipment operation at continuous sampling times in each area, a spatiotemporal prediction model is used to obtain the baseline environmental status of each area and the predicted duck flock risk for the preset time domain.

[0022] Based on the overall status of the duck house and historical equipment operation at multiple consecutive sampling times, the environmental status and duck flock risk within the future preset time domain are obtained through time series prediction models, state space models, or spatiotemporal graph prediction models, as shown in equation (7): in, The length of the historical observation window. To predict the time domain, This is a historical sequence of regulatory operations. This represents the current scene state.

[0023] S4: Strategy generation.

[0024] Based on the current state, baseline environmental state, control objectives, and remaining equipment regulation capacity, multiple candidate control sequences are generated, and candidate control sequences that violate equipment capacity boundaries, equipment interlock relationships, aquaculture safety boundaries, or control change rate constraints are eliminated.

[0025] No. The candidate combination control strategy is expressed as equation (8): in, For the first A candidate combination control strategy For operation type, For the sake of enforcement strength, To control the cycle number, For duration, This represents the number of operations included in the candidate strategy. The candidate combination control strategy is generated through operation combinations, execution intensity configuration, and control cycle arrangement, eliminating sequences that violate device interlocking relationships or safety constraints. The operation execution order is determined by the arrangement order of the control cycles, and the operation duration is determined by the number of consecutive control cycles occupied by the same or similar operations.

[0026] Based on the current scene and state characteristics, the state-related weights of each individual effect and interaction effect are obtained through a weight generation network, as shown in equations (9) and (10): in, This is the result of concatenating the current scene node features, state node features, and candidate operation features. For the first The weight of the single-item regulatory effect, For the first Weights of higher-order interaction effects.

[0027] S5: Effect prediction and evaluation.

[0028] Based on the dynamic control effect diagram, according to the response delay, duration, and effect decay process of each operation, the incremental effect of each candidate combination control sequence in future control cycles is predicted. This incremental effect is then superimposed onto the baseline environmental state to obtain the predicted state after intervention for each candidate combination control sequence, and a comprehensive evaluation value is calculated. Candidate combination control strategy The predicted incremental effect is obtained according to equation (11): in, Candidate Combination Control Strategy In the The predicted incremental effect of each control cycle For the first Item operation in the The single-item regulatory effect of a control cycle For the first The higher-order interaction relationship in the first The interaction effect of each control cycle and These are the corresponding effect weights. The number of operations included in the candidate combination control strategy. The number of activated higher-order interactions.

[0029] The predicted incremental effect is superimposed on the baseline predicted state under the condition of no new control intervention according to equation (12): in, Candidate Combination Control Strategy The corresponding predicted state after intervention, This represents the baseline predicted state under conditions of no new control interventions.

[0030] After obtaining the predicted state after intervention corresponding to the candidate combination control strategy, the comprehensive evaluation value is calculated. The comprehensive evaluation value is determined based on at least one of temperature deviation, humidity deviation, harmful gas risk, duck flock stress risk, equipment operating load, and energy consumption, and is calculated according to formula (13): in, To predict the time domain, For the cost of temperature control, The cost of humidity control The cost of air quality risks associated with ammonia and carbon dioxide. The cost of heat stress, cold stress, or airflow stimulation for duck flocks This comes at the cost of equipment load and frequent start-stop cycles. As a result of energy consumption, to These are evaluation weights determined based on the breeding stage and scenario conditions. Each evaluation weight is dynamically adjusted according to the current breeding stage and scenario conditions. For example, the air quality risk weight is increased in scenarios with rapidly rising ammonia levels, and the temperature and cold stress risk weights are increased in low-temperature scenarios during the brooding period.

[0031] S6: Rolling optimization and execution.

[0032] Under the constraints of minimum safe ventilation volume, wind speed in the duck activity area, pressure difference in the duck house, equipment control range, and equipment change rate, the target control sequence with the optimal comprehensive evaluation value is determined. Only the control operation corresponding to the current control cycle in the target control sequence is executed, and the state is reacquired and the target control sequence is solved again at the next control time. The safety constraints include at least Equation (14):

[0033]

[0034] in, This is the current ventilation volume. The minimum safe ventilation volume is determined based on the number of ducks, average weight, and air quality. Forecast wind speed in the area where ducks are active. This represents the maximum permissible wind speed for the current age of the duck flock. and These are the lower and upper limits of the control quantity for the corresponding equipment, respectively. For the first Each device at time The control quantity.

[0035] The candidate combined control strategy with the best comprehensive evaluation value and that satisfies the safety constraints is determined as the target control sequence according to equation (15): in, For the target control sequence, A set of control strategies to meet the safety boundaries and equipment capacity constraints of aquaculture.

[0036] The rolling optimization includes: at the current control moment, predicting and comprehensively evaluating the effects of multiple candidate combined control strategies within a preset future time domain, determining the target control sequence with the optimal comprehensive evaluation value and meeting safety constraints; executing only the first operation vector corresponding to the current control cycle in the target control sequence; at the next control moment, updating the comprehensive state of the duck house based on newly obtained environmental data, duck flock status, and equipment response, and regenerating and optimizing subsequent combined control strategies.

[0037] The current cycle operation in the target control sequence is converted into ventilation equipment control commands and sent to the executing device. The control commands include at least one or more of the following: fan start / stop commands, target fan speed, target air inlet opening, air inlet direction or guide device angle, fan running time, fan stop time, and combinations of zoned ventilation equipment. The equipment control commands are sent to the corresponding executing device via a programmable logic controller, edge controller, or industrial computer.

[0038] S7: Feedback Update.

[0039] The actual environmental response, duck flock response, and equipment response after control execution are obtained. The deviation between the actual response and the predicted response is calculated. Based on the normalized counterfactual contribution, the deviation is assigned to the corresponding directed control effect edge and higher-order interaction relationship. The learnable parameters are updated with a limited amplitude. The updated parameters are first entered into the model to be verified. After verification by prediction accuracy, safety constraints, and stability, the updated parameters replace the current running model. The safety hard constraints, equipment physical boundaries, and emergency control rules do not change during the online learning process.

[0040] After control is executed, actual environmental responses, duck flock behavior responses, equipment responses, and energy consumption are collected, and the deviation between the actual response and the predicted response is calculated. Based on the difference between the actual effect of the complete combined strategy and the predicted results after removing individual operations or higher-order interactions, the normalized counterfactual contribution of each individual operation and combined relationship is determined. When the confidence level of the corresponding effect parameter is not lower than a preset confidence threshold, and the number of historical samples in similar scenarios is not lower than a preset sample threshold, the actual response deviation is allocated to the corresponding directed control effect edge or higher-order interaction based on the normalized counterfactual contribution. When the confidence level or the number of historical samples does not meet the preset conditions, edge-level online updates are paused, and the corresponding control samples are stored in the calibration sample set.

[0041] Based on response bias, effect contribution, and data reliability, the effect strength, response delay, effect duration, and confidence level of directed control effect edges corresponding to soft prior knowledge, as well as the interaction effect strength and confidence level of higher-order interaction relationships, are updated with limited amplitude. When any of the following falls below a preset condition: the reliability of the actual response data, the confidence level of the corresponding effect parameter, or the sufficiency of historical samples, the online update of the corresponding parameter is paused. When the updated dynamic control effect map leads to an increase in prediction error or a rise in safety risk over multiple consecutive control periods, the corresponding effect parameters are restored to the stable version before the update.

[0042] The effects of safety hard constraints and physical locking do not change with online updates. The breeding safety boundary, equipment hard constraints, flock safety boundary, and orientation locking effect edges are not automatically modified during the online learning process. Hard constraint knowledge, minimum safe ventilation volume, equipment hard constraints, permissible wind speed for the flock, and emergency ventilation conditions remain constant, ensuring the safety and stability of the intelligent optimization process.

[0043] This invention also provides an intelligent ventilation control system for duck houses based on a dynamic regulation effect diagram, comprising: The data acquisition module is used to collect environmental data from multiple areas of the duck house, outdoor meteorological data, duck flock status data, and equipment operation data. The environmental data acquisition module includes temperature and humidity sensors, ammonia sensors, wind speed sensors, carbon dioxide sensors, differential pressure sensors, and litter moisture content sensors deployed in the front, middle, and rear of the duck house, as well as in various environmental monitoring areas at the duck flock's activity height.

[0044] The duck flock status acquisition module is used to acquire information on duck breeding batches and duck flock behavior. The module includes cameras and sound sensors installed on the top or side walls of the duck house to acquire information on duck distribution, aggregation, open-mouth breathing, wing-spreading behavior, and abnormal vocalizations.

[0045] The equipment status acquisition module is used to acquire the operating status of the fan and air intake equipment, including fan start / stop status, fan speed, fan current, fan power, air inlet opening, air inlet direction, equipment fault status, and cumulative equipment operating time.

[0046] The aquaculture knowledge processing module is used to process the experience of aquaculture experts into scenarios, control operations, expected effects, side effects, safety constraints, and operational interactions, forming aquaculture knowledge units.

[0047] The dynamic regulation effect graph learning module is used to construct and update scene nodes, state nodes, operation nodes, response nodes, directed weighted edges, and higher-order interaction relationships, and to perform the dynamic regulation effect graph construction and update operation in step S2.

[0048] There is no new control intervention baseline prediction module, which is used to construct the overall status of the duck house and predict the environmental status of the duck house and the risk to the duck flock in the future preset time domain, and to perform the baseline prediction operation in step S3.

[0049] The combined strategy generation module is used to generate candidate combined control strategies that include operation type, execution intensity, execution order and duration, and executes the candidate combined control strategy generation operation in step S4.

[0050] The effect prediction and rolling optimization module is used to predict the regulatory effect of candidate combination control strategies, calculate the comprehensive evaluation value, and determine the target combination control strategy under the constraints of aquaculture safety boundary and equipment capacity, and execute the effect prediction and rolling optimization operations in steps S5 and S6.

[0051] The equipment control module is used to convert the current cycle operation in the target sequence into equipment control commands and send them to the execution equipment. The control commands include at least one or more of the following: fan start / stop command, fan target speed, air inlet window target opening, air inlet direction or guide device angle, fan running time, fan stop time, and combination of zoned ventilation equipment.

[0052] The feedback update module is used to update the learnable parameters in the dynamic regulation effect diagram based on the deviation between the actual response and the predicted response, and to perform the feedback update operation in step S7.

[0053] The modules described above can be deployed on the same industrial computer or edge computing device, or they can be deployed separately on the duck house field controller and the remote server. Functions involving minimum safety ventilation, equipment protection, and emergency exhaust should be prioritized for deployment on the field controller to maintain basic ventilation safety in the event of a network outage.

[0054] Compared with the prior art, the present invention has the following beneficial effects: (1) Enhanced aquaculture knowledge and adaptive scenario control. This invention integrates aquaculture experience, historical monitoring data, equipment operation records, and control responses to learn the actual control effects of various ventilation operations under different seasons, batches, and duck flock conditions, avoiding the solidification of experience into fixed thresholds or control rules. Through the dynamic perception of scenario nodes and state nodes, the effect intensity of the same control operation is adaptively adjusted according to the current scenario state, improving the scenario adaptability of control decisions.

[0055] (2) Modeling of combined effects improves prediction accuracy. This invention uses a dynamic control effect diagram to characterize the impact of a single operation on the environment, flock status, equipment load, and energy consumption, and describes the synergy, cancellation, and timing effects between multiple operations through high-order interaction relationships. The actual effect of the combined operation is calculated using formula (4) to calculate the interaction effect, which compensates for the deviation of simple superposition prediction and improves the accuracy of combined control effect prediction.

[0056] (3) Autonomous strategy optimization and multi-objective collaborative decision-making. Based on the real-time status and future trends of the duck house, this invention autonomously generates candidate strategy combinations that include operation type, execution intensity, execution order and duration. Through the comprehensive evaluation function of formula (13), it weighs multiple dimensions such as temperature, humidity, air quality, duck stress, equipment load and energy consumption, and performs rolling optimization under the constraints of breeding safety boundary and equipment capacity, thus realizing multi-objective collaborative decision-making.

[0057] (4) Closed-loop online learning for continuous adaptation and optimization. This invention updates the single-item control effect, combined interaction effect, and their confidence level based on the deviation between the actual response and the predicted response, enabling the system to continuously adapt to changes in duck house structure, equipment performance, and breeding batches. Through limit update and parameter rollback mechanisms, the learning ability is guaranteed while avoiding model degradation caused by abnormal data.

[0058] (5) Balancing safety and adaptability. This invention constructs a multi-layered safety protection system by constraining the minimum safe ventilation volume (Formula 14a), the permissible wind speed for duck flocks (Formula 14b), and the equipment capacity (Formula 14c). By combining the separation design of hard constraint knowledge and soft prior knowledge, abnormal data identification, and parameter rollback mechanism, the invention ensures the safety and stability of the intelligent optimization process and improves the adaptability, synergy, and reliability of duck house environmental control. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the overall process of the intelligent ventilation control method for duck houses based on dynamic regulation effect diagrams of the present invention. Figure 2 This is a schematic diagram of the node, edge, and higher-order interaction structure of the dynamic regulation effect diagram of the present invention. Figure 3 This is a schematic diagram of the structure of the modulation effect learning network of the present invention; Figure 4 This is a flowchart of the candidate combination control strategy generation and rolling optimization process of the present invention; Figure 5 This is a flowchart of the actual control response evaluation and effect diagram online update process of the present invention; Figure 6 This is a schematic diagram of the composition and structure of the intelligent ventilation control system for duck houses of the present invention; Figure 7 This is a real-time overview interface diagram of the intelligent ventilation control system of the present invention; Figure 8 This is a diagram of the duck house zoned environmental monitoring interface of the present invention; Figure 9 This is a diagram of the interface for collaborative control and command issuance of the ventilation equipment according to the present invention; Figure 10 This is a diagram of the interface for candidate combination strategy generation and rolling optimization in this invention; Figure 11 This is a diagram of the control response evaluation, alarm, and operation log interface of the present invention; Figure 12 This is a diagram showing the interface for configuring aquaculture knowledge parameters and safety boundaries in this invention. Detailed Implementation

[0060] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are merely preferred embodiments of the present invention and do not constitute any limitation on the scope of protection of the present invention.

[0061] Example 1: This embodiment is compared with Figure 1 The overall flowchart shown and Figure 6 The system composition shown illustrates the overall deployment method and key parameter configuration of this invention in a large-scale duck house.

[0062] like Figure 6 As shown, the system of this invention includes a data acquisition module, a duck flock status acquisition module, an equipment status acquisition module, aquaculture knowledge processing module, a dynamic regulation effect map learning module, a baseline prediction module without new control intervention, a combined strategy generation module, an effect prediction and rolling optimization module, an equipment control module, and a feedback update module. Each module is deployed in an edge controller (model: Nvidia Jetson orin nano) and communicates with sensors and actuators via industrial Ethernet.

[0063] The duck house has the following structural parameters: 60 m long, 15 m wide, and 3.5 m high eaves, with an effective rearing area of ​​approximately 900 m². 2 The ventilation system employs a combination of longitudinal and minimum ventilation, with the main ventilation fan installed on one side of the gable wall and air inlets distributed on both longitudinal walls. When the duck house holds 8000 ducks, the average stocking density is approximately 8.9 ducks / m². 2 When the average weight of the ducks is 2.5 kg / bird, the corresponding live load per unit area is approximately 22.2 kg / m². 2 The area was divided into three longitudinal zones—front, middle, and rear—based on distances of 0–20 m, 20–40 m, and 40–60 m from the main air intake. Environmental monitoring points were set up at the ducks' activity height and the upper airflow height within the shed, forming a total of six basic environmental monitoring zones. Figure 7 and Figure 8 As shown, the system's real-time overview interface and zoned environmental monitoring interface provide a visual display of the overall environmental status of the duck house and its zones, allowing managers to intuitively view various environmental parameters and duck flock risk information. Additional monitoring points can be set up for areas with uneven airflow distribution, damp bedding, or frequent duck gatherings.

[0064] Each monitoring area is equipped with the following sensors: temperature and humidity sensor (Model: Sensirion SHT35, accuracy ±0.2°C, ±2%RH), ammonia sensor (Model: Alphasense NH3-B1, range 0~100ppm, accuracy ±2ppm), carbon dioxide sensor (Model: SenseAir S8, range 0~5000ppm, accuracy ±30ppm), wind speed sensor (Model: RMYoung 81000, accuracy ±0.1m / s), and differential pressure sensor (Model: Setra 264, range ±250Pa, accuracy ±1Pa). The moisture content of the bedding material is monitored in situ using a near-infrared moisture meter (Model: Bruker TGO), with data collected every 4 hours. The sensor sampling period is set to 5 seconds. Data is collected via RS-485 bus to the edge controller for preprocessing, including median filtering for noise reduction, outlier removal (3σ principle), and linear interpolation for missing values.

[0065] The outdoor weather station (model: Davis Vantage Pro2) collects outdoor temperature, humidity, wind speed, wind direction, rainfall and light intensity in real time, with a sampling period of 60 seconds.

[0066] The duck flock behavior perception system includes five Hikvision DS-2CD2646FWD-LZS cameras, evenly suspended from the ceiling of the duck house at a height of 3m above the ground, covering the entire shed's field of view. The cameras continuously capture overhead images at a sampling frame rate of 30fps. An edge controller deploys a duck flock target detection model based on YOLOv8, providing real-time output: duck flock aggregation level (number of ducks per unit area, unit: ducks / m²). 2 The features extracted included: the percentage of ducks breathing with their mouths open (number of ducks breathing with their mouths open / total number of ducks × 100%), the percentage of ducks with their wings spread (number of ducks with their wings spread / total number of ducks × 100%), the activity level (rate of change of duck position, unit: m / s), and the percentage of ducks far from the air intake area (number of ducks in the area > 40m from the air intake / total number of ducks × 100%). The feature extraction frame interval was 30 seconds.

[0067] Ventilation equipment configuration: 6 variable frequency fans (model: EM-50, 2.2kW per unit, air volume range 0~45000m³ / h). 3 / h, speed range 0~1450rpm), 4 fixed frequency fans (model: EM-40, single unit power 1.5kW, air volume 35000m³ / h). 3 / h), 24 air inlet window actuators (model: Siemens GIB135.1E, opening range 0~100%, response time ≤150s), and 6 sets of air diversion devices (angle adjustment range -30°~+30°, electric actuator response time ≤30s).

[0068] The system sampling period is set to 5 seconds, the control optimization period to 2 minutes, the prediction time domain H to 30 minutes (15 control cycles), and the historical observation window L to 60 minutes. These parameter configurations are determined based on the size of the duck house and the equipment response speed, and can be adjusted accordingly for different duck houses.

[0069] like Figure 1 As shown, the system operates according to the following process: S1 Data acquisition and knowledge acquisition → S2 Effect map construction → S3 Baseline prediction → S4 Strategy generation → S5 Effect prediction and evaluation → S6 Rolling optimization and execution → S7 Feedback update, forming a closed-loop control.

[0070] Example 2: Candidate strategy generation and rolling optimization in winter brooding scenarios.

[0071] This example uses the winter brooding stage as an example. The ducks are 7 days old, with an average weight of about 0.15 kg / duck, a stock size of 8000 ducks, and an effective rearing area of ​​about 900 m². 2 The average stocking density in the entire shed is approximately 8.9 birds / m². 2 The outdoor temperature is 2°C, and the outdoor wind speed is 3–5 m / s. The target temperature range corresponding to the current age is determined by the age-temperature curve, the flock's aggregation status, and activity behavior; in this example, it is set to 28–30°C. The system simultaneously maintains control objectives such as minimum safe ventilation, air quality, and preventing cold air from directly entering the ducks' activity area.

[0072] (a) Current state acquisition (S1, time t=0) The real-time data of each sensor are shown in Table 1: Table 1 Real-time values ​​of environmental parameters in each monitoring area Duck flock behavior: 2% of ducks were breathing with their mouths open (normal), and the aggregation level was 1.8 ducks / m². 2 (High, normal threshold ≤1.5), wingspan ratio 1% (normal), activity level 0.12m / s (low).

[0073] Current equipment status: 2 variable frequency fans are running at 40% speed (approximately 580 rpm), 1 fixed frequency fan is running, the average opening of the air inlet window is 35%, and the angle of the air guide device is uniformly 0° (horizontal air outlet).

[0074] (ii) Baseline Prediction (S2) A GCN-GRU spatiotemporal graph neural network is used to predict the baseline for the next 30 minutes. The input consists of multivariate time-series data from six regions (12 sampling points) over the past 60 minutes and device status data. The output is the predicted environmental status for each region for the next 30 minutes. This prediction corresponds to equation (7): in, (Historical observation window length, corresponding to 60 minutes) (Prediction time domain, corresponding to 30 minutes) This is a historical sequence of regulatory operations. This represents the current scene state.

[0075] Taking the ammonia concentration in the middle-lower layer as an example, the prediction results are shown in Table 2: Table 2. Baseline Predicted Ammonia Concentration Values ​​in the Middle-Lower Layer Without additional control measures, the ammonia concentration in the middle-lower zone will reach 34.2 ppm in 30 minutes, exceeding the safe threshold (25 ppm) for the brooding period, and the cold stress risk index for the duck flock will rise from the current 0.2 to 0.55 (threshold 0.5).

[0076] (III) Real-time assignment of scene nodes, state nodes and effect graphs (S3) Based on the current scenario (winter brooding, 7 days old, 0.15 kg weight, density 55 chicks / m²), 2 The relevant edge weights in the dynamic regulation effect diagram are calculated using equation (3): Operating node: "Increase fan speed" Reaching the "Ammonia Concentration" state node Take edge weights as an example. Figure 2 As shown, the operation nodes in the effect diagram Connect to the state node via directed weighted edges. (Ammonia concentration), the edge weight is dynamically determined by equation (3): Scene status The feature vector is [winter, brooding period, age 7 days, density 55], representing the overall condition of the duck house. This includes current temperature and humidity, ammonia concentration, and duck flock behavior. The effect graph learning module outputs the following based on statistical data from historical similar scenarios (winter brooding, outdoor temperature 0-5°C): This means that, in the current scenario, for every 1% increase in the fan speed, the ammonia concentration in the middle-lower layer decreases by 0.58 ppm.

[0077] Similarly, the edge weight from the same operation to the "temperature" state node is: That is, for every 1% increase in fan speed, the regional temperature decreases by 0.035°C. The edge weights to the "cold stress risk" response node are: The cold stress risk index increases with increasing rotational speed.

[0078] "Increase air intake window opening" operation The edge weights for "ammonia concentration" are: "Raise the air intake position" (adjust the airflow guide angle by +15°, i.e., direct airflow upwards) operation The marginal weight for "cold stress risk" is: (iv) Candidate strategy generation and calculation of higher-order interaction effects (S4) like Figure 4 As shown in Table 3, based on the current state, baseline prediction, and control targets (ammonia concentration ≤25ppm, temperature ≥22°C, cold stress risk ≤0.5), five candidate combined control strategies are generated. Table 3 Candidate Combination Control Strategies Each candidate strategy is expressed in the form of equation (8): like Figure 10 As shown, the candidate combination strategy generation and rolling optimization interface visualizes the five candidate strategies and their status in the steps of baseline prediction, candidate generation, safety pruning, and effect evaluation. Operators can intuitively view the comprehensive evaluation scores and safety constraint verification status of different strategies (such as U1 to U5).

[0079] Candidate strategy For example, ,in .

[0080] like Figure 2 As shown at the bottom, the higher-order interaction relationship connects both operation A (increasing fan speed) and operation B (increasing inlet window opening) to the ammonia concentration state node. The calculation of this interaction effect is based on equation (4): The set of operations .

[0081] Prediction of single-item regulation effect (taking ammonia concentration change as an example): Operation A single effect (fan speed increased by 15%): Operation B single-item effect (increase air inlet window opening by 10%): The cumulative effect of single-item effects: However, actual operating data shows that there is a synergistic effect when the fan speed increase and the air inlet window opening are simultaneously executed in winter. Based on equation (4), the interaction effect is statistically analyzed from similar historical scenarios: Historical data shows that when the fan speed is increased by 15% and the air inlet opening is increased by 10% simultaneously, the actual ammonia reduction is 11.7 ppm. Therefore: The interaction effect is negative and in the same direction, indicating a synergistic effect between the two operations in ammonia removal (the actual effect is better than simple superposition). Meanwhile, the interaction effect has a positive impact on temperature reduction (an offsetting effect): increasing the fan speed leads to a temperature decrease of 0.525°C, increasing the air inlet opening increases the air volume, causing an additional temperature decrease of 0.12°C, but the increased air inlet opening reduces air resistance, improving fan efficiency, resulting in an actual temperature decrease of 0.58°C. The interaction effect... That is, the synergistic effect reduces the temperature drop by 0.065°C compared to the simple summation.

[0082] This synergistic effect is more pronounced in summer ventilation (interaction effect approximately -3.2 ppm), and weakens in winter due to lower intake air temperatures. The effect diagram automatically retrieves the corresponding interaction effect value based on the current scenario.

[0083] (V) Effect Prediction and Comprehensive Evaluation (S5) For candidate strategies Calculate the first according to formula (11) The predicted incremental effect of each control cycle ( (i.e., 10 minutes in) The weight coefficients are generated by the weights in equations (9)-(10) to produce the network output. Equations (9) and (10) are as follows: in This is the result of concatenating the current scene node features, state node features, and candidate operation features. For the first The weight of the single-item regulatory effect, For the first Weights of higher-order interaction effects.

[0084] In the current scenario, the weight generation network output is: , , (Ammonia has a higher weighting than temperature because the current problem is excessive ammonia levels, while the temperature is still within an acceptable range.)

[0085] Substitution formula (11): This increment is superimposed on the baseline state (at the 5th control cycle, i.e., 10 minutes, the baseline ammonia concentration is predicted to be 25.2 ppm), according to equation (12): Candidate strategy It is predicted that the ammonia concentration in the middle and lower layers will drop to 15.67 ppm after 10 minutes.

[0086] The comprehensive evaluation value of each candidate strategy within the complete time domain (30 minutes, 15 control cycles) is calculated using equation (13): Parameter settings: , , , , , (Air quality risk has the highest weight in the winter brooding scenario, followed by temperature). The comprehensive evaluation values ​​of each candidate strategy are shown in Table 4: Table 4. Overall Evaluation Values ​​of Each Candidate Strategy The one with the lowest overall evaluation value (8.5) and that satisfies all safety constraints was identified as the target control sequence.

[0087] (vi) Safety constraint verification (S6) right Verify the security constraints of equation (14): Current ventilation volume: The current air volume is .

[0088] Minimum safe ventilation Calculate according to formula (16): in Minimum safe ventilation volume (m³) 3 / h), The age coefficient is calculated as follows: 0.3 for the brooding period, 0.6 for the growth period, and 1.0 for the fattening period. This refers to the number of animals in stock. This refers to the average weight (kg). In this example, the chicks were 7 days old during the brooding period. , , kg, .satisfy .

[0089] Predicted wind speed in the duck activity area: After implementation, the predicted wind speed in the duckling activity area is 0.45 m / s, which is the upper limit of the permissible wind speed during the brooding period (7 days old). ,satisfy .

[0090] Equipment control parameters: Fan speed increased to 55% (maximum 100%), air inlet opening 45% (maximum 100%), satisfying the constraints.

[0091] (vii) Execution and Feedback Updates (S7) System Execution Operation during the current control cycle (minutes 0-2): The fan speed linearly increases to 55% within 2 minutes (i.e., The incremental pressure is applied evenly over 120 seconds), and the air intake window opening increases synchronously to 45% (response delay approximately 18 seconds). Figure 9 As shown, the ventilation equipment control interface clearly displays the operating parameters and command issuance process of the above-mentioned equipment, supporting the coordinated control of equipment such as variable frequency fans, fixed frequency fans, air inlets, and air diversion devices. After 2 minutes of execution, the actual ammonia concentration collected in the middle-lower layer was 23.0 ppm, and the deviation from the predicted value of 23.5 ppm in equation (12) was: The actual ammonia concentration decreased slightly faster than predicted. This bias, assessed with confidence (current effect parameter confidence 0.85, higher than the threshold of 0.70; number of historical samples for similar scenarios 127, higher than the threshold of 50), triggered an online update: the edge weight from operation A to the ammonia concentration was increased from -0.58 to -0.60 (maximum limit -0.65), with the interaction effect... The value has been adjusted from -1.8 to -2.0 (maximum amplitude limit -2.5). The updated effect diagram will be used for policy optimization in the next control cycle.

[0092] Example 3: Regulation effect under high temperature and high humidity conditions in summer.

[0093] This example uses the summer fattening stage (ducks 35 days old, approximately 2.5 kg / duck, 8000 ducks in stock, stocking density 55 ducks / m²). 2For example, refer to Figure 2 The effect diagram shown illustrates the adaptive changes in edge weights under high-temperature conditions. Outdoor temperature: 35°C; humidity: 80%; southerly wind: level 2.

[0094] Current status: Indoor temperature 31.5°C (central area), relative humidity 78%, ammonia concentration 8.5 ppm, carbon dioxide concentration 1800 ppm. 42% of the ducks are breathing with their mouths open (significantly higher than the normal threshold of 15%), 35% have their wings spread (higher than the normal threshold of 20%), and the flock density is 1.2 ducks / m². 2 (Normal), activity level 0.35m / s (slightly high).

[0095] Current equipment status: 4 variable frequency fans are running at 85% speed (approximately 1230 rpm), 3 fixed frequency fans are running, the average opening of the air inlet window is 75%, the air guide device angle is -10° (downward airflow), and the wet curtain water pump is fully open.

[0096] like Figure 2 As shown, the directed weighted edges from each operation node to the state node in the effect diagram change with the scene. Taking the edge weight from "increase fan speed" to "temperature" as an example: in summer, the outdoor temperature is higher than the indoor temperature, so increasing ventilation not only does not lower the temperature but also introduces hot air, so the edge weight is -0.01 (close to 0); while the edge weight from "open evaporative cooling pad" to "temperature" is -0.15°C / % power, and the edge weight from "increase air inlet window opening" to "wind speed" is +0.008(m / s) / % opening.

[0097] Candidate generation strategy: Increase the fan speed (+10%) and increase the air inlet window opening (+10%) to ensure uniform airflow throughout the area; Enhanced front-zone evaporative cooling pads (pump power +20%) + prioritized rear-zone fans, providing differentiated cooling for different zones; Increase the fan speed (+15%) and adjust the angle of the air guide device (+20°, directing air upwards to avoid the ducks' activity layer) to reduce the perceived temperature for the ducks.

[0098] Effect diagrams predict the incremental effects of each strategy. For example: increasing the fan speed by 15% increases the regional wind speed by 0.25 m / s; adjusting the airflow angle ensures that the cold air falls away from the duck activity layer (0.3~1.2 m above the ground), resulting in a wind speed increase of only 0.08 m / s in the duck activity area. The predicted proportion of open-mouth breathing decreases from 42% to 18% (the effect diagram is predicted through the "wind speed-felt temperature-duck behavior" path), and the proportion of wingspan decreases to 12%.

[0099] Calculation of overall evaluation value: (Temperature has the highest weight in high-temperature scenarios). , , (Heat stress has the highest weight). , .final The one with the lowest overall evaluation score (7.2) was selected as the target strategy.

[0100] After 3 minutes of execution, the actual temperature dropped by 1.8°C (predicted drop of 2.1°C), and the proportion of people breathing through their mouths decreased to 22% (predicted 18%). Bias analysis: Insufficient temperature decrease ( The heat stress relief in the duck flock was less than expected. Based on deviation analysis, the system found that the current power of the evaporative cooling pad pump had reached 100% (the upper limit of the hard constraint), making further improvement in cooling capacity impossible. However, the contribution of adjusting the flow angle was lower than the predicted value (actual contribution 0.4°C vs. predicted 0.6°C). Therefore, the confidence level of the edge weight from "flow angle adjustment" to "temperature" was reduced from 0.75 to 0.65 (below the update threshold of 0.70, so the edge update was paused), and the cooling effect of the flow guiding device was marked as having high uncertainty in this scenario. Subsequent strategies will reduce the dependence on this operation and increase the control weight of the evaporative cooling pad and water curtain. Simultaneously, the system recorded this deviation sample in the calibration sample set (number SC-2026-07-22-003) for unified calibration after accumulating more similar scenario samples.

[0101] Example 4: Model training and parameter calibration for effect maps.

[0102] This embodiment is compared with Figure 3 The shown control effect graph learning network structure illustrates the training process of the effect graph model before deployment and the periodic calibration mechanism after deployment.

[0103] (a) Construction of training dataset Continuous operational data for 10 months, from September 2025 to June 2026, was extracted from the historical operational database of the target duck house. The sampling period was 5 seconds, and after downsampling to a 2-minute interval, approximately 216,000 valid samples were obtained. The training set (70%), validation set (15%), and test set (15%) were divided according to the following rules: Training set: September 2025 to April 2026 (approximately 151,200 samples) Validation set: May 2026 (approximately 32,400 samples) Test set: June 2026 (approximately 32,400 samples) Each training sample is a triplet (pre-control state) ,operate Response after control The pre-control state includes environmental data from 6 areas and duck flock behavior data from the past 60 minutes (30 sampling points). Record the actual control operations performed and the response after control. Record the changes in state within 30 minutes after control is implemented.

[0104] (II) Initial Graph Structure Construction Based on aquaculture knowledge units, an initial effect diagram structure (S2) is constructed. Figure 2 Based on the experience of aquaculture experts, the initial directed weighted edges shown in Table 5 are established: Table 5 Initial List of Directed Weighted Edges Establish initial high-order interaction relationships: Increased fan speed + increased air inlet opening: Synergistic effect of ammonia removal (in Equation 4) absolute value ) Increased fan speed + larger air inlet opening: Temperature drop offsets the effect. Increased fan speed + evaporative cooling pad activation: a synergistic cooling effect (summer) Increased fan speed + downward airflow angle: Synergistic effect of cold stress in duck flocks (winter) Rear fan + front air inlet: Synergistic effect of airflow organization (III) Boundary weight learning and calibration like Figure 3 As shown, edge weight parameters are learned through a graph attention network. The edge weights are calculated based on the change from "fan speed increase" to "ammonia concentration". For example, extract all samples from the training set that "only increase the fan speed (no other operations)", and group them by scenario for statistical analysis: Table 6. Marginal weights for fan speed increase to ammonia concentration in different scenarios Scene features were fitted using the weighted least squares method. Relationship function with edge weight The weights are the sample confidence scores (based on sensor data completeness and consistency of manual annotation). The training loss function is shown in equation (17): in, For the loss of border rights training, For the first The weight of each sample (based on sensor data completeness and consistency of human annotation). To predict edge weights, For label edge weight, This is the regularization coefficient (taken as 0.001). These are the model parameters.

[0105] (iv) Determination of confidence threshold Confidence thresholds for each effect parameter are determined based on the validation set. This is achieved using directed weighted edges. Taking the example of (fan speeding up to ammonia), the prediction error of this edge on the validation set at different confidence levels is shown in Table 7: Table 7. Prediction errors at different confidence levels A confidence threshold of 0.70 was set (i.e., edges with a confidence level ≥ 0.70 can be used for online updates, while edges with a confidence level < 0.70 are paused for updates). Under this threshold, the coverage of trusted edges (confidence level ≥ 0.70) on the validation set was 73.2%, and the weighted average prediction error was 1.08 ppm, which meets the accuracy requirements for duck house environmental control (ammonia concentration control accuracy ± 3 ppm).

[0106] (v) Hard constraint boundary calibration Based on the hard constraint knowledge in the aquaculture knowledge unit, define a safety boundary that cannot be modified online: 1. The minimum safe ventilation volume shall be calculated according to formula (16): ,in The values ​​were 0.3 for the brooding period, 0.6 for the growth period, and 1.0 for the fattening period.

[0107] 2. Maximum permissible wind speed for duck flocks (divided by age): 1-7 days old: 0.5 m / s; 8-14 days old: 0.8 m / s; 15-28 days old: 1.2 m / s; >28 days old: 1.8 m / s.

[0108] 3. Equipment hard constraints: fan start-stop interval ≥ 60 seconds, single continuous operation time ≤ 120 minutes, fan speed daily change rate ≤ 10% / min, air inlet window opening change rate ≤ 5% / s.

[0109] 4. Emergency safety conditions: When the instantaneous ammonia concentration is ≥50ppm, the emergency exhaust mode will be activated without conditions; when the temperature is ≥38°C or ≤0°C, the safety protection mode will be activated without conditions.

[0110] The aforementioned hard constraints are written into the security monitoring module during system initialization and are not modified during the online learning process. A separate hard constraint watchdog thread is deployed in the edge controller, checking sensor readings every 200ms. Figure 12 As shown, the system parameter and safety boundary configuration interface is used to set the breeding scenario (such as season, age, weight), control and prediction cycle, as well as hard constraint safety boundaries that cannot be modified online (such as minimum safe ventilation volume, maximum allowable wind speed, emergency exhaust conditions, etc.) to ensure that the system operates safely and stably within the boundaries.

[0111] (vi) Regular calibration after deployment The system performs offline calibration every 7 days, using the accumulated calibration sample set (samples corresponding to edges with confidence < 0.70 or historical sample count < 50) accumulated over the past 7 days to uniformly retrain the relevant edge weights. During the calibration process, if the confidence of an edge fails to improve to above 0.55 in 3 consecutive calibrations, it is marked as an "uncertain edge," and its weight is reduced in the strategy generation (multiplied by a coefficient of 0.3~0.5). At the same time, an early warning is issued to the aquaculture management personnel for manual confirmation.

[0112] Example 5: Detailed process for control response evaluation and online updating of effect maps.

[0113] This embodiment is compared with Figure 5 The flowchart shown illustrates the actual control response evaluation and effect map online update process, detailing the complete process of control response evaluation, deviation allocation, and effect map online update in S7.

[0114] (a) Background of the Control Event Scenario: Winter brooding period, current moment The system executes the target strategy. Operation A: Increase fan speed by 15%, increase inlet window opening by 10% (Operation B), and increase guide angle by 15° (Operation C). These three operations are executed simultaneously for 30 minutes. The effect diagram predicts the change in ammonia concentration 30 minutes later before execution. ppm, temperature change °C.

[0115] (II) Actual response collection time) After 30 minutes, the actual data collected showed that the ammonia concentration decreased from 22.3 ppm to 12.8 ppm. ppm; actual change in temperature as it drops from 23.0°C to 22.2°C. °C.

[0116] (III) Deviation Calculation Ammonia deviation rate: Temperature deviation rate: .

[0117] (iv) Contribution Analysis The contribution of each operation and interaction is determined through counterfactual inference. From the effect diagram of the complete combined strategy, individual operations or higher-order interactions are removed one by one, and the difference in predictions before and after removal is calculated. After removing operation A (fan speed increase), the predicted decrease in ammonia is -3.2 ppm, and the temperature decrease is -0.15°C. → Ammonia contribution of operation A: ppm (contribution rate 71.4%) → Temperature contribution of operation A: °C (contribution rate 72.7%) After removing operation B (increasing the air inlet window opening), the predicted decrease in ammonia is -9.0 ppm, and the temperature decrease is -0.50°C. → Ammonia contribution of operation B: ppm (contribution rate 19.6%) → Temperature contribution of operation B: °C (contribution rate 9.1%) Remove higher-order interaction relationships After (synergistic effect), the predicted decrease in ammonia is -9.5 ppm (cumulative effect of individual factors) → Contribution of interaction effect: ppm (contribution rate 15.2%) (v) Deviation Allocation Based on the contribution ratio, the ammonia deviation will be... ppm allocated to each effect parameter: Operation A allocation deviation: ppm Operation B allocation deviation: ppm Interaction effect allocation bias: ppm (vi) Update of effect parameters The current confidence levels of all effect parameters are ≥0.70, meeting the conditions for online updates. Operation A edge weight update: current Based on the direction of the deviation, the actual effect is stronger than predicted (a greater decrease), so the absolute value should be increased: ; Constrained by the amplitude of soft prior knowledge ( ),Pick ; Operation B edge weight update: current Updated to: ; Higher-order interaction relationship update: Current Updated to: ; Confidence update: Based on the accuracy of this prediction and sample consistency, the confidence level has been increased from 0.85 to 0.87.

[0118] (vii) Handling of deviations that do not meet update conditions If the confidence level of a certain effect parameter is below 0.70 (for example, the confidence level of the newly introduced guide angle operation C is only 0.55), then online updates for that edge are paused, and the biased samples are stored in the calibration sample set (number SC-2026-07-22-018). During weekly offline calibration, the edge weights are retrained using accumulated similar scenario samples (at least 50). Figure 11 As shown, the alarm and operation log interface records detailed operation logs of these control events (such as policy execution, feedback update, and deviation allocation) as well as information on the sample set to be calibrated (such as SC-2026-07-22-018), providing reliable data support for system operation status backtracking, parameter anomaly investigation, and subsequent offline calibration.

[0119] (viii) Parameter rollback mechanism If the updated effect plot shows an increase in cumulative prediction error (the moving average of the prediction error increases by more than 20% compared to before the update) or triggers a safety warning (such as the actual temperature drop exceeding 1.5 times the predicted value) within the subsequent 5 consecutive control periods, parameter rollback will be triggered and executed according to formula (18): in, These are the parameter values ​​after rollback. The current parameter value. This is the most recent stable version (the version with a prediction error of <10% in the past 7 days). It is mixed in a 7:3 ratio to preserve recent learning outcomes while preventing overfitting. A rollback event (number RB-2026-07-22-001) is also recorded for offline analysis.

[0120] Example 6: Comparative experiment with traditional fixed threshold control method (corresponding) Figure 1 ).

[0121] This embodiment demonstrates the beneficial effects of the present invention through comparative experiments. The method of the present invention was alternately operated in the same duck house (according to...). Figure 1 The complete process shown) and the traditional fixed threshold control method were used for 7 days each (July 1 to 14, 2026), with similar outdoor conditions (average temperature 28.3°C vs 28.7°C, relative humidity 72% vs 74%), and the ducks were from the same batch (28-35 days old).

[0122] (I) Description of Traditional Methods Traditional fixed threshold control method: When the temperature is ≥28°C, start the fans (3 fans operating at full frequency); when the temperature is ≥30°C, increase to 6 fans; when the ammonia concentration is ≥20ppm, increase ventilation (add 2 fans); when the humidity is ≥80%, start the evaporative cooling pads. Each control logic operates independently, without combined optimization or rolling prediction.

[0123] (II) Comparison Results Table 8. Comparative experimental results of the method of the present invention and the traditional method (III) Effect Analysis The method of the present invention improves the temperature control accuracy by 56%, mainly due to the predictive baseline prediction of Equation (7) and the incremental effect prediction of Equations (11)-(12): the system predicts the temperature change trend 30 minutes in advance through Equation (7) and starts the coordinated cooling strategy before the temperature rises, avoiding the lag phenomenon of "starting only when the temperature exceeds the threshold" in the traditional method.

[0124] The peak concentration of ammonia was reduced by 44%, thanks to the interaction effect modeling of Equation (4) and the incremental effect prediction of Equation (11): the synergistic effect of increasing the fan speed and increasing the air inlet opening improved the ammonia removal efficiency by about 28% (actual measurement: increasing the fan speed alone can reduce ammonia by 5.2 ppm, increasing the air inlet opening alone can reduce ammonia by 2.1 ppm, and the actual reduction during execution is 9.8 ppm, with a synergistic contribution of 2.5 ppm).

[0125] The 18% reduction in energy consumption is attributed to the equipment load cost in the comprehensive evaluation function of equation (13). and energy consumption costs The introduction of this technology avoids the crude control approach of "multiple devices operating simultaneously" in traditional methods, and prioritizes the selection of equipment combinations with high energy efficiency ratios.

[0126] The daily weight gain of the duck flock increased by 10.9% (from 52.5g / d to 58.2g / d), which is attributed to the comprehensive improvement in environmental stability and comfort, thus verifying the economic benefits of the invention in actual aquaculture production.

[0127] Example 7: Adaptive changes in edge weights of effect graphs under different scenarios (corresponding) Figure 2 ).

[0128] This embodiment refers to Figure 2 The effect graph structure shown demonstrates the scene adaptability of equation (3) of this invention by comparing the edge weights of the same directed weighted edge in different scenarios.

[0129] Select the directed weighted edge from "fan speed increase" to "ammonia concentration". (i.e., operation A to the ammonia response node), extract the edge weights (all statistical means) for different seasons and different duck flock ages from the effect graph database, as shown in Table 9: Table 9 Comparison of edge weights for "increased fan speed → ammonia concentration" under different scenarios Results analysis: The same operation (increasing fan speed by 1%) resulted in significant differences in ammonia removal efficiency under different scenarios, ranging from -0.35 ppm / % (summer + brooding period) to -0.72 ppm / % (winter + fattening period), with an absolute difference of 106%. If a fixed control rule is adopted (such as "increasing fan speed by 2% for every 1 ppm increase in ammonia"), there may be over-ventilation in winter (leading to a sudden drop in temperature), and insufficient ventilation in summer (failure to meet ammonia control standards).

[0130] This invention uses dynamic edge weight assignment in equation (3) This allows the same operation to automatically apply the corresponding effect strength from Table 9 in different scenarios, improving the adaptability of the control strategy. For example, in the summer + brooding scenario, the system automatically reduces the expected ammonia removal effect of fan speed increase (side weight -0.35), and correspondingly increases the weight of opening the wet curtain or increasing the opening of the air inlet window, avoiding over-reliance on inefficient fan ventilation.

[0131] Example 8: Comparison of prediction accuracy of effect diagrams before and after online learning.

[0132] This embodiment is compared with Figure 5 The feedback update process shown is compared with the change in prediction accuracy before and after online learning of the effect map to verify the actual effect of the S7 feedback update mechanism.

[0133] On the first day after system deployment (December 1, 2025, the effect map was initialized only by aquaculture knowledge units and historical data, without any online updates) and the 30th day (December 30, 2025, after 28 online updates and 3 offline calibrations), prediction accuracy tests were conducted. Test method: For 48 consecutive hours, the deviation between the system's predicted ammonia concentration in the middle-lower layer over 30 minutes and the actual collected value after 30 minutes was recorded every 2 minutes. The test results are shown in Table 10. Table 10 Comparison of prediction accuracy before and after online learning Example of increased confidence level for the effect: "Increased fan speed + increased air inlet opening" higher-order interaction effect (in equation (4) The confidence level of the data was increased from the initial value of 0.60 (soft prior knowledge confidence level) to 0.88. The initial interaction effect value of -1.5ppm (expert experience estimate) was adjusted to -1.9ppm (statistical learning result based on actual operating data) after 28 online updates, which more accurately reflects the true synergistic effect of this combined operation in the target duck house.

[0134] The experimental data above demonstrate that the closed-loop learning mechanism of this invention can continuously improve the prediction accuracy of the effect map, enabling the system to continuously adapt to the specific structure and equipment performance characteristics of the duck house during long-term operation.

[0135] It should be noted that the above embodiments are merely preferred embodiments of the present invention, used to explain the technical principles and application effects of the present invention, and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions, modifications, or improvements made based on the technical concept of the present invention within the spirit and principles of the present invention should be included within the scope of protection of the present invention. In practical applications, those skilled in the art can make adaptive adjustments to the relevant parameters according to the specific test materials, working conditions, and accuracy requirements, all of which fall within the scope of protection of the present invention.

Claims

1. A method for intelligent ventilation control in duck houses based on dynamic regulation effect diagrams, characterized in that, Includes the following steps: S1 Data Acquisition and Knowledge Acquisition: Acquire breeding knowledge units, historical environmental data of duck houses, outdoor meteorological data, duck flock status data, historical equipment operation data, and response data before and after operation execution; the breeding knowledge units shall at least include applicable scenarios, control operations, expected effects, safety constraints, equipment interlock relationships, and operation sequence relationships; S2 Effect Graph Construction: A dynamic regulation effect graph is constructed based on the aquaculture knowledge units and historical operational data. The effect graph includes scenario nodes, state nodes, operation nodes, and response nodes, directed regulation effect edges connecting operation nodes and state or response nodes, and higher-order interaction relationships connecting at least two operation nodes and at least one state or response node. The directed regulation effect edges include at least the effect direction, effect intensity, response delay, effect duration, effect decay parameter, and applicable scenario. The higher-order interaction relationships are used to characterize the synergistic effect, offsetting effect, and temporal effect among multiple operations. S3 Baseline Prediction: Obtain the current environmental status of the duck house, the status of the duck flock, the outdoor weather status, and the equipment operation status. Under the condition of maintaining the current equipment operation status and the preset safe ventilation control unchanged, and without superimposing candidate incremental operations, predict the baseline environmental status in the future preset time domain. S4 Strategy Generation: Based on the current state, baseline environmental state, control objectives, and remaining equipment regulation capacity, generate multiple candidate combination control sequences and eliminate candidate combination control sequences that violate equipment capacity boundaries, equipment interlock relationships, aquaculture safety boundaries, or control change rate constraints. S5 Effect Prediction and Evaluation: Based on the dynamic regulation effect diagram, according to the response delay, duration and effect decay process of each operation, predict the incremental effect of each candidate combination control sequence in each future control cycle, superimpose the incremental effect onto the baseline environmental state, obtain the predicted state after intervention corresponding to each candidate combination control sequence, and calculate the comprehensive evaluation value. S6 Rolling Optimization and Execution: Under the constraints of minimum safe ventilation, wind speed in duck activity area, pressure difference in duck house, equipment control range and equipment change rate, determine the target control sequence with the optimal comprehensive evaluation value, execute only the control operation corresponding to the current control cycle in the target control sequence, and reacquire the state and solve the target control sequence at the next control time. S7 Feedback Update: Obtain the actual environmental response, duck flock response, and equipment response after control execution; calculate the deviation between the actual response and the predicted response; allocate the deviation to the corresponding directed control effect edge and higher-order interaction relationship based on the normalized counterfactual contribution; and perform a limited update on the learnable parameters; the updated parameters first enter the model to be verified, and then replace the current running model after passing the verification of prediction accuracy, safety constraints, and stability; the safety hard constraints, equipment physical boundaries, and emergency control rules do not change with the online learning process.

2. The method according to claim 1, characterized in that, In step S1, the historical data of the duck house includes one or more of the following: temperature, relative humidity, ammonia concentration, carbon dioxide concentration, wind speed, pressure difference, and bedding moisture content; the historical operating data of the equipment includes one or more of the following: fan start / stop status, fan speed, fan current, fan power, air inlet window opening, and air inlet direction; the response data includes one or more of the following: changes in environmental conditions, changes in duck flock behavior, changes in equipment load, and energy consumption.

3. The method according to claim 1, characterized in that, In step S1, the aquaculture knowledge unit further includes the initial environmental state, operation execution intensity, operation duration, potential side effects, and knowledge confidence. The aquaculture knowledge unit includes hard constraint knowledge and soft prior knowledge. The hard constraint knowledge is used to determine the aquaculture safety boundary, equipment capability boundary, equipment interlocking relationship, and direction locking effect edge that cannot be modified online. The soft prior knowledge is used to initialize the learnable parameters of the directed weighted edge and higher-order interaction relationship, and allows for limited updates based on the operating data.

4. The method according to claim 1, characterized in that, In step S2, the scenario nodes include at least one of the following: duck house season, breeding batch, duck age, weight, and breeding density; the status nodes include at least one of the following: duck house temperature, relative humidity, ammonia concentration, carbon dioxide concentration, wind speed, pressure difference, bedding moisture content, and duck behavior status; the operation nodes include at least one of the following: fan start / stop, fan speed adjustment, air inlet window opening adjustment, air inlet direction adjustment, intermittent ventilation, and zoned ventilation; and the response nodes include at least one of the following: environmental status change, duck behavior change, equipment load change, and energy consumption.

5. The method according to claim 1, characterized in that, In step S2, the current edge weight of the directed weighted edge is determined according to the following formula (3): ； in, For the current operation node To the status node or response node The right to the side; As a scene state, the scene nodes include at least one of the following: duck house season, breeding batch, duck age, weight, and breeding density; The overall status of the duck house includes at least one of the following: duck house temperature, relative humidity, ammonia concentration, carbon dioxide concentration, wind speed, pressure difference, bedding moisture content, and duck flock behavior. For the first The control operation includes at least one of the following: fan start / stop, fan speed adjustment, air inlet window opening adjustment, air inlet direction adjustment, intermittent ventilation, and zoned ventilation; the directed weighted edge includes at least a start node, an end node, an effect direction, an effect intensity, a response delay, an effect duration, and an applicable scenario; the higher-order interaction relationship includes at least the set of participating control operations, the operation execution order, the operation time interval, the interaction effect direction, the interaction effect intensity, and the interaction effect confidence.

6. The method according to claim 1, characterized in that, In step S2, the operation set The interaction effect is determined according to the following equation (4). ； in, For operation set The interaction effect vector, For operation set The actual state change vector after execution relative to the baseline without any new control intervention. For the first in the operation set The vector of the single-item regulatory effect of the operation; when When the direction of the single-item regulatory effect is consistent with the absolute value of the effect, it is determined to be a synergistic effect. When the overall regulatory effect is weakened or its direction is changed, it is determined to be an offsetting effect.

7. The method according to claim 1, characterized in that, In step S4, the duck house is divided into multiple environmental monitoring areas, with each area serving as a spatial node. Area connection edges are established based on the locational relationships and airflow connectivity between the areas. Based on environmental data, duck flock status, and equipment operation at continuous sampling times in each area, a spatiotemporal prediction model is used to obtain the baseline environmental status and duck flock risk prediction results for each area within a preset future time domain. The candidate combination control strategy is expressed as follows (8): ； in, For the first A candidate combination control strategy For operation type, For the strength of execution, To control the cycle number, For duration, The number of operations included in the candidate strategy; the candidate combination control strategy is generated by operation combination, execution intensity configuration and control cycle arrangement, and sequences that violate device interlock relationships or security constraints are eliminated.

8. The method according to claim 1, characterized in that, In step S5, the candidate combination control sequence The predicted incremental effect is obtained according to the following equation (11): ； in, Candidate combination control sequence In the The predicted incremental effect of each control cycle For the first Item operation in the The single-item regulatory effect of a control cycle For the first The higher-order interaction relationship in the first The interaction effect of each control cycle and These are the corresponding effect weights. The number of operations included in the candidate combination control sequence. The number of activated higher-order interaction relationships; the predicted incremental effect is superimposed on the baseline predicted state under the condition of no new control intervention according to the following formula (12): ； in, Candidate combination control sequence The corresponding predicted state after intervention, The baseline predicted state under no new control intervention conditions is obtained; the predicted state after intervention corresponding to the candidate combination control sequence is obtained; the comprehensive evaluation value is determined based on at least one of temperature deviation, humidity deviation, harmful gas risk, duck flock stress risk, equipment operating load and energy consumption, and the evaluation weights are dynamically adjusted according to the current breeding stage and scenario status.

9. The method according to claim 1, characterized in that, In step S6, the comprehensive evaluation value is determined according to the following formula (13): ； in, Candidate combination control sequence The overall evaluation value, To predict the time domain, For the price of temperature control, For the cost of humidity control, The cost of air quality risks associated with ammonia and carbon dioxide. The cost of heat stress, cold stress, or airflow stimulation for duck flocks This comes at the cost of equipment load and frequent start-stop cycles. As a result of energy consumption, to The evaluation weights are determined based on the breeding stage and the state of the scene; the safety constraints include at least the following formula (14): ； ； ； in, This is the current ventilation volume. The minimum safe ventilation volume is determined based on the number of ducks, their average weight, and air quality. Forecast wind speed in the area where ducks are active. This represents the maximum permissible wind speed for the current age of the duck flock. and These are the lower and upper limits of the control quantity for the corresponding equipment, respectively. For the first Each device at time The control quantity; the candidate combination control sequence with the best comprehensive evaluation value and that meets the safety constraints is determined as the target control sequence according to the following formula (15): ； in, For the target control sequence, A set of control strategies to meet the constraints of aquaculture safety boundaries and equipment capabilities; the rolling optimization includes comprehensively evaluating candidate strategies in the future time domain at the current control time to determine the target sequence, performing only the current cycle operation, and re-collecting the state and updating the solution at the next time.

10. A smart ventilation control system for duck houses based on a dynamic control effect diagram, characterized in that, include: The data acquisition module is used to collect environmental data from multiple areas of the duck house, outdoor weather data, duck flock status data, and equipment operation data. A dynamic regulation effect diagram construction module is used to construct and store the dynamic regulation effect diagram as described in any one of claims 1 to 9 based on aquaculture knowledge units and historical data; The baseline prediction module is used to predict the baseline environmental state in a preset time domain in the future based on the current state. The candidate strategy generation module is used to generate multiple candidate combination control sequences; The effect prediction and rolling optimization module is used to predict the incremental effect of each candidate sequence based on the dynamic control effect diagram, calculate the comprehensive evaluation value, and determine the target control sequence through rolling optimization under safety constraints. The device control module is used to convert the current cycle operation in the target sequence into device control commands and send them to the execution device; The feedback update module is used to update the learnable parameters in the dynamic regulation effect diagram based on the deviation between the actual response and the predicted response. The system is configured to perform the method according to any one of claims 1 to 9.

Images