Livestock breeding environment dynamic prediction and regulation method based on machine learning

Abstract

The invention belongs to the technical field of environment regulation and control, and discloses a livestock breeding environment dynamic prediction and regulation and control method based on machine learning. Comprising the following six key steps: firstly, preprocessing multi-source data, and constructing a standardized feature data set; secondly, performing time sequence feature extraction and environment-biological response correlation analysis to obtain a multi-factor interaction network; then, constructing a dynamic coupling model of the environmental parameters and the production efficiency; a multi-target reinforcement learning regulation and control model is constructed based on the animal husbandry production efficiency prediction space; predictive environmental parameter regulation and risk assessment are carried out in combination with real-time monitoring data; and finally, intelligent environment intervention is executed, and a regulation and control strategy is iteratively optimized through production index feedback. According to the method, the breeding environment is converted from passive response to active prediction, complex correlation understanding between environmental factors and biological indexes is established, and multi-target balance optimization of production efficiency, energy consumption and animal welfare is achieved.

Description

Technical Field

[0001] This invention relates to the field of environmental control technology, and more specifically, to a method for dynamic prediction and control of livestock farming environment based on machine learning. Background Technology

[0002] The livestock farming environment is a key factor affecting the healthy growth and production performance of livestock and poultry, playing a decisive role in farming efficiency and animal welfare. With the intensive and large-scale development of modern animal husbandry, environmental control technology has evolved from simple manual adjustment to automated control. Traditional environmental control originated from basic temperature and humidity control, mainly relying on simple ventilation and heating equipment. Automated environmental control systems have been gradually promoted in large-scale farms, achieving simple feedback control based on preset thresholds. In recent years, with the development of Internet of Things (IoT) technology, various environmental sensors have been widely used in farms, providing technical support for real-time monitoring of the farming environment.

[0003] In actual production, traditional livestock farming environmental control methods exhibit multiple technical bottlenecks. Data generated by numerous sensors within farms often exists in "data silos," with environmental parameters such as temperature, humidity, and gas concentration isolated from production data like animal weight and feed intake, failing to form an integrated information network. Farm managers frequently face a dilemma of data overload coupled with insufficient insight, relying on simplified linear models to make judgments in the face of environmental fluctuations, struggling to grasp the complex interactions between factors such as temperature, humidity, and airflow. For example, during hot summer months, managers often rely solely on increased ventilation to cool down, neglecting the potential for increased humidity and uneven airflow distribution to exacerbate heat stress. Farms generally use fixed environmental parameter settings, unable to dynamically adjust according to animal growth stages and individual differences, leading to energy waste and decreased growth performance. Traditional systems excessively pursue maximizing single production indicators (such as daily weight gain), neglecting the balance between energy consumption and animal welfare, resulting in a high-energy-consuming, high-stress production model. When extreme weather or disease outbreaks occur, existing technologies can only respond passively, unable to predict and implement preventative measures, causing significant losses to the livestock industry. More importantly, traditional control systems lack the ability to learn and iteratively optimize themselves, and cannot extract valuable control experience from historical data. This leads to the repeated occurrence of the same errors, resulting in a long-term stagnation in the management level of the breeding environment, which restricts the development of modern animal husbandry.

[0004] In view of this, the present invention proposes a machine learning-based method for dynamic prediction and regulation of livestock farming environment to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a method for dynamic prediction and control of livestock farming environment based on machine learning, comprising:

[0006] Step S1: Perform multi-source data preprocessing on multi-dimensional sensor data of the breeding environment, animal physiological status monitoring data, feed consumption data, historical yield data and meteorological data to obtain a standardized breeding environment characteristic dataset;

[0007] Step S2: Based on the standardized aquaculture environment feature dataset, perform time-series feature extraction and environment-biological response correlation analysis to obtain a multi-factor interaction network of the aquaculture environment;

[0008] Step S3: Based on the multi-factor interaction network of the breeding environment, construct a dynamic coupling model of environmental parameters and production efficiency to obtain the prediction space of livestock production efficiency;

[0009] Step S4: Based on the livestock production efficiency prediction space, construct a multi-objective reinforcement learning regulation model and optimize the strategy under constraints to obtain an adaptive environment regulation strategy library.

[0010] Step S5: Based on the adaptive environmental control strategy library, combined with the collected real-time environmental monitoring data, predictive environmental parameter control and risk assessment are carried out to obtain a dynamic control scheme for the aquaculture environment;

[0011] Step S6: Based on the aforementioned dynamic control scheme for the livestock breeding environment, implement intelligent environmental intervention and iterate and optimize the strategy through production indicator feedback to obtain the optimal control mode for the livestock breeding environment.

[0012] The technical effects and advantages of the machine learning-based method for dynamic prediction and control of livestock farming environment in this invention are as follows:

[0013] This invention integrates multi-source data and advanced algorithms to transform livestock farming environmental regulation from traditional, experience-based, and static management to intelligent, dynamic, and precise control. By establishing a complex network of relationships between environmental factors and biological indicators, it enables a deeper understanding of the impact mechanisms of different environmental conditions on livestock growth and development, thereby achieving environmental management that better meets biological needs. Predictive regulation capabilities allow farms to shift from passively responding to environmental changes to proactively preventing and intervening in advance, reducing the adverse effects of environmental fluctuations on production. Multi-objective balance optimization effectively reduces energy consumption and significantly improves animal welfare while increasing production efficiency, achieving a synergistic improvement in economic and ecological benefits. The adaptability of this invention allows it to flexibly adjust strategies according to different farming stages, climatic conditions, and production goals, greatly enhancing its ability to cope with complex and changing environments. Through continuous data collection and strategy iteration, the system can continuously improve itself, forming a virtuous cycle of optimization and achieving continuous improvement in regulation effectiveness. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a machine learning-based method for dynamic prediction and control of livestock farming environment according to the present invention.

[0015] Figure 2 This is a detailed flowchart illustrating the implementation steps of step S3 of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This application provides a machine learning-based method for dynamic prediction and control of livestock farming environment. The implementing entities of this machine learning-based method include, but are not limited to: environmental monitoring equipment, data processing servers, artificial intelligence computing platforms, sensor network nodes, and environmental control equipment, which can be considered as general computing nodes in this application. The data processing platform includes, but is not limited to: at least one of an environmental monitoring system, a livestock production management system, and a machine learning model training system.

[0018] Please see Figure 1 This invention provides a method for dynamic prediction and control of livestock farming environment based on machine learning, including the following steps:

[0019] Step S1: Perform multi-source data preprocessing on multi-dimensional sensor data of the breeding environment, animal physiological status monitoring data, feed consumption data, historical yield data and meteorological data to obtain a standardized breeding environment characteristic dataset;

[0020] Step S2: Based on the standardized aquaculture environment feature dataset, perform time-series feature extraction and environment-biological response correlation analysis to obtain the multi-factor interaction network of the aquaculture environment;

[0021] Step S3: Construct a dynamic coupling model of environmental parameters and production efficiency based on the multi-factor interaction network of the breeding environment to obtain the prediction space of livestock production efficiency;

[0022] Step S4: Based on the livestock production efficiency prediction space, construct a multi-objective reinforcement learning regulation model, and optimize the strategy under constraints to obtain an adaptive environmental regulation strategy library.

[0023] Step S5: Based on the adaptive environmental control strategy library, combined with the collected real-time environmental monitoring data, predictive environmental parameter control and risk assessment are carried out to obtain a dynamic control scheme for the aquaculture environment;

[0024] Step S6: Implement intelligent environmental intervention based on the dynamic control scheme of the breeding environment, and optimize the strategy iteratively through production indicator feedback to obtain the optimal control mode of the livestock breeding environment.

[0025] This invention preprocesses multi-source data, including multi-dimensional sensor data, animal physiological state monitoring data, feed consumption data, historical yield data, and meteorological data from the aquaculture environment, to establish a standardized aquaculture environment characteristic dataset. This dataset provides a reliable data foundation for subsequent analysis. Based on this standardized dataset, time-series feature extraction and environment-biological response correlation analysis reveal the complex relationships between environmental factors and biological indicators. The multi-factor interaction network of the aquaculture environment helps to understand the mutual influence between environmental parameters and their comprehensive effect on production efficiency. By constructing a dynamic coupling model between environmental parameters and production efficiency, the invention closely links changes in environmental state with biological responses. The livestock production efficiency prediction space supports the prediction of production performance under different environmental conditions. Based on a multi-objective reinforcement learning control model, the invention considers multiple objectives, including production efficiency, energy consumption, and animal welfare. The adaptive environmental control strategy library can select the most suitable control strategy according to different situations. Combining real-time environmental monitoring data for predictive environmental parameter control and risk assessment improves the foresight of environmental management. The dynamic control scheme for the aquaculture environment achieves precise adjustment of environmental parameters. Intelligent environmental intervention and strategy iterative optimization ensure continuous improvement of control strategies. The optimal control mode achieves high efficiency, low consumption, and sustainable development of livestock production.

[0026] In this embodiment of the invention, the detailed implementation steps of step S1 include:

[0027] Missing values ​​are detected in multidimensional sensor data to obtain data integrity assessment results. Then, a time-series interpolation algorithm is used to fill in the missing values ​​based on the data integrity assessment results to obtain complete sensor data.

[0028] The complete sensor data, animal physiological state monitoring data, feed consumption data, historical yield data and meteorological data were denoised to obtain a smoothed dataset. The data denoising process used the adaptive Kalman filter algorithm.

[0029] The smoothed dataset is subjected to time scale unification processing to obtain synchronous time series data, and the synchronous time series data is subjected to dimension normalization to obtain standardized time series data.

[0030] Outlier detection and processing are performed on standardized time series data to obtain anomaly-corrected data. Seasonal decomposition is then performed on the anomaly-corrected data to obtain trend, periodic and random components.

[0031] By integrating trend, periodic, and random components with the original standardized time-series data, a multi-level feature representation is constructed to obtain a standardized aquaculture environment feature dataset.

[0032] In this embodiment, multi-dimensional sensor data of the aquaculture environment are first collected, including environmental parameters such as temperature, humidity, light intensity, and gas concentration (e.g., ammonia, carbon dioxide). This ensures that the sensors are reasonably arranged to cover the entire aquaculture space. Missing value detection is performed on the collected multi-dimensional sensor data, and data quality assessment tools are used to check the data integrity, including calculating the missing rate for each sensor and analyzing missing pattern (e.g., random missing values, block missing values, or time-series missing values). Data integrity assessment results are generated, and based on these results, appropriate temporal interpolation algorithms are selected to fill in the missing values. For example, linear interpolation or spline interpolation can be used for short-term random missing values, while seasonal ARIMA models or machine learning methods (e.g., random forest) can be used for long-term or regular missing values. Predictive imputation is performed on the forest and K nearest neighbors to generate complete sensor data. Noise removal is then applied to the complete sensor data and other data sources (including animal physiological state monitoring data, feed consumption data, historical yield data, and meteorological data) using an adaptive Kalman filter algorithm. This algorithm automatically adjusts filtering parameters based on noise characteristics to adapt to different noise patterns from various data sources. Environmental parameters such as temperature and humidity are modeled using a state-space model. Random noise is filtered out through a prediction-update iterative process to generate a smoothed dataset. The smoothed dataset undergoes time-scale unification processing to address the issue of inconsistent sampling frequencies across different data sources; for example, sensor data may be sampled at the minute level, while physiological monitoring data may be sampled at the hourly or daily level. A unified time-scale framework is designed, and interpolation or downsampling methods are used to align data from different time scales to the same time scale, generating synchronous time-series data. This synchronous time-series data undergoes dimensional normalization to address the issue of different units and orders of magnitude for parameters such as temperature (°C), humidity (%), and ammonia concentration (ppm). Standardization methods (such as Z-score normalization or Min-Max normalization) are employed to transform each parameter to the same numerical range, ensuring comparability of different parameters in subsequent analyses, generating standardized time-series data. Outlier detection and processing are then performed on the standardized time-series data, using statistical methods (such as the 3σ rule and IQR method) or machine learning methods (such as isolated forest and single-class SVM) to identify anomalies. The system processes detected outliers, replacing them with the average, median, or predicted values ​​of neighboring values ​​to generate corrected outlier data. This corrected data is then subjected to seasonal decomposition, breaking down the time series into trend components (long-term trends), periodic components (daily or seasonal cycles), and random components (random fluctuations). Classical time series decomposition methods (such as STL decomposition and X-12-ARIMA) or modern methods (such as wavelet decomposition and empirical mode decomposition) are used for seasonal analysis to extract trend, periodic, and random components. These components are then integrated with the original standardized time series data to construct a multi-level feature representation that includes not only the original data but also its trend, periodic, and random fluctuation characteristics.This multi-level representation can more comprehensively describe the dynamic characteristics of the aquaculture environment, generating a standardized aquaculture environment feature dataset. This dataset provides a high-quality data foundation for subsequent time-series feature extraction and correlation analysis.

[0033] In this embodiment of the invention, the detailed implementation steps of step S2 include:

[0034] A sliding window segmentation method was applied to the standardized aquaculture environment feature dataset to obtain a multi-scale time window sequence. Statistical features and frequency domain features were then extracted from the multi-scale time window sequence to obtain a comprehensive time series feature set.

[0035] Principal component analysis was used to reduce the dimensionality of the comprehensive time series feature set to obtain a key feature subset. Based on the key feature subset, a Granger causality test model between environmental factors and biological indicators was established to obtain the causal relationship network.

[0036] The causal relationship network and the expert knowledge base were integrated and verified to obtain the preliminary structure of the environment-organism interaction. The Bayesian network learning algorithm was then applied to the preliminary structure of the environment-organism interaction to obtain the conditional probability distribution table.

[0037] A dynamic Bayesian network was constructed based on the conditional probability distribution table to obtain a model of the impact of time-varying environmental factors. Sensitivity analysis was performed on the model to obtain the ranking of key influencing factors.

[0038] By constructing a weighted directed graph structure using the ranking of key influencing factors and integrating temporal association rules, a multi-factor interaction network of the aquaculture environment is obtained.

[0039] In this embodiment, a sliding window segmentation technique is applied to the standardized aquaculture environment characteristic dataset. Time windows of different scales (such as hourly, daily, and weekly) are designed, and the time series is segmented by sliding these windows. The window size is flexibly set according to the analysis objectives and data characteristics. The window overlap rate is typically set to 30%-50% to capture temporal continuity features, generating a multi-scale time window sequence. Temporal features are extracted from this multi-scale time window sequence, including statistical features (such as mean, variance, skewness, kurtosis, quantiles, etc.) and frequency domain features (spectral features and energy distribution obtained through Fourier transform or wavelet transform). Complex features such as entropy values ​​(sample entropy, approximate entropy), Hurst exponent, etc., can also be extracted. To assess the complexity of time series data, all features are combined to form a comprehensive time series feature set. Principal component analysis (PCA) is then used to reduce the dimensionality of this feature set. The covariance matrix between features is calculated, and the correlation and redundancy among features are analyzed. Principal components are extracted, and their dimensions are determined based on the cumulative explained variance (typically selecting principal components that explain 80%-90% of the variance). The original features are projected onto the principal component space to obtain the dimensionality-reduced feature representation, generating a key feature subset. Based on this key feature subset, a Granger causality test model is established between environmental factors and biological indicators to analyze the causal relationship between environmental parameters (such as temperature, humidity, and gas concentration) and biological indicators (such as growth rate, feed conversion ratio, and physiological state). The relationship was investigated by performing Granger causality tests with different lag periods to determine which environmental factors significantly affected which biological indicators. A causal relationship network was constructed to represent the causal association between environmental factors and biological indicators. This network was then integrated and validated with knowledge provided by domain experts. Experts in animal husbandry, animal physiology, and other fields were invited to review and refine the causal relationship network. The network structure was adjusted based on scientific literature and practical experience. Statistical analysis results were cross-validated with expert knowledge to obtain a preliminary structure of the environment-biological interaction. Bayesian network learning algorithms were applied to this preliminary structure, and structure learning algorithms (such as K2 algorithm and MCMC algorithm) were used to optimize the network structure. Parametric learning algorithms were then used... (e.g., maximum likelihood estimation, Bayesian estimation) learns conditional probability distributions, considers the interactions between environmental factors and their combined effects on biological indicators, generates a conditional probability distribution table, representing the probability distribution of each node under the conditions of its parent node, constructs a dynamic Bayesian network based on the conditional probability distribution table, extends the static Bayesian network to the time domain, simulates the dynamic impact of environmental factors on biological indicators over time, introduces the transition probability in the time dimension, captures the dynamic characteristics of the environment-biological system, constructs a time-varying environmental factor impact model, performs sensitivity analysis on the time-varying environmental factor impact model, calculates the degree of influence of each environmental factor on biological indicators, and uses analysis of variance or sensitivity indices (such as the Sobol index) to quantify the intensity of the impact.

[0040] Environmental factors are ranked according to their impact intensity to obtain a ranking of key impact factors. A weighted directed graph structure is then constructed using this ranking, where nodes represent environmental factors and biological indicators, edges represent impact relationships, edge weights represent impact intensity, and direction represents the causal direction of the impact. Temporal association rules (such as frequent pattern mining and sequence association rules) are integrated to enrich the graph structure information. Regularities and patterns discovered in time series are incorporated to construct a multi-factor interaction network for the aquaculture environment. This network comprehensively describes the complex interaction relationships between environmental factors and between environmental factors and biological indicators, providing a structural foundation for the subsequent construction of dynamic coupling models.

[0041] In the embodiments of the present invention, see Figure 2 The diagram below illustrates the detailed implementation steps of step S3. In this embodiment of the invention, the detailed implementation steps of step S3 include:

[0042] The multi-factor interaction network of the aquaculture environment is transformed into a matrix representation to obtain the interaction coefficient matrix. Based on the interaction coefficient matrix, an environmental state transition equation is constructed to obtain an environmental dynamic system model.

[0043] Physiological parameters and corresponding environmental conditions of livestock and poultry at different growth stages were collected to obtain a set of stage-specific adaptive indicators. Based on the set of stage-specific adaptive indicators, a biological response function was established to obtain a multi-stage physiological response model.

[0044] By coupling and integrating the environmental dynamic system model with the multi-stage physiological response model, a two-way interaction model between the environment and organisms is obtained. Then, parameter sensitivity analysis is performed on the two-way interaction model between the environment and organisms to identify key regulatory nodes.

[0045] A deep neural network predictor is constructed based on key control nodes to obtain the mapping function from environmental parameters to production efficiency. The Monte Carlo method is then used to perform uncertainty analysis on the mapping function to obtain the prediction confidence interval.

[0046] By integrating the mapping function with the prediction confidence interval and adding seasonal and breeding cycle factors, a complete dynamic performance prediction model is constructed, resulting in a prediction space for livestock production performance.

[0047] In this embodiment, the multi-factor interaction network of the breeding environment is transformed into a matrix representation. An adjacency matrix is ​​constructed to represent network connectivity, and a weight matrix represents interaction strength. A Laplace matrix can also be constructed to analyze network structural characteristics. By combining these matrix representations, an interaction coefficient matrix is ​​constructed. This matrix quantifies the degree of mutual influence between various environmental factors. Based on the interaction coefficient matrix, an environmental state transition equation is constructed. A state-space model is used to describe the dynamic changes of environmental parameters over time. The state equation describes the internal state changes of the system, and the observation equation describes the relationship between the state and the observed values. Considering the influence of external inputs (such as control actions and weather changes) on the system, an environmental dynamic system model is constructed. This model can predict the trajectory of environmental parameter changes under specific conditions and control strategies. Data on livestock (such as pigs, chickens, cattle, etc.) at different growth stages (such as juvenile stage) are collected. Physiological parameter data (including weight, body temperature, heart rate, respiratory rate, and immune indicators) of livestock and poultry during growth, fattening, and egg / milk production periods are collected. Corresponding environmental condition data, such as temperature, humidity, and gas concentration, are also recorded. The adaptability and sensitivity differences of livestock and poultry to environmental conditions at different growth stages are analyzed, and a set of stage-specific adaptability indicators is constructed. Based on this set, a biological response function is established to describe the impact of environmental condition changes on biological indicators. Nonlinear functions (such as Logistic function, Gompertz function, etc.) or machine learning models (such as support vector regression, random forest regression, etc.) are used to fit the environment-biological response relationship. Considering the specific needs and response characteristics of different growth stages, a multi-stage physiological response model is constructed. This model can predict changes in the physiological state of livestock and poultry under specific environmental conditions. For example, the formula for the biological response function is:

[0048] Where G(t) represents a biological indicator (such as body weight), a is the asymptotic value (representing the maximum theoretical value that the biological indicator can reach over time), b is the displacement parameter, which determines the position of the growth curve on the time axis and reflects the developmental state of the animal at birth or in the initial stage; c is the growth rate parameter, which controls the growth rate and the steepness of the curve. The larger the value, the faster the organism grows and the earlier it approaches the asymptotic value; t is time.

[0049] This paper integrates an environmental dynamic system model with a multi-stage physiological response model, using the output of the environmental model as the input of the biological model. Changes in the biological state can, in turn, affect the environment (e.g., metabolic heat, exhaled gases). A bidirectional feedback mechanism is established, considering the co-evolutionary process of the environment-biological system. An environment-biological bidirectional interaction model is constructed, and parameter sensitivity analysis is performed. Local sensitivity analysis methods (e.g., partial derivative method) or global sensitivity analysis methods (e.g., Sobol method, FAST method) are used to evaluate the importance of model parameters, identifying the parameters and nodes with the greatest impact on system behavior. These parameters and nodes are key points for environmental regulation, resulting in key regulatory nodes. Based on these key regulatory nodes, a deep neural network predictor is constructed. Deep learning architectures (e.g., multilayer perceptron, recurrent neural network, convolutional neural network) are used to establish a mapping relationship from environmental parameters to production efficiency indicators (e.g., weight gain rate, egg production rate, feed conversion rate). Historical datasets containing environmental parameters and their corresponding data are used when training the network. The corresponding production efficiency records are used to train a mapping function from environmental parameters to production efficiency. The Monte Carlo method is employed to analyze the uncertainty of this mapping function. Multiple random samplings are used to simulate the uncertainties in the model input (such as measurement errors and environmental fluctuations), analyzing how these uncertainties propagate to the model output. The probability distribution, variance, and standard deviation of the predicted values ​​are calculated, and prediction confidence intervals are constructed to reflect the reliability and uncertainty range of the prediction results. The mapping function and prediction confidence intervals are integrated, and seasonal factors (such as environmental changes throughout the year) and breeding cycle factors (such as batch turnover and cleaning / disinfection) are considered to seasonally adjust and periodically correct the model, enhancing its adaptability to time patterns. A complete dynamic efficiency prediction model is constructed, which can predict production efficiency performance under different combinations of environmental conditions, generating a livestock production efficiency prediction space. This prediction space can be visualized as a multidimensional surface or heat map, intuitively displaying the complex relationship between environmental parameters and production efficiency, providing a decision-making basis for subsequent reinforcement learning regulation.

[0050] In this embodiment of the invention, the detailed implementation steps of step S4 include:

[0051] By defining the environmental regulation state space, action space, and reward function, a basic framework for reinforcement learning is obtained. The reward function comprehensively considers three dimensions: production efficiency, energy consumption, and animal welfare.

[0052] A deep Q-network model is constructed based on the basic framework of reinforcement learning to obtain a value assessment system. A dual network architecture is adopted to reduce the valuation bias and obtain a stable learning structure.

[0053] Safety constraints on environmental parameters and equipment operation constraints are formulated to obtain a set of constraints. This set of constraints is then incorporated into the strategy optimization process, and a constrained strategy gradient algorithm is designed to obtain a constrained optimization framework.

[0054] By using a constrained optimization framework, multiple rounds of strategy exploration and evaluation are conducted in the livestock production efficiency prediction space to obtain a Pareto optimal solution set. Based on the Pareto optimal solution set, a multi-scenario regulation strategy mapping is constructed to obtain the strategy selection mechanism.

[0055] Based on different breeding stages, different climate conditions and different production goals, the optimal strategy under the systematic organization strategy selection mechanism is obtained to obtain an adaptive environmental regulation strategy library.

[0056] In this embodiment, an environmental regulation state space is defined for the reinforcement learning problem, including a combination of environmental parameters (such as temperature, humidity, ventilation rate, etc.) and livestock physiological state indicators (such as body temperature, activity level, etc.). An action space is defined, including executable environmental regulation operations (such as adjusting the temperature setpoint, changing the ventilation rate, controlling the light intensity, etc.). A reward function is defined, comprehensively considering three dimensions: production efficiency (such as weight gain rate, egg / milk production), energy consumption (such as electricity, fuel consumption), and animal welfare (such as stress level, behavioral diversity). By weighted combination of indicators from these three dimensions, a comprehensive evaluation index is formed, and a basic reinforcement learning framework is constructed, including a state transition function, a reward calculation function, and a target function. A deep Q-network is then constructed based on this basic reinforcement learning framework. The DQN (Deep Q-function Network) model uses deep neural networks (such as multilayer perceptrons or convolutional neural networks) to approximate the Q-function, which represents the long-term expected reward of taking an action in a specific state. The network is trained to learn the optimal policy by minimizing temporal difference errors, constructing a value evaluation system capable of assessing the long-term value of different environmental control strategies. A dual-network architecture (TargetNetwork and CurrentNetwork) is employed to reduce the bias and variance of Q-value estimation. TargetNetwork parameters are periodically copied from CurrentNetwork and kept fixed for a period of time, reducing the risk of oscillations and divergence during training and forming a stable learning structure. Safety constraints are established for environmental parameters, such as... Temperature range (considering the comfort zone of livestock and poultry), humidity range (avoiding excessive dryness or humidity), and upper limits for gas concentration (such as ammonia and carbon dioxide) are used to formulate equipment operation constraints, such as limits on equipment start-up and shutdown frequency (avoiding frequent start-ups and shutdowns), upper limits for energy consumption, and equipment operating range. These safety and operational constraints are integrated into a constraint set, which is then incorporated into the strategy optimization process. Constrained strategy gradient algorithms, such as Constrained Policy Optimization (CPO), Lagrange methods, or penalty term methods, are designed to ensure that the generated strategy satisfies all constraints, avoiding unsafe or infeasible control strategies. A constraint optimization framework is constructed that can optimize strategy performance while satisfying constraints. This framework is then used to predict livestock production efficiency within the livestock production efficiency prediction space. Multi-round strategy exploration and evaluation are conducted, using Monte Carlo tree search or model-based reinforcement learning methods to sample and evaluate strategies in the policy space. The performance of strategies across multiple objective dimensions (production efficiency, energy consumption, animal welfare) is assessed, trade-offs between strategies are identified, and a set of strategies not comprehensively outperformed by any other strategy is selected, forming a Pareto optimal solution set. Based on this Pareto optimal solution set, a multi-scenario control strategy mapping is constructed. The most suitable strategy selection under different scenarios (such as extreme weather, disease outbreaks, energy price fluctuations, etc.) is analyzed, and a scenario-strategy mapping table is established. This enables the system to intelligently select the best strategy based on the current situation, forming a strategy selection mechanism. This mechanism is tailored to the specific needs of different breeding stages (such as juvenile stage, growth stage, egg / milk production stage, etc.).Environmental challenges under different climatic conditions (such as high temperatures in summer, low temperatures in winter, and seasonal transitions) and the priorities of different production objectives (such as maximizing output, optimizing efficiency, and minimizing costs) are systematically organized into an optimal strategy selection mechanism. This results in a strategy matrix covering various possible scenarios and demand combinations, constructing an adaptive environmental control strategy library. This library can automatically select the most suitable environmental control strategy based on the real-time status and needs of the farm, providing strategic support for subsequent predictive environmental parameter control.

[0057] In this embodiment of the invention, the detailed implementation steps of step S5 include:

[0058] Receive real-time environmental monitoring data stream, obtain the current environmental state vector, and retrieve the adaptive environmental control strategy library based on the current environmental state vector to obtain a candidate control strategy set;

[0059] By combining short-term meteorological forecast data and historical operation data of the farm, the changes in environmental parameters within the next time period T (e.g., 24-72 hours) are predicted to obtain environmental trend prediction results.

[0060] Based on the environmental trend prediction results, a prospective simulation evaluation of the candidate control strategy set is carried out to obtain the multi-strategy effectiveness prediction results. The multi-strategy effectiveness prediction results are then comprehensively ranked to obtain the optimal control sequence.

[0061] Risk assessment under extreme conditions is performed on the optimal control sequence to obtain an emergency response plan. The optimal control sequence and the emergency response plan are then integrated to obtain a complete control scheme.

[0062] The complete control plan is transformed into a set of equipment execution instructions and suggestions for human intervention, resulting in a dynamic control plan for the aquaculture environment.

[0063] In this embodiment, a real-time environmental monitoring data stream is received via a sensor network, including real-time measurements of environmental parameters such as temperature, humidity, gas concentration, and light intensity, as well as physiological states and behavioral indicators of livestock and poultry (such as body temperature and activity levels). The received data is preprocessed (denoising, anomaly detection, etc.) and features are extracted to construct a current environmental state vector. This vector comprehensively describes the current state of the breeding environment. Based on the current environmental state vector, an adaptive environmental control strategy library is retrieved. The nearest neighbor algorithm or similarity matching method is used to find the preset strategy that best matches the current state, or a strategy selection rule (such as a decision tree or rule engine) is used to select a suitable strategy, generating a candidate control strategy set. This set contains multiple potentially applicable control strategies. Short-term meteorological forecast data (such as temperature, humidity, wind speed, and rainfall) is obtained from meteorological departments or meteorological data service providers for the next 1-3 days. Historical operational data from the farm is also acquired, including environmental changes and equipment response data under similar past conditions. Combining meteorological forecasts and historical data, time series prediction models (such as ARIMA and LSTM) are used to predict changes in environmental parameters over the next 24-72 hours. The impacts of external meteorological changes, internal heat balance, and livestock metabolic activities on the environment are considered to generate environmental trend prediction results. These results describe the possible trends in environmental parameters in the short term. Based on the environmental trend prediction results, a prospective simulation evaluation is performed on each strategy in the candidate control strategy set. An environmental-biological model is used to evaluate these strategies. The coupled model simulates future environmental parameter changes and production efficiency under different strategies. Evaluation indicators include production efficiency (such as expected weight gain and egg production rate), energy consumption, and animal welfare indicators. It generates multi-strategy efficiency prediction results, which are then comprehensively ranked. Multi-objective optimization methods (such as weighted summation and analytic hierarchy process) are used to comprehensively consider various indicators, ranking them according to the farm's current priorities (such as efficiency priority or cost priority). The top-ranked strategies are selected to form the optimal control sequence, representing the optimal combination of control strategies for a future period starting from the current moment. The optimal control sequence undergoes risk assessment under extreme conditions, simulating extreme situations (such as equipment failure, extreme weather, and power outages). This study assesses the robustness and responsiveness of the strategy, identifies potential risks and system vulnerabilities, designs alternative solutions and emergency interventions for extreme situations, and formulates an emergency response plan. This plan includes risk identification, early warning indicators, and emergency response measures. The optimal control sequence is integrated with the emergency response plan to construct a complete control scheme encompassing conventional control strategies and abnormal situation response measures. This enhances the system's resilience and adaptability, ensuring the maintenance of a suitable aquaculture environment under various conditions. The complete control scheme is translated into specific equipment execution instruction sets, such as temperature setpoints for temperature control equipment, wind speed adjustments for ventilation systems, and lighting plans for lighting systems. Standardized control instructions are generated based on equipment control interfaces and communication protocols, while also generating suggestions for manual intervention.This guides aquaculture personnel on necessary manual operations and observations, leading to the development of a dynamic environmental control plan. This plan includes both automated control commands and suggestions for manual intervention, ensuring the accuracy and reliability of environmental control.

[0064] In this embodiment of the invention, the detailed implementation steps of step S6 include:

[0065] The dynamic control scheme of the aquaculture environment is decomposed into temperature control commands, humidity control commands, gas concentration control commands and light control commands to obtain a multi-dimensional environmental control command set;

[0066] The intelligent control system forwards multi-dimensional environmental control command sets to the corresponding execution devices to achieve automated environmental parameter adjustment, and records the device operating status and environmental parameter change curves during the execution process to obtain intervention execution data;

[0067] Real-time collection of livestock and poultry behavioral characteristics, physiological indicators and production data to obtain a biological response dataset, and comparison and analysis of the biological response dataset with production targets to obtain performance evaluation results;

[0068] The reward function of the reinforcement learning model is fine-tuned based on the performance evaluation results to obtain an optimized value evaluation system. The optimized value evaluation system is then used to update the policy network parameters to obtain an iterative optimization policy.

[0069] The iterative optimization strategy is fed back to the adaptive environmental control strategy library for updating, and after multiple rounds of iteration, a best practice experience library is formed to obtain the optimal control mode of the livestock breeding environment.

[0070] In this embodiment, the dynamic control scheme for the aquaculture environment is decomposed into specific control instructions for multiple subsystems. For the temperature control system, parameters such as temperature setpoint, heating / cooling rate, and temperature fluctuation range are generated. For the humidity control system, humidity setpoint and humidification / dehumidification control instructions are generated. For the gas concentration control system, ventilation frequency, air exchange volume, and air filtration parameters are generated. For the light control system, parameters such as light intensity, light duration, and spectral distribution are generated. These subsystem instructions are integrated into a multi-dimensional environmental control instruction set to ensure coordinated operation of each subsystem. The multi-dimensional environmental control instruction set is distributed to various execution devices through an intelligent control system (such as a PLC or IoT control platform). These execution devices include environmental control equipment such as heaters, air coolers, humidifiers, exhaust fans, and lighting equipment. A communication link (such as an industrial bus or wireless network) is established between the control center and the execution devices to achieve reliable transmission and execution confirmation of instructions, enabling precise automatic adjustment of environmental parameters. Simultaneously, data during the execution process is recorded, including equipment start-up data. The system collects real-time data on equipment operation status (stop status, operating parameters, energy consumption data, etc.) and environmental parameters (temperature, humidity, gas concentration, etc.) to form intervention execution data. This data records the entire process of environmental regulation in detail. Livestock behavior monitoring systems (such as video analysis and sound analysis) collect real-time data on livestock behavior characteristics, such as activity patterns, feeding behavior, and resting behavior. Physiological monitoring devices (such as body temperature sensors and heart rate monitors) collect real-time physiological indicators, such as body temperature, respiratory rate, and heart rate. Production management systems collect production data, such as weight gain, egg / milk production, and feed consumption. These data are integrated to form a biological response dataset, which reflects the livestock's response to environmental regulation. The biological response dataset is compared and analyzed with pre-set production targets to calculate the achievement rate of key performance indicators (KPIs), such as weight gain rate, egg production rate, and feed conversion rate. This assesses the impact of the regulation strategy on production efficiency and generates an efficiency evaluation result, which quantifies the actual effect of the current regulation strategy.

[0071] Based on the performance evaluation results, the reward function of the reinforcement learning model is fine-tuned. The weights of each indicator in the reward function are adjusted according to the actual production performance. For example, the proportion of production efficiency in the total reward is increased, or the balance between different production indicators is adjusted. The structure and parameters of the reward function are optimized so that it more accurately reflects the actual production goals, thus forming an optimized value evaluation system.

[0072] Specifically, the formula for fine-tuning the reward function is as follows:

[0073] Where Rnew and Rold are the reward functions before and after fine-tuning, respectively, and α is the learning rate. It is the gradient of policy performance with respect to the reward function.

[0074] The policy network parameters of the reinforcement learning model are updated using an optimized value assessment system and newly collected data. Policy gradient methods (such as REINFORCE, PPO, etc.) or value iteration methods (such as DQN, DDPG, etc.) are used to update the policy network, enhancing the model's adaptability and optimization capabilities to the actual environment. Iterative optimization strategies are generated and fed back to the adaptive environment control strategy library to update the policy parameters and selection rules in the strategy library. Poorly performing strategies are eliminated, while high-performing strategies are retained and strengthened. Through a continuous cycle of policy evaluation, optimization, and update, the quality and adaptability of the strategy library are continuously improved. After multiple rounds of iterative optimization, successful experiences and effective practices are summarized to form a best practice experience library, which includes optimal control modes for different livestock species, different growth stages, different seasons, and different production goals. These modes have been verified in practice and have stable and reliable effects, forming the optimal control mode for the livestock breeding environment. This mode can achieve a balance between maximizing production efficiency and minimizing energy consumption while ensuring animal welfare, representing the highest level of intelligent livestock breeding environment control.

[0075] In this embodiment of the invention, step S1 involves performing time-scale unification processing on the smoothed dataset to obtain synchronized time-series data, including:

[0076] Identify the original sampling frequency and recording timestamp of various data sources to obtain a data time characteristic table, and design the minimum common time granularity based on the data time characteristic table to obtain a standardized time scale;

[0077] The high-frequency sampled data is downsampled, and the sliding window averaging method is used to generate data points that conform to the standardized time scale, resulting in a downsampled dataset.

[0078] Spline interpolation is performed on the low-frequency sampled data to generate intermediate data points on a standardized time scale, thus obtaining the up-frequency dataset;

[0079] Gaussian process regression is applied to irregularly sampled data to reconstruct the time series data, resulting in a regularized dataset.

[0080] The down-frequency dataset, up-frequency dataset, and regularized dataset are aligned and integrated according to a standardized time scale to obtain synchronized time-series data.

[0081] In this embodiment, the original sampling frequency and time characteristics of various data sources are first comprehensively checked and recorded. For example, sensor data may be sampled at the second or minute level, animal physiological state data may be sampled at the hour level, feed consumption and production data may be recorded at the daily level, and meteorological data may be recorded at the hourly or three-hour level. Information such as the sampling interval, time span, and timestamp format for each type of data is recorded to generate a data time characteristic table. This table comprehensively describes the time attributes of each data source. Based on the data time characteristic table, considering the analysis requirements and computational resource constraints, a suitable minimum common time granularity is designed. This granularity should be fine enough to retain important temporal dynamic characteristics, but not so fine that it leads to data redundancy and computational burden. Typically, a minimum common time granularity of 1 / 2 is used. Choose a minute, hour, or more suitable time granularity to generate a standardized time scale, which serves as the baseline time axis for all data alignment. Downsample data with a sampling frequency higher than the standardized time scale, such as downsampling second-level sensor data to hourly levels. Use a sliding window averaging method to calculate the average of the original data within each standard time scale. The window size is set to the standard time granularity (e.g., 1 hour). Windows can overlap or not. Other statistics within the window, such as maximum, minimum, and standard deviation, can be calculated as needed to form a downsampled dataset with a time resolution consistent with the standardized time scale. Upsample data with a sampling frequency lower than the standardized time scale are then processed. Frequency processing, such as upscaling daily data to hourly levels, employs spline interpolation methods (e.g., cubic spline interpolation) to generate smooth interpolation curves between known data points. The spline curve values ​​are taken at each time point on a standardized time scale, ensuring the interpolation results maintain the overall data balance (e.g., daily total feed consumption remains constant), forming an up-frequency dataset. This dataset fills the gaps in low-frequency data on the standard time scale. For irregularly sampled data (e.g., data with variable intervals or significant missing data), Gaussian process regression is applied for time series reconstruction. Gaussian process regression can handle irregularly sampled data and provide estimates of prediction uncertainty. A suitable kernel function (e.g., RBF kernel, Matér kernel, etc.) is then used. (e.g., n-core) captures the temporal correlation of data, predicts data values ​​on a standardized time scale, and simultaneously obtains an assessment of the uncertainty of the prediction, forming a regularized dataset. This dataset transforms irregular data into regular time series. The down-frequency dataset, up-frequency dataset, and regularized dataset are aligned and integrated according to the standardized time scale to ensure that all data sources have corresponding data values ​​at each standard time point. A unified data table or matrix is ​​constructed, where rows represent time points and columns represent different variables or features. A final data quality check is performed to ensure that the integrated data is continuous and consistent in time, forming synchronous time series data. This data provides a temporally aligned foundation for subsequent feature extraction and modeling.

[0082] In this embodiment of the invention, step S2 involves using principal component analysis to reduce the dimensionality of the comprehensive time-series feature set, resulting in a key feature subset, including:

[0083] Calculate the feature correlation matrix of the comprehensive time series feature set to obtain the feature dependency graph, and perform feature clustering based on the feature dependency graph to obtain feature groups;

[0084] Calculate the variance contribution rate of features within each feature group to obtain the ranking of feature importance within the group, and select representative features from each group to obtain a preliminary feature set;

[0085] Apply principal component analysis to the preliminary feature set to calculate eigenvectors and eigenvalues, and obtain the principal component transformation matrix;

[0086] Based on the principal component transformation matrix, the original features are mapped to the principal component space to obtain the principal component scores. Then, the key principal components are selected according to the cumulative variance contribution rate to obtain the dimensionality-reduced feature representation.

[0087] By combining the dimensionality-reduced feature representation with key biological indicators selected by domain experts, a final feature set is constructed, resulting in a subset of key features.

[0088] In this embodiment, the correlation matrix between features in the comprehensive time-series feature set is first calculated. The degree of correlation between features is quantified using metrics such as Pearson correlation coefficient, Spearman rank correlation coefficient, or mutual information, generating an n×n correlation matrix, where n is the number of features. Each element in the matrix represents the degree of correlation between two features. The correlation matrix is ​​visualized as a feature dependency graph, and the strength of the correlation between features is intuitively displayed using heatmaps, network graphs, etc. High correlation indicates potential information redundancy between features. Feature clustering is then performed based on the feature dependency graph. Hierarchical clustering, K-means clustering, or spectral clustering methods are used to group similar features into the same group. Correlation is used as a similarity metric during clustering, and the clustering is adjusted accordingly. Parameters (such as the number of clusters) are used to obtain reasonable feature groupings, forming feature clusters. Each cluster contains highly correlated features. The variance contribution rate of each feature within each feature cluster is calculated to assess the explanatory power of each feature for the overall variation. The calculation method can be univariate ANOVA or the correlation between the feature and the target variable (such as a production efficiency indicator). Features within the clusters are ranked according to their variance contribution rates to identify the most representative features in each cluster, forming a feature importance ranking within the clusters. Representative features are selected from each feature cluster, usually the feature with the highest variance contribution rate in each cluster, or the feature most meaningful in terms of business understanding. This ensures that the selected feature set covers the main dimensions of the original feature space, forming a preliminary feature set. The feature set has already removed most redundant features. Principal Component Analysis (PCA) is applied to the preliminary feature set. First, the data is standardized to ensure consistency in the dimensions of each feature. The covariance matrix or correlation matrix is ​​calculated, and eigenvalue decomposition is performed to obtain eigenvectors and their corresponding eigenvalues. The eigenvectors represent the direction of the principal components, and the eigenvalues ​​represent the variance of the principal components. The eigenvectors are sorted by eigenvalue, and a principal component transformation matrix is ​​constructed. This matrix contains the sorted eigenvectors as column vectors. Based on the principal component transformation matrix, the preliminary feature set is mapped to the principal component space. The calculation formula is: Principal Component Score = Preliminary Feature Set × Principal Component Transformation Matrix. Each principal component is a linear combination of the original features. The variance of each principal component is calculated. The variance contribution rate and cumulative variance contribution rate are used to determine the proportion of total variance explained by the top k principal components. The variance contribution rate is calculated as the sum of eigenvalues ​​and the total variance. Key principal components are selected based on the cumulative variance contribution rate, typically choosing the top k components with a cumulative contribution rate of 85%-95%. Alternatively, a scree plot can be used to determine the number of principal components, forming a dimensionality-reduced feature representation. This representation retains the main information of the original data but significantly reduces the dimensionality. This dimensionality-reduced feature representation is then combined with key biological indicators selected by domain experts (such as animal husbandry or animal physiology experts). These key biological indicators may not be prominent in statistical analysis but have significant biological meaning, such as specific physiological indicators, behavioral characteristics, or production efficiency indicators.By integrating data-driven statistical analysis results with domain expertise, a final feature set was constructed, forming a key feature subset. This subset retains the main variation information of the data and includes biologically important indicators, providing a scientific foundation for subsequent establishment of causal relationship models between environmental factors and biological indicators.

[0089] In this embodiment of the invention, step S4 utilizes a constrained optimization framework to conduct multiple rounds of strategy exploration and evaluation within the livestock production efficiency prediction space to obtain a Pareto optimal solution set, including:

[0090] Design a multi-objective evaluation function to transform production efficiency, energy consumption rate and animal welfare index into scalar reward signals, and obtain a comprehensive evaluation mechanism;

[0091] A strategy exploration tree is constructed within the livestock production efficiency prediction space, and a Monte Carlo tree search algorithm is used to sample strategies to obtain a set of candidate strategies.

[0092] Based on a comprehensive evaluation mechanism, each strategy in the candidate strategy set is simulated and evaluated, and multi-objective performance indicators are calculated to obtain the strategy evaluation matrix.

[0093] Based on the policy evaluation matrix, non-dominated ranking is performed to select the set of policies that are not completely surpassed by any other policy, thus obtaining the Pareto front policy.

[0094] Cluster analysis is applied to Pareto front strategies to identify policy families with similar characteristics, and representative policies are selected from each policy family to obtain a simplified set of Pareto optimal solutions.

[0095] In this embodiment, a multi-objective evaluation function is first designed to quantify production efficiency indicators, including key production indicators such as daily weight gain rate, feed conversion rate, and egg / milk production; energy consumption rate indicators, including energy consumption per unit output, equipment operating power, and total energy cost; and animal welfare indices, including behavioral diversity scores, stress level indicators, and health status scores. A weighted summation method is designed to integrate these three types of indicators into a scalar reward signal. The weights can be flexibly adjusted according to the farm's priorities. Alternatively, other multi-objective to single-objective methods, such as Chebyshev weighted methods or goal programming, can be used to form a comprehensive evaluation mechanism. A strategy exploration tree is constructed within the livestock production efficiency prediction space, with the prediction space serving as the state space. Possible environmental parameter adjustment schemes are used as the action space. An initial policy tree structure is constructed, where nodes represent states, edges represent actions, and paths represent policies. Monte Carlo Tree Search (MCTS) algorithm is used for policy sampling. MCTS consists of four steps: selection (selecting the most valuable path starting from the root node), expansion (adding new nodes), simulation (randomly simulating from the new nodes to the final state), and backpropagation (updating node values). Appropriate exploration / exploitation balance parameters (such as UCB parameters) are set to ensure that both new policies are explored and known good policies are utilized. A candidate policy set is generated through multiple iterations. This set contains representative samples from the policy space. Each policy in the candidate policy set is simulated and evaluated based on a comprehensive evaluation mechanism. The strategy execution process is simulated using an environment-biology coupling model. Simulation results are collected, and performance metrics for each strategy across multiple objective dimensions are calculated: production efficiency metrics (e.g., weight gain rate, output), energy consumption metrics (e.g., total energy consumption, energy efficiency ratio), and animal welfare metrics (e.g., comfort score). This forms a strategy evaluation matrix, where rows represent different strategies and columns represent different performance metrics. Non-dominated ranking is performed based on the strategy evaluation matrix, defining dominance relationships as follows: if strategy A is not inferior to strategy B on all objectives and is superior to B on at least one objective, then A dominates B. A fast non-dominated ranking algorithm (such as the one used in NSGA-II) is used to identify all non-dominated strategies. These strategies constitute the Pareto front, meaning no strategy is superior to B without sacrificing its own performance. Pareto front strategies are a set of strategies that improve a particular objective while sacrificing other objectives. These non-dominated strategies are selected to form the Pareto front strategy set, which represents the optimal choice among different trade-offs among multiple objectives. Cluster analysis is applied to the Pareto front strategies using appropriate clustering algorithms (such as K-means, hierarchical clustering, etc.) to cluster the strategies in the objective space. The number of clusters can be set according to actual needs, such as selecting 5-10 representative strategy families. The characteristics of each cluster are analyzed, such as the "high output, low energy consumption" family, the "balanced" family, the "welfare priority" family, etc. A representative strategy is selected from each strategy family. This can be the strategy closest to the cluster center or the strategy with the best comprehensive index, ensuring that the selected strategy represents the characteristics of the family.This results in a concise Pareto optimal solution set, which includes representative strategies for different trade-offs, providing diverse yet streamlined options for selecting control strategies.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0097] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0098] In the description of this invention, it should be understood that the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0099] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0100] In the description of this invention, "several" means one or more, and "a large number" means two or more.

[0101] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0102] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0103] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for dynamic prediction and control of livestock farming environment based on machine learning, characterized in that, include: Step S1: Perform multi-source data preprocessing on multi-dimensional sensor data of the breeding environment, animal physiological status monitoring data, feed consumption data, historical yield data and meteorological data to obtain a standardized breeding environment characteristic dataset; Step S2: Based on the standardized aquaculture environment feature dataset, perform time-series feature extraction and environment-biological response correlation analysis to obtain a multi-factor interaction network of the aquaculture environment; Step S3: Based on the multi-factor interaction network of the breeding environment, construct a dynamic coupling model of environmental parameters and production efficiency to obtain the prediction space of livestock production efficiency; Step S4: Based on the livestock production efficiency prediction space, construct a multi-objective reinforcement learning regulation model and optimize the strategy under constraints to obtain an adaptive environment regulation strategy library. Step S5: Based on the adaptive environmental control strategy library, combined with the collected real-time environmental monitoring data, predictive environmental parameter control and risk assessment are carried out to obtain a dynamic control scheme for the aquaculture environment; Step S6: Based on the aforementioned dynamic control scheme for the livestock breeding environment, implement intelligent environmental intervention and iterate and optimize the strategy through production indicator feedback to obtain the optimal control mode for the livestock breeding environment.

2. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 1, characterized in that, The process involves multi-source data preprocessing of multi-dimensional sensor data of the aquaculture environment, animal physiological state monitoring data, feed consumption data, historical yield data, and meteorological data to obtain a standardized aquaculture environment characteristic dataset, including: Missing values ​​are detected in the multidimensional sensor data to obtain a data integrity assessment result. Then, a time-series interpolation algorithm is used to fill in the missing values ​​based on the data integrity assessment result to obtain complete sensor data. The complete sensor data, animal physiological state monitoring data, feed consumption data, historical yield data and meteorological data are subjected to data denoising processing to obtain a smoothed dataset. The data denoising processing adopts an adaptive Kalman filter algorithm. The smoothed dataset is subjected to time scale unification processing to obtain synchronous time series data, and the synchronous time series data is subjected to dimension normalization to obtain standardized time series data. Outlier detection and processing are performed on the standardized time series data to obtain anomaly-corrected data, and seasonal decomposition is performed on the anomaly-corrected data to obtain trend, periodic and random components. By integrating the aforementioned trends, periods, and random components with the original standardized time-series data, a multi-level feature representation is constructed to obtain a standardized aquaculture environment feature dataset.

3. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 2, characterized in that, The process of extracting time-series features and performing environment-biological response correlation analysis based on the standardized aquaculture environment feature dataset yields a multi-factor interaction network of the aquaculture environment, including: A sliding window segmentation is applied to the standardized aquaculture environment feature dataset to obtain a multi-scale time window sequence. Statistical features and frequency domain features are extracted from the multi-scale time window sequence to obtain a comprehensive time series feature set. Principal component analysis was used to reduce the dimensionality of the comprehensive time series feature set to obtain a key feature subset. Based on the key feature subset, a Granger causality test model between environmental factors and biological indicators was established to obtain a causal relationship network. The causal relationship network is integrated and verified with the expert knowledge base to obtain a preliminary structure of environment-biological interaction. A Bayesian network learning algorithm is then applied to the preliminary structure of environment-biological interaction to obtain a conditional probability distribution table. A dynamic Bayesian network is constructed based on the conditional probability distribution table to obtain a time-varying environmental factor influence model. Sensitivity analysis is then performed on the time-varying environmental factor influence model to obtain a ranking of key influencing factors. By ranking the key influencing factors, a weighted directed graph structure is constructed, and temporal association rules are integrated to obtain a multi-factor interaction network of the aquaculture environment.

4. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 3, characterized in that, The dynamic coupling model of environmental parameters and production efficiency is constructed based on the multi-factor interaction network of the breeding environment to obtain the prediction space of livestock production efficiency, including: The multi-factor interaction network of the aquaculture environment is transformed into a matrix representation to obtain the interaction coefficient matrix. Based on the interaction coefficient matrix, an environmental state transition equation is constructed to obtain an environmental dynamic system model. Physiological parameters and corresponding environmental conditions of livestock and poultry at different growth stages are collected to obtain a set of stage-specific adaptive indicators. Based on the set of stage-specific adaptive indicators, a biological response function is established to obtain a multi-stage physiological response model. The environmental dynamic system model is coupled and integrated with the multi-stage physiological response model to obtain an environment-biological two-way interaction model. Parameter sensitivity analysis is then performed on the environment-biological two-way interaction model to identify key regulatory nodes. A deep neural network predictor is constructed based on the key control nodes to obtain the mapping function from environmental parameters to production efficiency. The Monte Carlo method is then used to perform uncertainty analysis on the mapping function to obtain the prediction confidence interval. By integrating the mapping function with the prediction confidence interval and adding seasonal and breeding cycle factors, a complete dynamic efficiency prediction model is constructed to obtain the livestock production efficiency prediction space.

5. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 4, characterized in that, Based on the livestock production efficiency prediction space, a multi-objective reinforcement learning regulation model is constructed, and policy optimization under constraints is performed to obtain an adaptive environment regulation policy library, including: By defining the environmental regulation state space, action space, and reward function, a basic framework for reinforcement learning is obtained. The reward function comprehensively considers three dimensions: production efficiency, energy consumption, and animal welfare. Based on the aforementioned reinforcement learning framework, a deep Q-network model is constructed to obtain a value evaluation system. A dual network architecture is then used to reduce the evaluation bias, resulting in a stable learning structure. Safety constraints on environmental parameters and equipment operation constraints are formulated to obtain a set of constraints. The set of constraints is then incorporated into the strategy optimization process to design a constrained strategy gradient algorithm, thus obtaining a constrained optimization framework. The constrained optimization framework is used to conduct multiple rounds of strategy exploration and evaluation in the livestock production efficiency prediction space to obtain a Pareto optimal solution set. Based on the Pareto optimal solution set, a multi-scenario regulation strategy mapping is constructed to obtain a strategy selection mechanism. Based on different breeding stages, different climate conditions and different production goals, the optimal strategies under the aforementioned strategy selection mechanism are systematically organized to obtain an adaptive environmental regulation strategy library.

6. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 5, characterized in that, The method, based on the adaptive environmental control strategy library and combined with collected real-time environmental monitoring data, performs predictive environmental parameter control and risk assessment to obtain a dynamic control scheme for the aquaculture environment, including: Receive real-time environmental monitoring data stream, obtain the current environmental state vector, and retrieve the adaptive environmental control strategy library based on the current environmental state vector to obtain a candidate control strategy set; By combining short-term weather forecast data and historical operation data of the farm, the changes in environmental parameters within the next time period T are predicted, and the environmental trend prediction results are obtained. Based on the environmental trend prediction results, a prospective simulation evaluation is performed on the candidate control strategy set to obtain multi-strategy effectiveness prediction results. The multi-strategy effectiveness prediction results are then comprehensively ranked to obtain the optimal control sequence. A risk assessment under extreme conditions is performed on the optimal control sequence to obtain an emergency response plan, and the optimal control sequence is integrated with the emergency response plan to obtain a complete control scheme. The complete control scheme is transformed into a set of equipment execution instructions and suggestions for human intervention, resulting in a dynamic control scheme for the aquaculture environment.

7. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 6, characterized in that, The method of implementing intelligent environmental intervention based on the dynamic control scheme of the breeding environment, and iteratively optimizing the strategy through production indicator feedback to obtain the optimal control mode of the livestock breeding environment, includes: The dynamic control scheme for the aquaculture environment is decomposed into temperature control commands, humidity control commands, gas concentration control commands, and light control commands to obtain a multi-dimensional environmental control command set. The intelligent control system forwards the multi-dimensional environmental control instruction set to the corresponding execution device to realize automated environmental parameter adjustment, and records the device operating status and environmental parameter change curves during the execution process to obtain intervention execution data; Real-time collection of livestock and poultry behavioral characteristics, physiological indicators, and production data yields a biological response dataset. This dataset is then compared and analyzed with production targets to obtain performance evaluation results. Based on the performance evaluation results, the reward function of the reinforcement learning model is fine-tuned to obtain an optimized value evaluation system. The optimized value evaluation system is then used to update the policy network parameters to obtain an iterative optimization policy. The iterative optimization strategy is fed back to the adaptive environmental control strategy library for updating, and after multiple rounds of iteration, a best practice experience library is formed to obtain the optimal control mode of the livestock breeding environment.

8. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 2, characterized in that, The process of unifying the time scale of the smoothed dataset to obtain synchronized time-series data includes: Identify the original sampling frequency and recording timestamp of various data sources to obtain a data time characteristic table, and design the minimum common time granularity based on the data time characteristic table to obtain a standardized time scale; The high-frequency sampled data is downsampled, and the sliding window averaging method is used to generate data points that conform to the standardized time scale, resulting in a downsampled dataset. Spline interpolation is performed on the low-frequency sampled data to generate intermediate data points on a standardized time scale, thus obtaining the up-frequency dataset; Gaussian process regression is applied to irregularly sampled data to reconstruct the time series data, resulting in a regularized dataset. The down-frequency dataset, up-frequency dataset, and regularized dataset are aligned and integrated according to a standardized time scale to obtain synchronized time-series data.

9. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 3, characterized in that, The dimensionality reduction process performed on the comprehensive time-series feature set using principal component analysis yields a key feature subset, including: Calculate the feature correlation matrix of the comprehensive time series feature set to obtain the feature dependency graph, and perform feature clustering based on the feature dependency graph to obtain feature groups; Calculate the variance contribution rate of features within each feature group to obtain the ranking of feature importance within the group, and select representative features from each group to obtain a preliminary feature set; Applying principal component analysis to the preliminary feature set, we calculate the eigenvectors and eigenvalues ​​to obtain the principal component transformation matrix; Based on the principal component transformation matrix, the original features are mapped to the principal component space to obtain the principal component scores. Then, the key principal components are selected according to the cumulative variance contribution rate to obtain the dimensionality-reduced feature representation. The dimensionality-reduced feature representation is combined with key biological indicators selected by domain experts to construct the final feature set, thus obtaining the key feature subset.

10. The method for dynamic prediction and control of livestock farming environment based on machine learning according to claim 5, characterized in that, The constrained optimization framework is used to conduct multiple rounds of strategy exploration and evaluation within the livestock production efficiency prediction space to obtain a Pareto optimal solution set, including: Design a multi-objective evaluation function to transform production efficiency, energy consumption rate and animal welfare index into scalar reward signals, and obtain a comprehensive evaluation mechanism; A strategy exploration tree is constructed within the livestock production efficiency prediction space, and a Monte Carlo tree search algorithm is used to sample strategies to obtain a set of candidate strategies. Based on a comprehensive evaluation mechanism, each strategy in the candidate strategy set is simulated and evaluated, and a multi-objective performance index is calculated to obtain a strategy evaluation matrix. Based on the policy evaluation matrix, non-dominated ranking is performed to select a set of policies and obtain the Pareto front policies. Cluster analysis is applied to the Pareto front strategy to identify strategy families, and representative strategies are selected from each strategy family to obtain the Pareto optimal solution set.

Images