A Machine Learning-Based Optimization Control Method for High-Pressure Spray Deodorization in Poultry Farms

By optimizing the high-pressure spray deodorization system of poultry farms using machine learning models, the problems of poor deodorization effect and resource waste in traditional methods have been solved, achieving efficient and stable deodorization control.

CN122195173BActive Publication Date: 2026-11-17JIANGSU INST OF POULTRY SCI
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610302188.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-03-12
Publication Date
2026-11-17
Estimated Expiration
2046-03-12

AI Technical Summary

Technical Problem

Traditional high-pressure spray deodorization systems in poultry farms suffer from poor deodorization effects, low control precision, and high resource consumption. They also cannot adapt to the dynamic changes in farm conditions, leading to waste of water and electricity.

Method used

A machine learning-based high-pressure spray deodorization optimization control method is adopted. By acquiring sample spray deodorization data, a direct decision model and an environmental correction mapping model are trained. These models are then deployed on edge management terminals and high-pressure spray deodorization terminals to optimize spray parameters in real time. Combined with parameter importance analysis and a weighted scoring mechanism, the optimal deodorization effect is achieved.

Benefits of technology

It improves the accuracy and efficiency of deodorization, reduces operating costs, and ensures the stability of deodorization effects and the optimized use of resources, as well as adaptability and response speed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122195173B_ABST
    Figure CN122195173B_ABST
Patent Text Reader

Abstract

This invention discloses a machine learning-based optimized control method for high-pressure spray deodorization in poultry farms, relating to the field of agricultural intelligent control. The method includes: acquiring sample spray deodorization data; training a direct decision model and an environmental correction mapping model based on the sample data; deploying the direct decision model and the environmental correction mapping model to the target poultry farming scenario, wherein the direct decision model is deployed at the edge management terminal of the target poultry farming scenario, and the environmental correction mapping model is deployed at the high-pressure spray deodorization terminal of the target poultry farming scenario; and acquiring the scene-specific farming control parameters of the target poultry farming scenario in real time and transmitting them to the edge management terminal and the high-pressure spray deodorization terminal for optimized deodorization control. This method solves the problems of poor deodorization effect, low control precision, and high resource consumption in the high-pressure spray deodorization process of poultry farms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent control technology, specifically to a machine learning-based optimized control method for high-pressure spray deodorization in poultry farms. Background Technology

[0002] Poultry farms generate a lot of odor during the production process, mainly from the decomposition products of poultry feces, such as ammonia (NH3) and hydrogen sulfide (H2S). This not only affects the surrounding environment of the farm, but may also have a negative impact on the health of poultry and production efficiency.

[0003] Traditional high-pressure spray deodorization systems often employ fixed time intervals or control methods based on simple environmental thresholds, lacking a comprehensive consideration of the odor generation mechanism. Fixed-time spraying may lead to over-spraying during periods of low odor generation, wasting water and energy; while insufficient spraying may occur during periods of high odor generation, resulting in unsatisfactory deodorization. Furthermore, traditional methods cannot adapt to the dynamically changing conditions of farms, ultimately leading to poor deodorization performance, low control precision, and high resource consumption.

[0004] Therefore, there is an urgent need for an intelligent deodorization control method that can balance control precision, response speed, and resource efficiency. Summary of the Invention

[0005] This application provides a machine learning-based optimized control method for high-pressure spray deodorization in poultry farms, aiming to solve the problems of poor deodorization effect, low control accuracy, and high resource consumption in existing technologies.

[0006] In view of the above problems, this application provides a machine learning-based optimized control method for high-pressure spray deodorization in poultry farms.

[0007] This application provides a machine learning-based optimized control system for high-pressure spray deodorization in poultry farms, including:

[0008] Obtain sample spray deodorization data, wherein the sample spray deodorization data includes sample aquaculture control parameters and sample spray control parameters;

[0009] Train a direct decision-making model and an environment-corrected mapping model based on sample data;

[0010] The direct decision-making model and the environmental correction mapping model are deployed to the target aquaculture scenario, wherein the direct decision-making model is deployed at the edge management terminal of the target aquaculture scenario, and the environmental correction mapping model is deployed at the high-pressure spray deodorization terminal of the target aquaculture scenario;

[0011] The system acquires the scene-specific aquaculture control parameters of the target aquaculture scenario in real time and transmits them to the edge management terminal and the high-pressure spray deodorization terminal for deodorization optimization control.

[0012] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0013] This application firstly deploys edge devices and embedded terminals in a hierarchical model to adapt to different computing power requirements and improve system flexibility. Secondly, based on parameter importance analysis and a weighted scoring mechanism, it prioritizes the adjustment of key parameters to achieve the optimal deodorization effect with the lowest energy consumption, making the correction strategy precise and efficient. Thirdly, it reduces costs through lightweight deployment and reduces computing resource requirements and hardware costs through knowledge distillation and mathematical model solidification. Embedded model response improves response speed and timely responses to environmental changes. Finally, a multi-level constraint sample selection mechanism ensures training data quality, improves model stability, and enhances data quality controllability. This addresses the problems of poor adaptability, low control precision, and high resource consumption in existing deodorization control methods, providing a machine learning-based optimized control method for high-pressure spray deodorization in poultry farms. Through hierarchical deployment and lightweight model design, it dynamically optimizes spray parameters based on farming inputs and environmental conditions, thereby improving deodorization effect, reducing operating costs, maintaining stable deodorization effect, and optimizing resource utilization. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A schematic diagram of the process for optimizing the control of high-pressure spray deodorization in poultry farms based on machine learning;

[0016] Figure 2 This is a schematic diagram illustrating the process of training a direct decision model and an environmental correction mapping model based on sample data in the optimized control of high-pressure spray deodorization in poultry farms using machine learning. Detailed Implementation

[0017] This application provides a machine learning-based optimized control method for high-pressure spray deodorization in poultry farms, which addresses the problems of poor deodorization effect, low control precision, and high resource consumption in existing technologies.

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0019] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0020] Examples, such as Figure 1 As shown, this application provides a machine learning-based optimized control method for high-pressure spray deodorization in poultry farms, the method comprising:

[0021] S10: Obtain sample spray deodorization data, wherein the sample spray deodorization data includes sample aquaculture control parameters and sample spray control parameters;

[0022] Furthermore, the aquaculture control parameters should include at least the total number of animals raised, the amount of feed given per unit of animal, the temperature of the aquaculture environment, and the humidity of the aquaculture environment.

[0023] Spray control parameters should include at least spray duration and spray interval.

[0024] In this embodiment of the application, the sample spray deodorization data refers to the data set related to spray deodorization collected from historical operation records or experimental data, including input features and output labels; the sample aquaculture control parameters refer to aquaculture-related variables that affect the generation and diffusion of odors, which may include total aquaculture quantity, unit feed amount, environmental temperature and humidity, etc.; the sample spray control parameters refer to the operation variables of the spray system, including spray duration, spray interval, spray volume, etc.

[0025] Specifically, based on real-time monitoring data from local farms and other external data, and by standardizing data formats and cleaning outliers, sample spray deodorization data, including sample farming control parameters and sample spray control parameters, is finally obtained.

[0026] Step S10 in the method provided in this application embodiment includes:

[0027] Obtain the current breeding control parameters of the target breeding scenario, and analyze and obtain strong sample constraints and weak sample constraints;

[0028] Based on the strong sample constraints, a class of sample data for the target aquaculture scenario is obtained by combining the preset strong constraint tolerance, and the strong constraint satisfaction of the class of sample data is calculated accordingly.

[0029] Based on the weak sample constraints, the first type of sample data is screened a second time using a preset weak constraint tolerance to obtain the second type of sample data for the target aquaculture scenario, and the weak constraint satisfaction of the second type of sample data is calculated accordingly.

[0030] Based on the satisfaction of the weak constraints and the satisfaction of the strong constraints, the two types of sample data are iteratively probabilistically selected to obtain the sample spray deodorization data;

[0031] Wherein, the tolerance of the strong constraint is smaller than the tolerance of the weak constraint.

[0032] In this embodiment of the application, the current breeding control parameters of the target breeding scenario are first obtained, and strong sample constraints and weak sample constraints are analyzed and obtained.

[0033] Among them, the current breeding control parameters refer to the actual operating conditions of the current breeding farm, including the total number of breeding animals, the amount of feed per unit, the temperature of the breeding environment, and the humidity of the breeding environment; strong sample constraints refer to the conditions that must be strictly met, which may include poultry breeds and feed composition, with a small tolerance range; weak sample constraints refer to the conditions that can be flexibly met, which may include climate zones and temperatures, with a large tolerance range.

[0034] Specifically, based on the current poultry breed, feed composition, and environmental temperature and humidity among the current breeding control parameters of the target breeding scenario, the characteristics of the current breeding farm are analyzed, and standard values ​​for strong and weak constraints are defined respectively.

[0035] Among them, strong sample constraints use poultry breeds and feed composition as strong constraints; weak sample constraints use climate or administrative divisions and poultry house temperature as weak constraints.

[0036] Strong sample constraints allow for precise control of aquaculture parameters, reducing the impact of strong influencing factors on deodorization effectiveness. Weak sample constraints control environmental factors in aquaculture parameters, minimizing the influence of external environmental factors on deodorization effectiveness. By limiting these constraints, the effectiveness of parameter regulation can be guaranteed.

[0037] For example, strong sample constraints are: poultry breed: laying hens; feed composition: moisture content 40%, protein content 25%. Weak sample constraints are: temperate monsoon climate, North China, poultry house temperature 25℃-30℃.

[0038] Then, based on weak sample constraints and combined with a preset weak constraint tolerance, the first type of sample data is further filtered to obtain the second type of sample data for the target aquaculture scenario. The preset weak constraint tolerance is relative to the strong constraint.

[0039] Secondly, based on strong sample constraints, combined with the preset strong constraint tolerance, a class of sample data for the target aquaculture scenario is obtained, and the strong constraint satisfaction of the class of sample data is calculated accordingly.

[0040] Among them, constraint tolerance is the maximum range that the strong constraint parameter is allowed to deviate from the target value. It is usually set to be small to ensure the matching of core features; sample data is a candidate dataset after preliminary screening by strong constraints; strong constraint satisfaction is a quantification of the degree of matching between the sample and the target field under strong constraints.

[0041] Specifically, data that meets the above conditions is selected from the historical database to form a sample set. The strong constraint satisfaction level is calculated for each sample in the first class. The similarity between the sample data in the first class and the strong constraint is calculated, and this similarity is used as the strong constraint satisfaction level of the sample data in the first class.

[0042] The strong constraint satisfaction rate for class-I sample data is calculated as follows: 1 - |Class-I sample data - Standard value of strong sample constraints| / Strong sample tolerance. If any strong constraint is not satisfied, the sample is directly excluded. The higher the strong constraint satisfaction rate, the closer the data calculated from class-I sample data to the strong sample constraints is.

[0043] For example, the strong constraint must be strictly met, with a tolerance that the poultry breeds must be completely identical, the difference in feed composition must not exceed ±3%, and the satisfaction level is 0.85.

[0044] Secondly, based on weak sample constraints and combined with the preset weak constraint tolerance, the first type of sample data is screened a second time to obtain the second type of sample data of the target aquaculture scenario, and the weak constraint satisfaction of the second type of sample data is calculated accordingly.

[0045] The second type of sample data is a refined dataset after secondary screening using weak sample constraints. The weak constraint satisfaction rate assesses the degree of matching between samples and environmental conditions. The tolerance for weak constraints is usually set relatively loosely to introduce diversity; therefore, the weak constraint satisfaction rate may fluctuate significantly.

[0046] Specifically, based on the first type of sample, a second type of sample data with the same target aquaculture scenario is obtained through secondary screening according to the range of weak constraint tolerance. Then, the similarity between the second type of sample data and the weak sample constraints is calculated.

[0047] The weak constraint satisfaction is calculated using a weighted similarity method: Weak constraint satisfaction = w1 × climate similarity + w2 × administrative division similarity + w3 × (1 - |ΔT| / T) range Climate similarity is calculated based on the Köppen climate classification, and administrative division similarity is calculated based on geographical distance. Here, w1 is the weight of the climate factor, w2 is the weight of the administrative division factor, and w3 is the weight of the temperature factor, and w1 + w2 + w3 = 1. ΔT is the difference between the sample temperature and the target temperature, T... range This refers to the temperature range within the weak constraint tolerance.

[0048] For example, if the temperature similarity of two types of samples is 0.7 and the humidity similarity is 0.9, then the weak constraint satisfaction is 0.8.

[0049] Finally, based on the satisfaction of weak and strong constraints, iterative probability selection was performed on the two types of sample data to obtain sample spray deodorization data.

[0050] Among them, the iterative probability selection is based on the satisfaction degree to calculate the selection probability, and the final sample set is constructed through multiple random selections; the sample capacity constraint is the preset final sample quantity requirement to ensure that the amount of data is sufficient and not overloaded.

[0051] Specifically, the satisfaction rates of strong and weak constraints are first normalized, converting the satisfaction rate of weak constraints into a basic selection probability. Simultaneously, a corresponding reinforcement coefficient is set based on the strong sample constraints. Then, a probability correction coefficient is calculated using the strong constraint satisfaction rate and the reinforcement coefficient. This correction coefficient is used to adjust the basic probability to obtain the final selection probability. Finally, a random number generator iteratively selects samples according to the probability.

[0052] In addition, the tolerance for strong constraints is smaller than that for weak constraints.

[0053] Specifically, strong sample constraints have stricter requirements, while weak sample constraints have broader requirements. Therefore, the tolerance range of strong constraints will be smaller than that of weak constraints.

[0054] In step S10 of the method provided in this application embodiment, the strong sample constraints include at least the characteristics of the cultured species and the characteristics of the feed composition; the weak sample constraints include at least the regional climate characteristics and the temperature characteristics of the cultured environment.

[0055] In this embodiment of the application, the characteristics of the breeding species refer to the specific biological classification, growth stage and physiological characteristics of poultry, such as different breeds such as broiler chickens, laying hens and breeding chickens, which can directly determine the metabolic rate, fecal composition and ammonia production rate of poultry.

[0056] Feed composition characteristics include the proportions of key nutrients such as crude protein, amino acids, and phosphorus in the feed. These directly affect the digestibility and absorption efficiency of poultry and the chemical composition of feces, thus determining the amount of malodorous substances such as ammonia and hydrogen sulfide produced.

[0057] Regional climate characteristics refer to the macro-climate type and seasonal variation patterns of the location of the aquaculture farm, which may include parameters such as annual precipitation, monsoon patterns, and average temperature.

[0058] The temperature characteristics of the breeding environment refer to the actual temperature distribution inside the poultry house, which can affect the metabolic rate of poultry, microbial activity, and the volatilization rate of malodorous substances.

[0059] Specifically, for strong sample constraints, the sample selection process first requires that the farmed species in the sample data be completely consistent with the target farm; only samples containing data of that species are selected, excluding data of other species. Subsequently, the detailed formulation of the feed used by the target farm is analyzed, and tolerance ranges for key components are set. For weak sample constraints, historical climate data of the region where the target farm is located is collected. In sample selection, sample data from regions with the same or similar climate types are prioritized to obtain regional climate characteristics. A tolerance range for temperature matching is then set to obtain the temperature characteristics of the aquaculture environment.

[0060] In step S10 of the method provided in this application embodiment, the two types of sample data are iteratively probabilistically selected based on the weak constraint satisfaction degree and the strong constraint satisfaction degree to obtain the sample spray deodorization data, including:

[0061] Normalize the satisfaction of the weak constraints and define the basic selection probability of the two types of sample data;

[0062] The strong constraint satisfaction is normalized, and the probability correction coefficient of the basic selection probability is defined in combination with the preset strengthening coefficient.

[0063] Based on the probability correction coefficient, the corresponding traversal correction of the basic selection probability is performed to obtain the final selection probability set;

[0064] Based on the final selection probability set, a random number generator is used to iteratively select from the two types of sample data until the preset sample capacity constraint is met, and the iterative selection result is output as the sample spray deodorization data.

[0065] In this embodiment of the application, the weak constraint satisfaction is first normalized, and the basic selection probability of the two types of sample data is defined.

[0066] Normalization scales the original data proportionally to make it fall into a uniform numerical range, eliminating the influence of dimensions and facilitating comparison and calculation between different indicators. The basic selection probability is the initial selection probability assigned to each type of sample data based on the normalized weak constraint satisfaction. The higher the satisfaction, the greater the basic probability.

[0067] Specifically, the weak constraint satisfaction of each sample in the two types of sample data is normalized. Then, the normalized value is directly defined as the basic selection probability of that sample.

[0068] For example, the satisfaction of the weak constraints is 0.8, which is normalized to 0.67.

[0069] Secondly, the strong constraint satisfaction is normalized, and the probability correction coefficient of the basic selection probability is defined in combination with the preset reinforcement coefficient.

[0070] Among them, the reinforcement coefficient is a preset multiplier factor greater than 1, which is used to amplify the influence of strong constraint satisfaction in probability correction and reflect the emphasis on core matching elements; the probability correction coefficient is a coefficient calculated by combining the normalized strong constraint satisfaction and the reinforcement coefficient, and is used to adjust the basic selection probability by weight.

[0071] Specifically, the strong constraint satisfaction of each type of sample data is normalized. Then, the normalized strong constraint satisfaction is multiplied by a preset reinforcement coefficient to obtain the probability correction coefficient for that sample. The formula can be expressed as: Probability Correction Coefficient = Normalized Strong Constraint Satisfaction × Reinforcement Coefficient.

[0072] For example, if the normalized sample has a strong constraint satisfaction of 0.85 and the reinforcement coefficient is set to 1.3, then its probability correction coefficient is 0.85 × 1.3 ≈ 1.1.

[0073] Next, based on the probability correction coefficient, the corresponding traversal correction of the basic selection probability is performed to obtain the final selection probability set.

[0074] The traversal correction refers to adjusting the basic selection probability of each binary sample data in turn to take into account the influence of strong constraint satisfaction. The final selection probability set is the final probability set of each binary sample data after correction.

[0075] Specifically, the base selection probability of each binary sample is multiplied by the corresponding probability correction coefficient to obtain the final selection probability of that sample. The final selection probabilities of all binary sample data are then summed to form the final selection probability set.

[0076] For example, if the basic selection probability of a certain type II sample data is 0.85 and the probability correction coefficient is 1.1, then its final selection probability is 1.1 × 0.85 ≈ 0.94.

[0077] Finally, based on the final selection probability set, a random number generator is used to iteratively select from the two types of sample data until the preset sample capacity constraint is met, and the iterative selection result is output as the sample spray deodorization data.

[0078] Specifically, the random number generator generates a random number between 0 and 1, and compares this random number with the probability in the final selection probability set. If the random number is less than the final selection probability of a sample, then that sample is selected. This process is repeated continuously, and after each selection, the number of selected samples is checked to see if the preset sample capacity constraint has been met.

[0079] In this embodiment, multi-level constraints are used to screen high-quality sample data to ensure the representativeness and generalization ability of the model training data. By standardizing the data format and cleaning outliers, the consistency and reliability of the data are ensured. At the same time, sufficient learning samples are provided for model training, enabling the model to capture the complex relationship between aquaculture conditions and spray parameters.

[0080] S20: Train the direct decision-making model and the environment-corrected mapping model based on the sample data;

[0081] In this embodiment, the direct decision model is a lightweight machine learning model used to directly calculate the baseline spray parameters based on the aquaculture control parameters; the environmental correction mapping model is a lightweight mathematical model used to dynamically correct the baseline parameters based on environmental changes.

[0082] Specifically, a complex teacher model is trained using sample data, and then its knowledge is transferred to a lightweight direct decision-making model and an environment-corrected mapping model through knowledge distillation techniques.

[0083] Step S20 in the method provided in this application embodiment includes:

[0084] Based on the sample data, a full-environment decision-making model is trained and validated, and the model is compressed using a knowledge distillation method to obtain the direct decision-making model.

[0085] Based on the preset standard environment constraints, the aquaculture environment residuals corresponding to the sample data are calculated and obtained;

[0086] The sample data is subjected to non-standard screening based on preset standard environment constraints, and the non-standard screening results are dimensionally pruned to obtain pseudo-standard data.

[0087] Input the pseudo-standard condition data into the direct decision model, obtain the pseudo-standard condition control parameters, and calculate the residual between the pseudo-standard condition control parameters and the corresponding sample spray control parameters in the sample data to obtain the parameter correction residual;

[0088] By combining the residuals of the aquaculture environment with the residuals of the parameter correction, an environment correction mapping model based on a regression model is constructed.

[0089] In this embodiment, firstly, a full-environment decision-making model is trained and validated based on sample data. During training, aquaculture control parameters from the sample data are used as inputs, and spraying parameters are used as outputs. By continuously adjusting the model's weights and biases, the model output is made as close as possible to the actual spraying parameters. During validation, methods such as cross-validation are used to divide the sample data into training and validation sets, and the model's performance on the validation set is evaluated to ensure that the model has good generalization ability. Then, knowledge distillation is used to compress the model to obtain a direct decision-making model.

[0090] Furthermore, knowledge distillation involves transferring knowledge from a complex, full-environment decision-making model to a lightweight, direct-decision decision-making model, allowing the direct-decision model to learn the output distribution of the full-environment decision-making model. The output of the full-environment decision-making model is used as a soft label, which, together with the output of the direct-decision model, constructs a loss function. Optimizing this loss function trains the direct-decision model, thereby reducing its complexity and computational cost while maintaining performance.

[0091] For example, the steps for training and obtaining a full-environment decision model based on a neural network are as follows:

[0092] First, data preparation involves collecting aquaculture control parameters based on sample data.

[0093] Secondly, in model construction, the number of nodes in the input layer is equal to the dimension of the input features. For example, if there are four features, namely total breeding volume, unit feeding amount, breeding environment temperature and humidity, then the input layer contains four nodes. Set 1-3 hidden layers, and adjust the number of nodes in each layer through experiments, such as 64, 32, etc. The activation function is ReLU. The output layer generally does not use an activation function. For example, if the output takes 2 nodes, directly output continuous values.

[0094] Next, for model training, the spray parameters are used as output. The Adam optimizer and mean squared error loss function are used to build the training framework. The batch size is set to 32 and the total number of training rounds is 50. An early stopping mechanism (patience=5) is introduced. When the validation set loss does not decrease for 5 consecutive rounds, the training process is automatically terminated, and the trained full-environment decision model is obtained.

[0095] Then, based on preset standard environmental constraints, the aquaculture environment residuals corresponding to the sample data are calculated. The preset standard environmental constraints are specific environmental conditions such as temperature, humidity, and air pressure. For each sample data, the difference between its actual aquaculture environment parameters and the standard environmental constraints is calculated as the aquaculture environment residual. By analyzing the aquaculture environment residuals, the degree of environmental deviation of the sample data is understood, providing a basis for subsequent model construction.

[0096] Next, the sample data was subjected to non-standard screening based on preset standard environmental constraints, and the screening results were then dimensionally pruned to obtain pseudo-standard data. Non-standard screening refers to selecting data from the sample data that shows significant differences between actual environmental parameters and standard environmental constraints. Dimensional pruning removes redundant information from the aquaculture environment dimensions. Since the direct decision-making model is a standard-condition-oriented model, environmental dimension information is not considered a variable in the direct decision-making model. By removing dimensions with low correlation to spray parameters from the non-standard data, redundant information is reduced, improving data quality and model training efficiency. The data after screening and pruning is the pseudo-standard data.

[0097] The pseudo-standard condition data is then input into the direct decision model to obtain the pseudo-standard condition control parameters. The residuals between these pseudo-standard condition control parameters and the corresponding sample spray control parameters in the sample data are calculated to obtain the parameter correction residuals. The pseudo-standard condition control parameters are the spray control parameters calculated by the direct decision model based on the pseudo-standard condition data. By calculating the residuals between these parameters and the sample spray control parameters, we can understand the prediction error of the direct decision model under different environments, providing information for the construction of an environment correction mapping model.

[0098] Finally, by combining the residuals of the aquaculture environment and the parameter correction residuals, an environmental correction mapping model based on a regression model is constructed. The regression model can be linear regression, multinomial regression, etc. Using the residuals of the aquaculture environment as input and the parameter correction residuals as output, the mapping relationship between the two is found by training the regression model. When the actual aquaculture environment changes, the environmental correction mapping model is used to dynamically correct the baseline spray parameters obtained from the direct decision-making model to adapt to different environmental conditions and improve the deodorization effect of the spray.

[0099] In step S20 of the method provided in this application embodiment, the step of combining the knowledge distillation method to perform model compression and obtain the direct decision model includes:

[0100] Based on preset standard condition environment constraints, first standard condition sample data is obtained by filtering from the sample data;

[0101] The first standard condition sample data is input into the full environment decision model to obtain the second standard condition sample data, wherein the second standard condition sample data uses the residual vector with the first standard condition sample data as a soft label;

[0102] Based on the knowledge distillation method, the full-environment decision model is compressed by combining the first standard condition sample data and the second standard condition sample data to obtain the direct decision model.

[0103] In this embodiment, firstly, firstly, standard-condition sample data is obtained by filtering from the sample data according to preset standard-condition environmental constraints. The preset standard-condition environmental constraints are a series of pre-defined and representative environmental conditions, such as specific temperature ranges, humidity ranges, and light intensity. During the filtering process, the environmental parameters of each sample in the sample data are checked one by one. Only when the environmental parameters of a sample completely meet the preset standard-condition environmental constraints is it included in the first standard-condition sample data set. This ensures that the data used in subsequent model training and knowledge distillation processes are consistent and representative, avoiding interference with model performance due to differences in environmental factors.

[0104] Secondly, the first standard case sample data is input into the full-environment decision model to obtain the second standard case sample data. Upon receiving the first standard case sample data, the full-environment decision model processes and analyzes the input data, thereby outputting the corresponding result, namely the second standard case sample data. The second standard case sample data is labeled with the residual vector between the first and second standard case sample data. The residual vector represents the difference between the first and second standard case sample data output by the full-environment decision model, reflecting the error situation of the full-environment decision model when processing the first standard case sample data. Using the residual vector as a soft label allows the subsequent direct-response decision model to learn the output distribution of the full-environment decision model and its error characteristics when processing data, thereby better simulating the performance of the full-environment decision model.

[0105] Finally, based on the knowledge distillation method, the full-environment decision-making model is compressed by combining the first and second standard-case sample data to obtain the direct-decision decision model. Knowledge distillation is a technique for transferring knowledge from complex models to simpler models, such as transferring knowledge from the full-environment decision-making model to the direct-decision decision model. During the transfer process, the first standard-case sample data is used as input, and the soft labels of the second standard-case sample data are used as the target to construct a loss function. By optimizing the loss function, the output of the direct-decision decision model is made as close as possible to the output of the full-environment decision-making model. During the optimization process, the parameters of the direct-decision decision model are continuously adjusted so that it gradually learns the knowledge of the full-environment decision-making model. After multiple iterations and optimizations, a lightweight direct-decision decision model is finally obtained, which reduces complexity and computational cost while maintaining certain performance.

[0106] In step S20 of the method provided in this application embodiment, the construction of the environment correction mapping model further includes:

[0107] Define a modified reward function, wherein the modified reward function includes at least a regulation contribution factor, a regulation space factor, and a regulation cost factor;

[0108] Configure and modify selection rules, including:

[0109] The environment correction mapping model is executed N times consecutively to obtain N alternative parameter correction residuals, where N is greater than or equal to 3;

[0110] Based on the modified reward function, calculate the modified reward value of the modified residuals of the N candidate parameters;

[0111] The residual of the alternative parameter corresponding to the highest corrected reward value is selected as the output of the environmental correction mapping model.

[0112] In this embodiment, a modified reward function is first defined. The regulation contribution factor measures the degree of contribution to improving the deodorization effect of the spray. For example, the increase in deodorization efficiency after adjusting spray parameters is related to the regulation contribution factor. The regulation space factor reflects the adjustable range of spray parameters under the current aquaculture environment. If, under certain environmental conditions, the adjustable range of parameters such as spray pressure and flow rate is large, then the value of the regulation space factor will be correspondingly large. The regulation cost factor considers the costs incurred in adjusting spray parameters, including energy consumption, pesticide usage, etc. The modified reward function is defined by comprehensively considering the regulation contribution factor, regulation space factor, and regulation cost factor, which allows for a more comprehensive evaluation of the merits of different parameter correction residuals.

[0113] Secondly, the selection rules for the correction are configured. The environment correction mapping model is executed N times consecutively, where N is greater than or equal to 3. Each execution yields the correction residuals of the candidate parameters. This allows for the acquisition of multiple possible parameter correction schemes under different simulation conditions, increasing the diversity of choices. Then, based on the correction reward function, correction reward values ​​are calculated for each of the N candidate parameter correction residuals. By comprehensively considering factors such as the contribution of regulation, the regulation space, and the regulation cost, each candidate parameter correction residual corresponds to a correction reward value. A higher correction reward value indicates that the candidate parameter correction residual is better under the comprehensive evaluation.

[0114] Finally, the residual of the alternative parameter with the highest correction reward value is selected as the output of the environmental correction mapping model. Thus, among the possible parameter correction schemes, the scheme that best balances deodorization effect, adjustability, and cost is selected, enabling the spray deodorization system to operate more efficiently and economically in different aquaculture environments.

[0115] In this embodiment, a hierarchical model design is used to process the main aquaculture factors using a direct decision-making model and to focus on environmental fluctuations using an environmental correction mapping model, thereby improving the overall accuracy and adaptability of the system.

[0116] S30: Deploy the direct decision-making model and the environmental correction mapping model to the target aquaculture scenario, wherein the direct decision-making model is deployed at the edge management terminal of the target aquaculture scenario, and the environmental correction mapping model is deployed at the high-pressure spray deodorization terminal of the target aquaculture scenario;

[0117] In this embodiment, the edge management terminal refers to a computing device deployed locally on the farm, which has medium computing power and is used to integrate multi-source data and run a direct decision-making model; the high-pressure spray deodorization terminal refers to the embedded controller of the spraying equipment, which is directly linked with the actuator and runs an environment correction mapping model.

[0118] Specifically, models are deployed according to the target aquaculture scenario, including software integration and hardware configuration. The direct decision-making model is deployed on edge devices to receive aquaculture data via API interfaces; while the environmental correction mapping model is embedded in the embedded controller to collect real-time environmental data through sensors.

[0119] Furthermore, regarding software integration, for the direct decision-making model, ensure good compatibility with the operating system and data interfaces of the edge management terminal. Write adaptation code to encapsulate the model into callable functions or services so that the edge management terminal can smoothly call the model to calculate baseline spray parameters. Simultaneously, test the API interface to ensure the accuracy and stability of data transmission, preventing data loss or errors when receiving aquaculture data.

[0120] For the environmental correction mapping model, the resource constraints of the embedded system are considered when embedding it into the embedded controller. The model is optimized to reduce its memory footprint and computational load, ensuring that it can run efficiently on the embedded controller of the high-pressure spray deodorization terminal.

[0121] For high-pressure spray deodorization terminals, select a suitable embedded controller with the computational capabilities and interface functions that match the environmental correction mapping model. Connect high-precision sensors to the embedded controller to ensure accurate acquisition of real-time environmental data, such as temperature, humidity, and air pressure.

[0122] S40: Real-time acquisition of scene aquaculture control parameters of the target aquaculture scene, and transmission to the edge management terminal and the high-pressure spray deodorization terminal for deodorization optimization control.

[0123] In this embodiment of the application, the scene-based aquaculture control parameters refer to the dynamic data of the aquaculture farm collected in real time; the deodorization optimization control refers to automatically adjusting the operation of the spray system according to the spray parameters output by the model.

[0124] Specifically, in the target aquaculture scenario, a sensor network is used to collect aquaculture control parameters in real time. Sensors can collect data such as total number of animals raised, feed intake per unit area, ambient temperature, and humidity. The collected data is transmitted to edge management terminals and high-pressure spray deodorization terminals via wired or wireless communication methods.

[0125] After receiving the aquaculture control parameters for the scenario, the edge management terminal runs the direct decision-making model to calculate the baseline spray parameters based on the input aquaculture control parameters. These baseline spray parameters represent a preliminary spray control scheme considering the main aquaculture factors, providing a foundation for subsequent optimization.

[0126] Step S40 in the method provided in this application embodiment includes:

[0127] Based on the scenario-based aquaculture control parameters and preset standard environment constraints, the scenario-based aquaculture environment residuals are calculated.

[0128] The dimensionality of the aquaculture control parameters in the scenario is clipped, and the clipping result is input into the direct decision model of the edge management terminal to obtain the ideal deodorization control parameters.

[0129] The ideal deodorization control parameters and the residual of the scenario breeding environment are synchronously transmitted to the high-pressure spray deodorization terminal, and the residual of the scenario breeding environment is input to the environmental correction mapping model to obtain the correction amount of the deodorization control parameters;

[0130] By combining the ideal deodorization control parameters with the correction amount of the deodorization control parameters, deodorization optimization control is carried out for the target breeding scenario.

[0131] In this embodiment, firstly, the residual of the aquaculture environment is calculated based on the scenario-based aquaculture control parameters and preset standard environmental constraints. The preset standard environmental constraints include specific temperatures, humidity, and air pressures. For each sample in the real-time collected scenario-based aquaculture control parameters, the difference between its actual aquaculture environment parameters and the standard environmental constraints is calculated as the scenario-based aquaculture environment residual. By understanding the residual of the scenario-based aquaculture environment, the degree of environmental deviation of the current target aquaculture scenario is understood, providing a basis for subsequent deodorization optimization control.

[0132] Secondly, the control parameters for the aquaculture scenario are dimensionally pruned, and the pruning results are input into the direct decision-making model on the edge management terminal to obtain the ideal deodorization control parameters. Dimensional pruning removes dimensions in the aquaculture scenario control parameters that have low correlation with the spray parameters, reducing redundant information in the data. Since the direct decision-making model is a standard-condition decision-making model, environmental information is not treated as a variable in the direct decision-making model; therefore, dimensional pruning can improve data quality and model training efficiency. The pruned aquaculture scenario control parameters are input into the direct decision-making model, and the ideal deodorization control parameters are calculated based on the input aquaculture control parameters to obtain a preliminary spray control scheme.

[0133] Then, the ideal deodorization control parameters and the residuals of the scenario-based aquaculture environment are synchronously transmitted to the high-pressure spray deodorization terminal, and the residuals of the scenario-based aquaculture environment are input into the environmental correction mapping model to obtain the correction amount of the deodorization control parameters. After receiving the ideal deodorization control parameters and the residuals of the scenario-based aquaculture environment, the high-pressure spray deodorization terminal inputs the residuals of the scenario-based aquaculture environment into the environmental correction mapping model. The environmental correction mapping model calculates the correction amount of the deodorization control parameters based on the input residuals of the scenario-based aquaculture environment and the mapping relationship obtained through previous training, reflecting the extent to which the ideal deodorization control parameters need to be adjusted under the current aquaculture environment.

[0134] Finally, by combining the ideal deodorization control parameters with the correction values, optimized deodorization control is achieved for the target aquaculture scenario. The final spray control parameters are obtained by adding the ideal deodorization control parameters to the correction values. The high-pressure spray deodorization terminal automatically adjusts the operation of the spray system based on these final spray control parameters, such as spray pressure, flow rate, and time, thereby achieving optimized deodorization control for the target aquaculture scenario.

[0135] For example, suppose a poultry farm has real-time monitoring and control parameters showing a total of 5000 poultry, a feed intake of 150 grams per bird per day, an ambient temperature of 28°C, and an ambient humidity of 70%. The preset standard environment constraints are a temperature of 25°C and a humidity of 65%. Then, calculating the residuals of the ambient environment, the temperature residual is 28-25=3°C, and the humidity residual is 70-65=5%.

[0136] Then, the aquaculture control parameters are dimensionally pruned to remove dimensions with low correlation to the spray parameters. The pruning results are then input into the direct decision-making model of the edge management terminal. The direct decision-making model calculates the ideal deodorization control parameters based on the input aquaculture control parameters, such as a spray pressure of 3 MPa and a spray flow rate of 10 liters per minute.

[0137] Subsequently, the ideal deodorization control parameters and the residuals of the scenario-based aquaculture environment are synchronously transmitted to the high-pressure spray deodorization terminal. The high-pressure spray deodorization terminal inputs the residuals of the scenario-based aquaculture environment into the environmental correction mapping model. Based on the mapping relationship obtained from previous training, the environmental correction mapping model calculates the correction amount of the deodorization control parameters, such as the spray pressure needing to be increased by 0.2 MPa and the spray flow rate needing to be increased by 1 liter per minute.

[0138] Finally, by combining the ideal deodorization control parameters and the deodorization control parameter correction, the final spray control parameters are obtained. The spray pressure becomes 3 + 0.2 = 3.2 MPa, and the spray flow rate becomes 10 + 1 = 11 liters per minute.

[0139] In summary, compared with existing technologies, this application dynamically adjusts spray parameters based on real-time changes in the aquaculture environment, thereby improving the deodorization effect of the spray and reducing costs, enabling the spray deodorization system to operate more efficiently and economically in different aquaculture environments.

[0140] Among them, a direct decision-making model and an environment-corrected mapping model are trained based on sample data, such as... Figure 2 As shown, it includes:

[0141] Based on the sample data, a full-environment decision-making model is trained and validated, and the model is compressed using a knowledge distillation method to obtain the direct decision-making model.

[0142] Based on the preset standard environment constraints, the aquaculture environment residuals corresponding to the sample data are calculated and obtained;

[0143] The sample data is subjected to non-standard screening based on preset standard environment constraints, and the non-standard screening results are dimensionally pruned to obtain pseudo-standard data.

[0144] Input the pseudo-standard condition data into the direct decision model, obtain the pseudo-standard condition control parameters, and calculate the residual between the pseudo-standard condition control parameters and the corresponding sample spray control parameters in the sample data to obtain the parameter correction residual;

[0145] By combining the residuals of the aquaculture environment with the residuals of the parameter correction, an environment correction mapping model based on a regression model is constructed.

[0146] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects:

[0147] In this embodiment, the application first deploys edge devices and embedded terminals in a hierarchical manner through model layering to adapt to different computing power requirements and improve system flexibility. Secondly, based on parameter importance analysis and a weighted scoring mechanism, key parameters are prioritized for adjustment to achieve optimal deodorization with minimal energy consumption, making the correction strategy precise and efficient. Thirdly, lightweight deployment reduces costs, and knowledge distillation and mathematical model solidification reduce computing resource requirements and hardware costs. Embedded model response improves response speed and timely responses to environmental changes. Finally, a multi-level constraint sample selection mechanism ensures training data quality, improves model stability, and enhances data quality controllability. This addresses the problems of poor adaptability, low control precision, and high resource consumption in existing deodorization control methods, providing a machine learning-based high-pressure spray deodorization optimization control method for poultry farms. Through hierarchical deployment and lightweight model design, spray parameters are dynamically optimized based on farming inputs and environmental conditions, thereby improving deodorization effectiveness, reducing operating costs, maintaining stable deodorization performance, and optimizing resource utilization.

[0148] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0149] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A machine learning-based optimized control method for high-pressure spray deodorization in poultry farms, characterized in that, include: Obtain sample spray deodorization data, wherein the sample spray deodorization data includes sample aquaculture control parameters and sample spray control parameters; Train a direct decision-making model and an environment-corrected mapping model based on sample data; The direct decision-making model and the environmental correction mapping model are deployed to the target aquaculture scenario, wherein the direct decision-making model is deployed at the edge management terminal of the target aquaculture scenario, and the environmental correction mapping model is deployed at the high-pressure spray deodorization terminal of the target aquaculture scenario; The scene aquaculture control parameters of the target aquaculture scene are acquired in real time and transmitted to the edge management terminal and the high-pressure spray deodorization terminal for deodorization optimization control; Among them, training the direct decision-making model and the environment-corrected mapping model based on sample data includes: Based on sample data, a full-environment decision-making model is trained and validated, and the model is compressed using a knowledge distillation method to obtain the direct decision-making model. Based on the preset standard environment constraints, calculate and obtain the aquaculture environment residuals corresponding to the sample data; The sample data is subjected to non-standard screening based on the preset standard environment constraints, and the non-standard screening results are dimensionally pruned to obtain pseudo-standard data. Input the pseudo-standard condition data into the direct decision model to obtain the pseudo-standard condition control parameters, and calculate the residual between the pseudo-standard condition control parameters and the corresponding sample spray control parameters in the sample data to obtain the parameter correction residual; By combining the aquaculture environment residuals and parameter correction residuals, an environment correction mapping model based on a regression model is constructed. The process includes real-time acquisition of scene-specific aquaculture control parameters for the target aquaculture scenario and transmission of these parameters to the edge management terminal and the high-pressure spray deodorization terminal for deodorization optimization control, including: Based on the scenario-based aquaculture control parameters and preset standard environment constraints, the scenario-based aquaculture environment residuals are calculated. The dimensionality of the aquaculture control parameters in the scenario is clipped, and the clipping result is input into the direct decision model of the edge management terminal to obtain the ideal deodorization control parameters. The ideal deodorization control parameters and the residual of the scenario breeding environment are synchronously transmitted to the high-pressure spray deodorization terminal, and the residual of the scenario breeding environment is input to the environmental correction mapping model to obtain the correction amount of the deodorization control parameters; By combining the ideal deodorization control parameters with the correction amount of the deodorization control parameters, deodorization optimization control is carried out for the target breeding scenario.

2. The optimized control method for high-pressure spray deodorization in poultry farms based on machine learning as described in claim 1, characterized in that, Obtain sample spray deodorization data, wherein the sample spray deodorization data includes sample aquaculture control parameters and sample spray control parameters, including: Obtain the current breeding control parameters of the target breeding scenario, and analyze and obtain strong sample constraints and weak sample constraints; Based on the strong sample constraints, a class of sample data for the target aquaculture scenario is obtained by combining the preset strong constraint tolerance, and the strong constraint satisfaction of the class of sample data is calculated accordingly. Based on the weak sample constraints, the first type of sample data is screened a second time using a preset weak constraint tolerance to obtain the second type of sample data for the target aquaculture scenario, and the weak constraint satisfaction of the second type of sample data is calculated accordingly. Based on the satisfaction of the weak constraints and the satisfaction of the strong constraints, the two types of sample data are iteratively probabilistically selected to obtain the sample spray deodorization data; Wherein, the tolerance of the strong constraint is smaller than the tolerance of the weak constraint.

3. The optimized control method for high-pressure spray deodorization in poultry farms based on machine learning as described in claim 2, characterized in that, The strong sample constraints include at least the characteristics of the cultured species and the characteristics of the feed composition; the weak sample constraints include at least the regional climate characteristics and the temperature characteristics of the culture environment.

4. The optimized control method for high-pressure spray deodorization in poultry farms based on machine learning as described in claim 2, characterized in that, Based on the satisfaction levels of the weak and strong constraints, iterative probability selection is performed on the two types of sample data to obtain the sample spray deodorization data, including: Normalize the satisfaction of the weak constraints and define the basic selection probability of the two types of sample data; The strong constraint satisfaction is normalized, and the probability correction coefficient of the basic selection probability is defined in combination with the preset strengthening coefficient. Based on the probability correction coefficient, the corresponding traversal correction of the basic selection probability is performed to obtain the final selection probability set; Based on the final selection probability set, a random number generator is used to iteratively select from the two types of sample data until the preset sample capacity constraint is met, and the iterative selection result is output as the sample spray deodorization data.

5. The optimized control method for high-pressure spray deodorization in poultry farms based on machine learning as described in claim 1, characterized in that, The direct decision model is obtained by combining knowledge distillation methods with model compression, including: Based on the preset standard condition environment constraints, the first standard condition sample data is obtained by filtering from the sample data; The first standard condition sample data is input into the full environment decision model to obtain the second standard condition sample data, wherein the second standard condition sample data uses the residual vector with the first standard condition sample data as a soft label; Based on the knowledge distillation method, the full-environment decision model is compressed by combining the first standard condition sample data and the second standard condition sample data to obtain the direct decision model.

6. The optimized control method for high-pressure spray deodorization in poultry farms based on machine learning as described in claim 1, characterized in that, The construction of the environment correction mapping model also includes: Define a modified reward function, wherein the modified reward function includes at least a regulation contribution factor, a regulation space factor, and a regulation cost factor; Configure and modify selection rules, including: The environment correction mapping model is executed N times consecutively to obtain N alternative parameter correction residuals, where N is greater than or equal to 3; Based on the modified reward function, calculate the modified reward value of the residuals of N alternative parameters; The residual corresponding to the candidate parameter with the highest correction reward value is selected as the output of the environment correction mapping model.

7. The optimized control method for high-pressure spray deodorization in poultry farms based on machine learning as described in claim 1, characterized in that, The breeding control parameters should include at least the total number of animals raised, the amount of feed per unit, the temperature of the breeding environment, and the humidity of the breeding environment.

8. The optimized control method for high-pressure spray deodorization in poultry farms based on machine learning as described in claim 1, characterized in that, Spray control parameters should include at least spray duration and spray interval.

Citation Information

Patent Citations

  • Waste gas deodorization system and method

    CN119345886A

  • Malodorous odor and organic waste gas treatment method for enhancing absorption through pretreatment

    CN120155035A