Livestock breeding risk intelligent assessment method and system based on multi-source data fusion

Abstract

The invention provides a livestock breeding risk intelligent assessment method and system based on multi-source data fusion, and the method comprises the steps: collecting livestock individual vital sign data, breeding environment parameters, management behavior data and risk-related historical data through Internet of Things equipment, and carrying out the data preprocessing to form a standardized multi-source data set; extracting risk features of individual, group and environment levels based on the data set, and fusing the risk features to form a multi-dimensional risk feature library; utilizing machine learning to construct a differentiated risk assessment model; analyzing the incidence relation between the risk factor and the actual event through the Bayesian network to calibrate the model; realizing livestock risk grade dynamic division and early warning based on the calibrated risk scoring system; and finally, generating intervention suggestions for risk quantitative evaluation, risk prevention and control decision and loss evaluation. According to the method, accurate evaluation of livestock breeding risks is realized, decision support is provided for breeding safety management, and the method has relatively high application value.

Description

Technical Field

[0001] The present invention relates to the field of smart agricultural technology, and in particular to a method and system for intelligent livestock breeding risk assessment based on multi-source data fusion. Background Art

[0002] Intelligent risk assessment technology for animal husbandry belongs to the technical field that combines smart agriculture and risk management technology. It aims to accurately assess various risks in the animal husbandry process through information and intelligent means, and provide a scientific basis for risk pricing, claims decision-making and risk prevention and control.

[0003] Traditional livestock farming risk assessments rely primarily on manual inspections and empirical judgment. For example, relevant agencies dispatch surveyors to regularly check livestock health and environmental conditions at farms, or develop regional risk rate tables based on historical statistical data. Another common approach is based on simple, single-data-source monitoring systems, such as using only livestock body temperature or ambient temperature for early warning, lacking a comprehensive analysis of multi-dimensional risk factors.

[0004] With the development of the Internet of Things and artificial intelligence (AI), livestock risk monitoring technology based on sensor networks and data analysis has emerged. This technology deploys multiple sensors within the farming environment to collect real-time data on livestock vital signs and environmental parameters. It then uses simple thresholds or basic statistical models to identify risks. However, this technology still suffers from issues such as insufficient data dimensionality, overly simplistic risk assessment models, and insufficient correlation analysis of risk factors. This makes it difficult to accurately assess livestock risks in complex scenarios and fails to effectively support the refined requirements of risk pricing and underwriting decisions.

[0005] The main problems faced by existing technologies include: (1) a single data source and a lack of comprehensive analysis of multi-source data such as livestock vital signs, environmental parameters, and management behaviors; (2) a simple risk assessment model, mainly based on fixed threshold judgments, which cannot adapt to the characteristics of livestock of different breeds and different growth stages; (3) a lack of in-depth analysis of the correlation between risk factors, making it difficult to discover potential complex risks; and (4) risk assessment results are not closely connected with risk management business processes, making it difficult to directly support risk pricing and underwriting decisions. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent livestock breeding risk assessment method and system based on multi-source data fusion to solve technical problems in the existing technology such as insufficient data dimensions, overly simple risk assessment models, insufficient analysis of risk factor correlations, and loose connection between risk assessment results and business processes.

[0007] To achieve the above objectives, the present invention provides the following technical solutions:

[0008] An intelligent assessment method for livestock breeding risks based on multi-source data fusion, including:

[0009] Obtain livestock individual vital signs data, breeding environment parameters, management behavior data, and risk-related historical data collected by multiple IoT devices. Through data cleaning, outlier processing, and feature extraction, a standardized multi-source data set is obtained.

[0010] Based on the standardized multi-source data set, individual-level health risk characteristics are extracted through time series analysis, group-level disease risk characteristics are extracted through cluster analysis, and environmental-level risk characteristics are extracted through correlation analysis. A multidimensional risk feature library is formed using a feature fusion algorithm.

[0011] Based on the multidimensional risk signature database, a machine learning method is used to construct differentiated risk assessment models for different livestock breeds and different growth stages, and generate model assessment results;

[0012] Based on the model assessment results and risk-related historical data, the correlation between risk factors and actual events is analyzed through Bayesian networks to calibrate the risk assessment model and output a calibrated risk scoring system;

[0013] Based on the calibrated risk scoring system, by calculating scores and setting differentiated thresholds for real-time monitoring data, dynamic classification of livestock risk levels and early warning triggering are achieved, thereby obtaining graded risk early warning results;

[0014] Based on the hierarchical risk warning results, by establishing a risk-loss mapping relationship, intervention suggestions are generated for risk quantification assessment, risk prevention and control decisions, and loss assessment, thereby realizing the integration of risk assessment results and risk management processes.

[0015] According to one embodiment of the present invention, extracting risk features at the environmental level through correlation analysis based on the standardized multi-source dataset includes:

[0016] Based on the standardized multi-source data set, a discriminant probability graphical model of the conditional neural field is constructed, and conditional probability calculation is performed on the multi-source sensor data to obtain an initial health state distribution;

[0017] Based on the initial health state distribution, local spatiotemporal features are extracted through a 3D convolutional neural network, and the spatiotemporal features are analyzed for long-range dependencies using a self-attention mechanism to generate a spatiotemporal feature expression;

[0018] Based on the spatiotemporal feature expression, the discriminant probability graphical model is optimized and trained by introducing the heat diffusion equation constraint and the energy conservation constraint as physical constraint regularization terms to obtain the optimized spatiotemporal health field;

[0019] The optimized spatiotemporal health field is sampled multiple times by a Monte Carlo dropout method to construct a posterior probability distribution of health status and obtain a health risk representation with uncertainty estimation;

[0020] Based on the health risk representation with uncertainty estimation, feature fusion is performed through an attention weighting mechanism to generate an enhanced risk feature representation.

[0021] According to one embodiment of the present invention, based on the hierarchical risk warning results, generating intervention suggestions for risk quantification assessment, risk prevention and control decision-making, and loss assessment by establishing a risk-loss mapping relationship includes:

[0022] Based on the graded risk warning results, a context vector is constructed containing basic information of the cattle, physiological indicators, nutritional parameters and environmental factors, and an evaluation function is constructed according to treatment effects, production performance changes and resource consumption to generate basic decision-making data;

[0023] Based on the decision-making basic data, a Q function network is constructed through a four-layer neural network, and a time difference learning method and a prediction model are used to deal with the delayed feedback problem to generate intervention evaluation indicators;

[0024] Based on the intervention evaluation metric, a variant of Thompson sampling is used to maintain the posterior distribution of the action-context combination, and a historical influence attenuation factor is introduced to adjust the exploration strategy and generate dynamic decision parameters.

[0025] Based on the dynamic decision parameters, a hierarchical decision structure is constructed through a two-stage decision method to classify and select intervention measures and generate a specific intervention plan, wherein the specific intervention plan includes a number of intervention suggestions;

[0026] Based on the specific intervention plan, a hybrid decision-making model is achieved by combining rule guidance and data-driven through a progressive weight adjustment mechanism.

[0027] According to one embodiment of the present invention, extracting the disease risk characteristics at the population level through cluster analysis based on the standardized multi-source dataset includes:

[0028] Based on the standardized multi-source dataset, a multi-layer heterogeneous risk association graph including livestock nodes, environmental nodes, and management behavior nodes is constructed, and edge connections between nodes are established by analyzing spatial proximity and association relationships to generate an initial graph structure;

[0029] Based on the initial graph structure, by designing node feature reconstruction, edge prediction and subgraph attribute prediction tasks, a self-supervised analysis of the data structure is performed to generate a sample set for training livestock breeding risks;

[0030] Based on the training sample set, the mutual information of node representation is decomposed into feature transfer information, structural common information and structural independent information through information theory method, and a balanced loss function is constructed;

[0031] Based on the loss function, information is transferred between different types of nodes and edges through the heterogeneous graph attention network, and the weight parameters are dynamically adjusted according to the characteristics of the graph structure to obtain an optimized graph representation;

[0032] Based on the optimized graph representation, a classifier is used to identify disease transmission paths and generate risk assessment results.

[0033] According to one embodiment of the present invention, the multiple IoT devices are used to collect the vital signs data of the individual livestock, including biochip ear tags and NB chip ear tags for collecting body temperature, activity level, number of ruminations and lying time, and environmental monitoring equipment for collecting temperature, humidity, ammonia concentration and light, wherein the management behavior data includes feeding frequency, feed ratio, immunization records and treatment records.

[0034] According to one embodiment of the present invention, extracting individual-level health risk characteristics through time series analysis based on the standardized multi-source dataset includes:

[0035] Based on body temperature and activity data, physiological rhythm characteristics are extracted through time series analysis;

[0036] Based on feeding and rumination data, behavioral characteristics are extracted through behavioral pattern analysis;

[0037] Based on the activity state data, state features are extracted through state transition analysis;

[0038] A health baseline model is constructed based on the physiological rhythm characteristics, behavioral characteristics and state characteristics.

[0039] According to one embodiment of the present invention, the method of constructing differentiated risk assessment models for different livestock breeds and different growth stages based on the multidimensional risk signature library using a machine learning method includes:

[0040] Construct a health risk assessment sub-model based on supervised learning algorithms;

[0041] Construct a mortality risk assessment sub-model based on survival analysis method;

[0042] Construct a production performance risk assessment sub-model based on regression analysis method;

[0043] Based on the ensemble learning method, the results of each sub-model are integrated to generate a risk assessment model.

[0044] According to one embodiment of the present invention, analyzing the correlation between risk factors and actual events through a Bayesian network, calibrating the risk assessment model, and outputting a calibrated risk scoring system includes:

[0045] Based on historical data, collect and organize information on risk event types, loss levels, and timing;

[0046] Based on the risk event information, construct a correlation data set between risk factors and actual events;

[0047] Based on the association data set, applying Bayesian network to perform causal association analysis;

[0048] Based on the results of causal association analysis, calculate the intensity of the risk factor's influence on the occurrence of the event;

[0049] Setting differentiated calibration parameters based on the intensity of the conditional influence;

[0050] Using the calibration parameters to correct and adjust the output results of the risk assessment model;

[0051] Based on the calibrated assessment results, a tiered risk scoring system is constructed;

[0052] In the hierarchical risk scoring system, weight coefficients are assigned to different risk types and levels to form a comprehensive risk scoring standard.

[0053] According to one embodiment of the present invention, the step of calculating scores and setting differentiation thresholds for real-time monitoring data includes:

[0054] Build a real-time data processing mechanism based on stream processing architecture;

[0055] Based on risk level classification, set up multi-level warning threshold strategies;

[0056] Calculate the rate of change and acceleration of risk scores based on real-time data;

[0057] Select the appropriate warning notification method based on risk scoring and warning strategy;

[0058] Based on the early warning results, record verification information and continuously optimize the early warning model.

[0059] According to one embodiment of the present invention, generating intervention suggestions for risk quantification assessment, risk prevention and control decision-making, and loss assessment by establishing a risk-loss mapping relationship includes:

[0060] Based on the risk assessment results, a correlation model between risk factors and loss extent is constructed;

[0061] Calculating risk quantification indicators based on the association model;

[0062] Generate risk prevention and control suggestions and intervention measures based on risk quantification indicators;

[0063] Based on historical data, analyze the consistency between actual events and risk monitoring records to obtain analysis results;

[0064] Based on the analysis results, targeted risk intervention recommendations are generated.

[0065] The present invention also provides an intelligent livestock breeding risk assessment device based on multi-source data fusion, comprising:

[0066] A data collection and preprocessing module is used to collect livestock individual vital sign data, breeding environment parameters, management behavior data, and risk-related historical data through multiple IoT devices, and clean, process outliers, and extract features from the multi-source data to obtain a standardized multi-source data set;

[0067] A risk feature extraction module is used to extract individual-level health risk features, group-level disease risk features, and environmental-level risk features based on the standardized multi-source data set, and to fuse the risk features to form a multidimensional risk feature library;

[0068] A risk assessment model construction module is used to construct differentiated risk assessment models for different livestock breeds and different growth stages based on the multidimensional risk feature library and generate model assessment results;

[0069] A risk calibration module is used to analyze the correlation between risk factors and actual events based on risk-related historical data, calibrate the risk assessment model, and output a calibrated risk scoring system;

[0070] A risk grading and early warning module is used to score and calculate real-time monitoring data based on the calibrated risk scoring system, and set differentiated risk thresholds to achieve dynamic classification and early warning triggering of livestock risk levels, thereby obtaining graded risk early warning results;

[0071] The risk management module is used to generate intervention suggestions for risk quantification assessment, risk prevention and control decisions and loss assessment based on the hierarchical risk warning results, so as to achieve the integration of risk assessment results and risk management processes.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] (1) Multi-source data fusion: The present invention collects multi-source data such as individual livestock vital signs, breeding environment parameters, management behaviors and historical data through IoT devices, which makes up for the lack of data dimensions in traditional methods and provides a more comprehensive basis for risk assessment.

[0074] (2) Multi-level risk feature extraction: The present invention extracts risk features from three levels: individual, group, and environment, and forms a multi-dimensional risk feature library through feature fusion algorithm to comprehensively capture risk signals at different levels and improve the accuracy of risk identification.

[0075] (3) Differentiated risk assessment model: The present invention constructs differentiated risk assessment models for different livestock breeds and different growth stages, overcoming the limitations of the one-size-fits-all approach of traditional methods and improving the accuracy of risk assessment.

[0076] (4) Risk calibration based on historical data: The present invention uses Bayesian networks to analyze the relationship between risk factors and actual events, and calibrates the risk assessment model to make the risk assessment results more consistent with the actual situation.

[0077] (5) Integration of risk assessment and management processes: Based on the hierarchical risk warning results, the present invention generates targeted intervention suggestions, realizes the deep integration of risk assessment and risk management processes, and improves the practicality and effectiveness of risk management. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments. The drawings herein are incorporated into and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and, together with the specification, are used to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without inventive effort.

[0079] Figure 1 This is a flow chart of an intelligent livestock breeding risk assessment method based on multi-source data fusion according to an embodiment of the present invention;

[0080] Figure 2 This is a structural diagram of an intelligent livestock breeding risk assessment device based on multi-source data fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the disclosure for which protection is sought, but merely represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present disclosure.

[0082] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not require further definition or explanation in subsequent drawings.

[0083] The term "and / or" herein simply describes an association relationship, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, and the existence of B alone. In addition, the term "at least one" herein refers to any combination of at least two of any one or more of a plurality of items. For example, "at least one of A, B, and C" can represent any one or more elements selected from the set consisting of A, B, and C.

[0084] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0085] See also Figure 1 , Figure 1 A flowchart of a method for intelligently assessing livestock breeding risk based on multi-source data fusion, provided in accordance with one embodiment of the present invention. The method comprises the following steps:

[0086] Step S101: Obtain individual livestock vital signs data, breeding environment parameters, management behavior data, and risk-related historical data collected by multiple IoT devices, and obtain a standardized multi-source data set through data cleaning, outlier processing, and feature extraction.

[0087] During the implementation of the present invention, step S101 is first performed, which involves acquiring individual livestock vital sign data, breeding environment parameters, management behavior data, and risk-related historical data collected by multiple IoT devices. Through data cleaning, outlier processing, and feature extraction, a standardized multi-source dataset is obtained. This step is the foundation of the entire risk assessment process. Through a network of IoT devices deployed in the livestock breeding environment, such as biochip ear tags, NB chip ear tags, and environmental sensors, livestock vital sign data such as body temperature and activity level, as well as parameters such as ambient temperature and humidity, are collected in real time. Simultaneously, the system also records breeding and management behavior data and historical risk records. This raw data undergoes preprocessing steps such as cleaning, outlier processing, and standardization to form a high-quality, uniformly formatted multi-source dataset, providing the data foundation for subsequent risk feature extraction.

[0088] Step S102: Based on the standardized multi-source data set, extract individual-level health risk characteristics through time series analysis, extract group-level epidemic risk characteristics through cluster analysis, extract environmental-level risk characteristics through correlation analysis, and use feature fusion algorithm to form a multidimensional risk feature library.

[0089] Next, step S102 is executed. Based on the standardized multi-source dataset, individual health risk characteristics are extracted through time series analysis, group disease risk characteristics are extracted through cluster analysis, and environmental risk characteristics are extracted through correlation analysis. A feature fusion algorithm is then used to form a multidimensional risk signature library. In this step, the system extracts and analyzes risk characteristics at multiple levels. At the individual level, physiological rhythm characteristics and behavioral pattern characteristics are extracted through analysis of livestock vital sign time series data to construct an individual health baseline model. At the group level, cluster analysis of the behavior of similar livestock groups within the farm is performed to identify possible disease transmission patterns. At the environmental level, correlation analysis is used to extract environmental risk characteristics and assess the impact of environmental factors on livestock health. Finally, the system integrates these different level characteristics through a feature fusion algorithm to form a multidimensional risk signature library that comprehensively reflects the risk status.

[0090] Step S103: Based on the multidimensional risk feature library, a machine learning method is used to construct differentiated risk assessment models for different livestock breeds and different growth stages, and a model assessment result is generated.

[0091] Then, step S103 is executed, that is, based on the multidimensional risk feature library, a differentiated risk assessment model is constructed for different livestock breeds and different growth stages using a machine learning method to generate a model assessment result. In this step, the system fully considers the specificity of different livestock breeds and growth stages and establishes a hierarchical model architecture. For different risk types such as health risk, death risk and output risk, the system constructs special sub-models, such as a health risk assessment model based on supervised learning, a death risk assessment model based on survival analysis, and an output risk assessment model based on regression analysis. These sub-models are integrated through an ensemble learning method to form a comprehensive risk assessment model, which outputs standardized risk scores and risk type judgment results. This differentiated model design can adapt to the specific needs of different breeding scenarios and improve the accuracy of risk assessment.

[0092] Step S104: Based on the model evaluation results and risk-related historical data, the correlation between risk factors and actual events is analyzed through a Bayesian network, the risk evaluation model is calibrated, and a calibrated risk scoring system is output.

[0093] Then, step S104 is executed, that is, based on the model evaluation results and risk-related historical data, the correlation between risk factors and actual events is analyzed through a Bayesian network, the risk assessment model is calibrated, and a calibrated risk scoring system is output. This step introduces historical risk event data to establish a correlation data set between risk factors and actual events, applies a Bayesian network to perform causal analysis, and identifies key risk factors and their weights. Based on the analysis results, the system calculates the calibration coefficient and corrects and adjusts the output results of the risk assessment model to ensure that the risk assessment results are more consistent with the actual risk situation. Through this calibration process based on historical data, the risk assessment model can continuously learn and improve, adapt to the dynamic changes in the breeding environment, and improve the accuracy of risk prediction.

[0094] Step S105: Based on the calibrated risk scoring system, by performing score calculation and differentiated threshold setting on the real-time monitoring data, dynamic classification of livestock risk levels and early warning triggering are achieved to obtain graded risk early warning results.

[0095] Then execute step S105, that is, based on the calibrated risk scoring system, by performing score calculation and differentiated threshold setting on the real-time monitoring data, the dynamic division of livestock risk levels and early warning triggering are realized, and a graded risk early warning result is obtained. In this step, the system constructs a real-time data processing mechanism based on the stream processing architecture to support risk calculation on different time scales. According to factors such as livestock breed and growth stage, the system sets a differentiated risk threshold system to divide the risk level into multiple levels. For real-time monitoring data, the system not only calculates the current risk score, but also analyzes the rate of change and acceleration of the risk score to predict the risk trend. When the risk score exceeds the corresponding threshold, the system will automatically trigger an early warning and select an appropriate early warning notification method according to the risk level to ensure that relevant personnel can obtain risk information in a timely manner and take measures.

[0096] Step S106: Based on the graded risk warning results, by establishing a risk-loss mapping relationship, intervention suggestions are generated for risk quantification assessment, risk prevention and control decision-making, and loss assessment, thereby realizing the integration of risk assessment results and risk management processes.

[0097] Finally, step S106 is executed, that is, based on the hierarchical risk warning results, by establishing a risk-loss mapping relationship, intervention suggestions are generated for risk quantification assessment, risk prevention and control decisions, and loss assessment, thereby realizing the integration of risk assessment results and risk management processes. This step is a key link in converting risk assessment results into practical application value. Based on the risk assessment results, the system constructs a correlation model between risk factors and loss levels, calculates risk quantification indicators, and provides data support for risk pricing. For identified risks, the system will generate targeted risk prevention and control suggestions and intervention measures, such as adjusting feeding and management methods, improving environmental conditions, or implementing disease prevention and control. At the same time, the system also supports risk underwriting decisions and verification to ensure the scientific nature and effectiveness of the risk management process. Through this deep integration of risk assessment and risk management, the system can provide comprehensive risk management solutions for breeding companies, reduce breeding risks, and improve breeding benefits.

[0098] The following describes in detail the various steps of the above method in conjunction with specific embodiments.

[0099] In one embodiment, extracting risk features at the environmental level through correlation analysis based on the standardized multi-source dataset includes:

[0100] A1. Based on the standardized multi-source data set, construct a discriminant probability graphical model of the conditional neural field, perform conditional probability calculation on the multi-source sensor data, and obtain an initial health state distribution;

[0101] A2. Based on the initial health state distribution, extract local spatiotemporal features using a 3D convolutional neural network, and use a self-attention mechanism to perform long-range dependency analysis on the spatiotemporal features to generate a spatiotemporal feature representation;

[0102] A3. Based on the spatiotemporal feature expression, the discriminant probability graphical model is optimized and trained by introducing heat diffusion equation constraints and energy conservation constraints as physical constraint regularization terms to obtain an optimized spatiotemporal health field.

[0103] A4. Sampling the optimized spatiotemporal health field multiple times using the Monte Carlo dropout method to construct a posterior probability distribution of health status and obtain a health risk representation with uncertainty estimation;

[0104] A5. Based on the health risk representation with uncertainty estimation, feature fusion is performed through an attention weighting mechanism to generate an enhanced risk feature representation.

[0105] In one embodiment of the present invention, when extracting risk features at the environmental level based on a standardized multi-source data set, it is first necessary to execute step A1, that is, based on the standardized multi-source data set, construct a discriminant probability graph model of the conditional neural field, perform conditional probability calculation on the multi-source sensor data, and obtain the initial health status distribution. In this step, the conditional neural field is a discriminant probability graph model that combines the structured prediction ability of the conditional random field and the feature learning ability of the neural network, and can effectively model continuous spatiotemporal data. The system represents the health status of livestock as a spatiotemporal field function H(t,l), where t represents the time dimension and l represents the spatial position, and is modeled by conditional probability P(H|X), where X represents the observation data set collected from multi-source sensors. This modeling method can make full use of the temporal continuity and spatial correlation of the data, laying the foundation for subsequent risk feature extraction.

[0106] Then execute step A2, that is, based on the initial health status distribution, extract local spatiotemporal features through a 3D convolutional neural network, and use the self-attention mechanism to perform long-distance dependency analysis on the spatiotemporal features to generate spatiotemporal feature expressions. In this step, the 3D convolutional neural network is specially designed to process three-dimensional data (two dimensions of time and space), and can effectively capture local spatiotemporal features. The network contains multiple layers of convolutional layers, and each layer uses a different number and size of convolution kernels for feature extraction. In order to capture long-range spatiotemporal dependencies, the system introduces a self-attention mechanism in the middle layer of the convolutional network, so that the model can pay attention to long-range correlations in the data. For example, when an abnormal body temperature occurs in a certain area, the self-attention mechanism can automatically associate the recent activity trajectory and environmental changes in the area, thereby more comprehensively understanding the potential health risk patterns.

[0107] Then, step A3 is executed, that is, based on the spatiotemporal feature expression, the discriminant probability graph model is optimized and trained by introducing the heat diffusion equation constraint and the energy conservation constraint as physical constraint regularization terms to obtain the optimized spatiotemporal health field. The innovation of this step is to introduce the principles of physics into the training process of the machine learning model, so that the model prediction results not only rely on the statistical laws of the data, but also conform to the basic physical laws. The heat diffusion equation constraint is used to simulate the spread of the disease in the population, and the energy conservation constraint is used to model the relationship between the energy consumption and health status of livestock. These physical constraints are added to the loss function in the form of regularization terms, such as λ·‖H(t,l)-R(t,l)‖ 2 , where R(t,l) is the theoretical expected distribution calculated from the physical equations. This allows the model to make reasonable predictions even in data-sparse areas, significantly improving the model’s generalization and prediction accuracy.

[0108] Then execute step A4, that is, use the Monte Carlo dropout method to sample the optimized spatiotemporal health field multiple times, construct the posterior probability distribution of the health state, and obtain a health risk representation with uncertainty estimation. In the field of risk assessment, uncertainty quantification is as important as risk estimation. The Monte Carlo dropout method is a practical Bayesian deep learning approximation method. By keeping the dropout layer (Dropout) in the neural network turned on during the inference phase, the same input is forward propagated multiple times (usually 50 times) to obtain multiple output samples. These samples constitute the posterior probability distribution of the health state, and the mean (as a risk prediction value) and standard deviation (as an uncertainty measure) can be calculated. This method makes the risk assessment result no longer a single point estimate, but contains reliable uncertainty information, providing a more comprehensive basis for decision-making, and is particularly suitable for dealing with decision-making problems in high-risk scenarios.

[0109] Finally, step A5 is executed, which involves performing feature fusion based on the health risk representation with uncertainty estimation through an attention weighting mechanism to generate an enhanced risk feature representation. This step fuses the previously generated health risk representation with uncertainty estimation with risk features from other sources. The system uses an attention weighting mechanism to dynamically adjust the weights of different features during the fusion process based on their reliability and relevance. For example, when the health risk forecast for a certain area has high uncertainty, the system will appropriately reduce its weight and increase the weights of other more reliable features. This adaptive fusion method can effectively handle the heterogeneity and uncertainty of multi-source data, generate a more robust enhanced risk feature representation, and provide high-quality input features for subsequent risk assessment models.

[0110] In one embodiment, based on the graded risk warning results, by establishing a risk-loss mapping relationship, generating intervention suggestions for risk quantification assessment, risk prevention and control decision-making, and loss assessment, the following steps are included:

[0111] B1. Based on the graded risk warning results, construct a context vector containing basic cattle information, physiological indicators, nutritional parameters, and environmental factors. Furthermore, construct an evaluation function based on treatment efficacy, production performance changes, and resource consumption to generate basic decision-making data.

[0112] B2. Based on the decision-making basic data, a Q-function network is constructed using a four-layer neural network. A temporal difference learning method and a prediction model are used to address the delayed feedback problem and generate intervention evaluation indicators.

[0113] B3. Based on the intervention evaluation metric, a variant of Thompson sampling is used to maintain the posterior distribution of the action-context combination. A historical influence attenuation factor is introduced to adjust the exploration strategy and generate dynamic decision parameters.

[0114] B4. Based on the dynamic decision parameters, construct a hierarchical decision structure using a two-stage decision-making method, classify and select intervention measures, and generate a specific intervention plan, which includes several intervention recommendations;

[0115] B5. Based on the specific intervention plan, a hybrid decision-making model is implemented by combining rule-based and data-driven approaches through a progressive weight adjustment mechanism.

[0116] In another embodiment of the present invention, when generating intervention recommendations based on graded risk warning results, step B1 is first performed. Based on the graded risk warning results, a context vector is constructed containing the cow's basic information, physiological indicators, nutritional parameters, and environmental factors. An evaluation function is then constructed based on treatment efficacy, changes in production performance, and resource consumption to generate decision-making foundation data. In this step, the system constructs a comprehensive context vector containing 32 features: basic cow information (such as age, lactation period, and milk production), physiological indicators (such as body temperature and rumination time), nutritional parameters (such as feed intake and feed type), and environmental factors (such as temperature, humidity, and hygiene score). Furthermore, the system designs a comprehensive evaluation function that considers multiple factors: treatment efficacy (such as the degree of somatic cell count recovery), changes in production performance (such as changes in milk production), resource consumption (such as treatment costs), and antibiotic use (reduced use is considered a positive contribution). This decision-making foundation data provides comprehensive information support for subsequent intervention decisions.

[0117] Next, step B2 is performed. Based on the decision-making data, a Q-function network is constructed using a four-layer neural network. Temporal-difference learning and a predictive model are used to address delayed feedback and generate intervention evaluation metrics. This step addresses the key challenge of delayed feedback on intervention effects in livestock farming. The system implements a four-layer neural network as the Q-function network. The input layer receives contextual features and action encodings, with two hidden layers (128 and 64 neurons, respectively, using the Reluctant Unit (ReLU) activation function) in the middle. The output layer predicts the expected reward value. To address the delayed feedback issue, the system innovatively employs temporal-difference learning and a "pseudo-reward" mechanism: before the actual reward is obtained, a pseudo-reward is generated using a predictive model for preliminary updates. Once the actual reward is available, the difference between the pseudo-reward and the actual reward is calculated and used for correction and update. This approach effectively alleviates the problem of decreased learning efficiency caused by delayed feedback, enabling the system to more rapidly learn the optimal intervention strategy.

[0118] Then, step B3 is executed. Based on the intervention evaluation metric, a variant of Thompson sampling is used to maintain the posterior distribution of the action-context combination. A historical influence reduction factor is introduced to adjust the exploration strategy and generate dynamic decision parameters. Thompson sampling is a classic exploration-exploitation balance algorithm. In this system, a variant of Thompson sampling is employed to maintain a posterior distribution for each action-context combination (using variational inference methods in Bayesian neural networks). Each decision is sampled from this distribution, and the action with the highest sample value is selected. The "historical influence reduction factor" β introduced by the system increases the influence of recent feedback on decision-making, adapting to dynamically changing environments. Furthermore, the system incorporates an uncertainty-based exploration enhancement mechanism: when faced with delayed feedback, the exploration tendency is appropriately increased to prevent over-exploitation of incorrect decisions. These dynamic decision parameters enable the system to continuously learn and adjust intervention strategies in complex and changing livestock farming environments.

[0119] Next, step B4 is executed. Based on the dynamic decision parameters, a hierarchical decision structure is constructed using a two-stage decision-making approach to categorize and select intervention measures and generate a specific intervention plan, which contains several intervention recommendations. Faced with a complex space of intervention measures, the system employs a hierarchical decision-making structure to simplify the decision-making process. In the first stage, the system categorizes intervention measures into several broad categories (such as observation, prevention, and treatment). In the second stage, the system selects specific intervention measures within the selected categories. For example, the treatment category might include options such as "local antibiotic therapy," "systemic antibiotic therapy," and "combination therapy." This hierarchical decision-making approach significantly improves learning efficiency, enabling the system to more quickly grasp the fundamental decision-making logic of "when to intervene" and "how to intervene." The resulting specific intervention plan contains multiple intervention recommendations, including both the intervention measures themselves and detailed information such as implementation timing, duration, and expected effects.

[0120] Finally, step B5 is executed, that is, based on the specific intervention plan, a hybrid decision-making model is implemented by combining rule-based and data-driven decision-making through a progressive weight adjustment mechanism. In the early stages of system operation, there is insufficient data to support purely data-driven decision-making, and the system mainly follows rules based on expert knowledge to guide decision-making; as data accumulates and learning deepens, the system gradually increases the weight of data-driven decision-making. This progressive weight adjustment mechanism is implemented through a time-dependent function, such as w(t) = min(0.2 + 0.8 × (1-e -t / τ ),1), where t represents the system runtime and τ is the adjustment rate parameter. This hybrid decision-making model combines the security of a rule-based system with the adaptability of a data-driven system, making it particularly suitable for sectors like animal husbandry, which require both specialized knowledge and face complex and dynamic environments. Through this organic combination of rule-based guidance and data-driven approaches, the system can continuously learn and optimize intervention strategies while ensuring safety, ultimately achieving more precise and effective risk management.

[0121] In one embodiment, extracting the disease risk characteristics at the population level through cluster analysis based on the standardized multi-source dataset includes:

[0122] C1. Based on the standardized multi-source dataset, construct a multi-layer heterogeneous risk association graph comprising livestock nodes, environmental nodes, and management behavior nodes. Edge connections between nodes are established by analyzing spatial proximity and association relationships to generate an initial graph structure.

[0123] C2. Based on the initial graph structure, perform self-supervised analysis of the data structure by designing node feature reconstruction, edge prediction, and subgraph attribute prediction tasks to generate a sample set for training livestock breeding risks;

[0124] C3. Based on the training sample set, decompose the mutual information of node representations into feature transfer information, structural common information, and structural independent information using information theory methods, and construct a balanced loss function;

[0125] C4. Based on the loss function, the heterogeneous graph attention network transmits information between different types of nodes and edges, and dynamically adjusts weight parameters according to the characteristics of the graph structure to obtain an optimized graph representation;

[0126] C5. Based on the optimized graph representation, use a classifier to identify the disease transmission path and generate a risk assessment result.

[0127] In another embodiment of the present invention, when extracting population-level disease risk characteristics through cluster analysis, step C1 is first performed. Based on the standardized multi-source dataset, a multi-layer heterogeneous risk association graph (MHRAG) is constructed, comprising livestock nodes, environmental nodes, and management behavior nodes. Edge connections between nodes are established by analyzing spatial proximity and association relationships, generating an initial graph structure. This step innovatively models the livestock breeding environment as a multi-layer heterogeneous risk association graph (MHRAG), comprising multiple types of nodes: livestock nodes (e.g., individual dairy cows, beef cattle, etc.), environmental nodes (e.g., environmental factors such as temperature, humidity, and ammonia concentration), and management behavior nodes (e.g., management activities such as feeding, immunization, and treatment). The system establishes contact relationships between livestock nodes by analyzing spatial proximity (e.g., establishing edge connections between livestock sharing the same fenced area), establishes exposure relationships between livestock and environmental nodes through physical location relationships, and establishes intervention relationships between livestock and management behavior nodes through operational records. Each node and edge is accompanied by a corresponding feature vector, such as physiological indicators of livestock nodes, parameter values of environmental nodes, and the strength and duration of edge relationships. This graph structure representation can intuitively capture complex risk associations and provide a structured basis for subsequent risk propagation analysis.

[0128] Next, step C2 is executed. Based on the initial graph structure, a self-supervised analysis of the data structure is performed by designing node feature reconstruction, edge prediction, and subgraph attribute prediction tasks to generate a training set of livestock breeding risk samples. Because labeled data (such as confirmed disease cases) is often scarce in livestock breeding environments, the system uses self-supervised learning to generate training samples. Three types of pre-training tasks are specifically designed: node feature reconstruction tasks require the model to predict hidden node features based on the graph structure and partial node features, such as predicting the body temperature of a particular cattle; edge prediction tasks require the model to determine whether a connection exists between two nodes, such as predicting whether two cattle have close contact; and subgraph attribute prediction tasks require the model to predict the overall properties of a subgraph consisting of a specific set of nodes, such as predicting the average body temperature trend of a herd of cattle in a fenced area. These tasks do not require external annotation, but instead leverage the structural characteristics of the data to create supervisory signals and generate a large number of training samples. In this way, the system can fully exploit the structural information inherent in the data, providing a rich set of training samples for subsequent graph representation learning.

[0129] Then, step C3 is performed. Based on the training sample set, information-theoretic methods are used to decompose the mutual information of node representations into feature-transmitted information, structurally shared information, and structurally independent information, thereby constructing a balanced loss function. Traditional graph neural networks often suffer from the "oversmoothing" problem, where the representations of connected nodes become too similar, resulting in reduced discriminative ability. The system innovatively applies information-theoretic methods to address this issue, decomposing the mutual information of node representations, I(X;Z), into four components: I(X;Z) = I(X;Z|Y) + I(X;Y;Z) - I(X;Y|Z) + I(Y;Z|X), where X is the input feature, Y is the graph structure, and Z is the learned representation. Each term has a clear meaning: I(X;Z|Y) represents information directly conveyed by features, I(X;Y;Z) represents information provided by both features and structure, I(X;Y|Z) represents redundant information not captured by the representations, and I(Y;Z|X) represents information provided independently by the structure. Based on this theoretical decomposition, the system designs a balanced loss function:

[0130] L=L pretrain -α xy I(X; Y; Z) + α y I(Y; Z|X)

[0131] where α xy and α y is an adjustable weight coefficient used to control the balance between different information components. This information-theoretic loss function design enables the model to effectively utilize graph structure information while preserving node distinguishability, solving a key challenge in representation learning.

[0132] Then, step C4 is executed, that is, based on the loss function, information is transferred between different types of nodes and edges through the heterogeneous graph attention network, and the weight parameters are dynamically adjusted according to the characteristics of the graph structure to obtain an optimized graph representation. The system uses the heterogeneous graph attention network (HGAT) to process heterogeneous graph data of multiple types of nodes and edges. The core of the network is the heterogeneous graph attention layer, which can use different parameters for information transfer according to different types of nodes and edges. In each attention layer, when node v receives information from its neighbor u, the attention coefficient is calculated as follows:

[0133] α vu =softmax(LeakyReLU(a T [W {phi(v)} h v ||W {phi(u)} h u ]))

[0134] Where φ(v) represents the type of node v, W {phi(v)}is a parameter matrix specific to this type, h v is the node feature, and a is the attention vector. The system also innovatively designs a dynamic weight adjustment mechanism that adaptively adjusts weight parameters based on the characteristics of the current graph structure (such as node degree distribution and clustering coefficient). For example, in areas with a higher risk of disease transmission, the weight of structural information is appropriately increased; in areas with better data quality, the weight of feature information is appropriately increased. This dynamic adjustment strategy enables the model to flexibly balance structural and feature information based on the characteristics of different regions and time periods, generating an optimal graph representation.

[0135] Finally, step C5 is executed: based on the optimized graph representation, a classifier is used to identify disease transmission pathways and generate risk assessment results. The system inputs the optimized graph representation into a specially designed disease transmission pathway identification model. This model combines the strengths of graph neural networks and epidemiological models, considering not only the characteristic representations of nodes but also the dynamics of disease transmission within the network. The system can identify potential "super-spreaders" (individuals that contribute significantly to disease transmission) and high-risk transmission pathways, predicting the potential spread of the disease within a population. Furthermore, the system can evaluate the effectiveness of different isolation and intervention strategies, providing management with informed decision-making support. The final risk assessment output includes an individual risk score (indicating the risk of infection for each livestock), a group risk level (indicating the risk status of the entire farm), a risk transmission heat map (visualizing the spatial distribution of risk), and early warning information (for high-risk individuals and regions). These assessment results provide precise risk monitoring and prevention guidance for livestock management, effectively enhancing disease prevention and control capabilities.

[0136] In one embodiment, the multiple IoT devices are used to collect the individual vital signs data of the livestock, including biochip ear tags and NB chip ear tags for collecting body temperature, activity level, number of ruminations and lying time, and environmental monitoring equipment for collecting temperature, humidity, ammonia concentration and light, wherein the management behavior data includes feeding frequency, feed ratio, immunization records and treatment records.

[0137] In one embodiment, extracting individual-level health risk characteristics through time series analysis based on the standardized multi-source dataset includes:

[0138] D1. Extract physiological rhythm characteristics based on body temperature and activity data through time series analysis;

[0139] D2. Analyze behavioral patterns based on feeding and rumination data to extract behavioral characteristics;

[0140] D3. Based on the activity state data, extract state features through state transition analysis;

[0141] D4. Construct a health baseline model based on the physiological rhythm characteristics, behavioral characteristics and state characteristics.

[0142] In another embodiment of the present invention, when extracting individual-level health risk characteristics through time series analysis, step D1 is first performed, namely, extracting circadian rhythm characteristics through time series analysis based on body temperature and activity data. Livestock's body temperature and activity level are important physiological indicators reflecting their health status. These indicators typically exhibit certain periodic variations, such as diurnal patterns and patterns of change after feeding. The system uses multiple time series analysis methods to extract these circadian rhythm characteristics, including statistical feature extraction (such as calculating body temperature statistics such as mean, variance, skewness, and kurtosis), time domain feature extraction (such as using autoregressive models to analyze the temporal dependence of body temperature), and frequency domain feature extraction (such as analyzing the periodic components of body temperature through Fourier transforms). In particular, the system can identify diurnal patterns of body temperature, including characteristics such as peak time, fluctuation amplitude, and stability; it can also capture regular changes in activity level, such as peak activity periods, rest periods, and activity intensity distribution. These circadian rhythm characteristics provide key input for baseline modeling of individual livestock health status and can effectively detect abnormal physiological changes.

[0143] Next, step D2 is performed, which involves extracting behavioral features through behavioral pattern analysis based on feeding and rumination data. Feeding and rumination behavior are important indicators of the health of the livestock's digestive system. The system extracts a rich set of behavioral features by analyzing data collected by devices such as mouth sensors and neck collar sensors. For feeding behavior, the system extracts features such as the number of daily feedings, the duration of each feeding session, feeding rate, total daily feeding time, feeding time distribution, and regularity of feeding intervals. For rumination, the system extracts features such as the number of daily ruminations, the duration of each rumination, rumination efficiency (rumination time per unit of feed intake), and rumination patterns (e.g., the proportion of nighttime ruminations). The system also analyzes the stability and changing trends of these behaviors, such as the coefficient of variation of feeding behavior over three consecutive days, which can reflect the stability of livestock behavioral patterns. These behavioral features directly reflect the digestive system function and nutritional status of livestock and are valuable for early detection of digestive system diseases and malnutrition.

[0144] Step D3 involves extracting state features through state transition analysis based on activity state data. The animal's activity states (e.g., standing, walking, lying down, feeding, etc.) and their transition patterns are crucial for assessing locomotor health, comfort, and overall well-being. The system analyzes accelerometer data to identify different activity states and conduct state transition analysis. Extracted state features include: the duration distribution of each state (e.g., average lying down time, maximum standing time), state transition frequency (e.g., the number of transitions from standing to lying down within a 24-hour period), state transition patterns (e.g., a probability matrix for transitions between specific states), and state distributions over different time periods (e.g., the proportion of lying down at night). The system also focuses on abnormal state behaviors, such as prolonged standing time (which may indicate hoof disease or joint pain), frequent posture changes (which may indicate discomfort), or abnormal resting behavior (which may indicate illness). These state features reflect the animal's behavioral habits and comfort, providing important insights for assessing locomotor health and overall welfare.

[0145] Finally, step D4 involves constructing a health baseline model based on the circadian, behavioral, and state characteristics. This health baseline model serves as a reference standard for assessing the health status of individual livestock and is crucial for identifying abnormal health states. The system employs a variety of methods to construct a personalized health baseline model, including statistical modeling (e.g., using historical data to calculate the normal range and fluctuation pattern of features), machine learning (e.g., using normal state data to train anomaly detection models), and time series modeling (e.g., using an autoregressive integrated moving average model to predict normal time series patterns). The health baseline model considers individual differences in livestock (e.g., age, breed, weight), physiological cycles (e.g., lactation and reproductive cycles), and environmental factors (e.g., seasonal variations and weather conditions) to establish a personalized health reference standard for each livestock. The system calculates the degree of deviation from the health baseline in real time, using methods such as the Z-score (the difference between the current value and the baseline mean divided by the standard deviation) or the Mahalanobis distance (a multidimensional deviation measure that considers inter-feature correlations), to generate a health risk score. This personalized health baseline-based risk assessment approach adapts to individual differences among livestock, improves the accuracy and sensitivity of anomaly detection, and provides a scientific basis for early identification of health risks.

[0146] In one embodiment, the multi-dimensional risk signature library is used to construct differentiated risk assessment models for different livestock breeds and different growth stages using a machine learning method, including:

[0147] E1. Construct a health risk assessment sub-model based on supervised learning algorithms;

[0148] E2. Construct a mortality risk assessment sub-model based on survival analysis methods;

[0149] E3. Construct a production performance risk assessment sub-model based on regression analysis;

[0150] E4. Based on the ensemble learning method, the results of each sub-model are integrated to generate a risk assessment model.

[0151] In an embodiment of the present invention in which a machine learning method is used to construct a differentiated risk assessment model, step E1 is first performed, i.e., a health risk assessment sub-model is constructed based on a supervised learning algorithm. Health risk assessment is a core link in livestock breeding risk management, focusing on the possibility of livestock diseases and their severity. The system uses a supervised learning method to construct the sub-model, with marked health abnormality cases in historical data (such as clinically diagnosed respiratory diseases, digestive system diseases, metabolic diseases, etc.) as training labels. The input features include multidimensional risk features extracted in the previous steps, covering multiple dimensions such as physiological indicators, behavioral patterns, and environmental exposure. The system tries a variety of supervised learning algorithms, including random forests (which can handle nonlinear relationships between features), gradient boosting trees (which are more robust to outliers), support vector machines (suitable for processing high-dimensional data), and deep neural networks (which can automatically learn feature representations). The performance of each algorithm is evaluated by five-fold cross validation, and the main considerations include accuracy, precision, recall, F1 score, and AUC value. Ultimately, the algorithm with the best performance is selected or a model integration approach is adopted. The output of this sub-model includes the current health risk probability of individual livestock (such as the probability of contracting a certain type of disease) and risk type identification (such as respiratory risk, digestive system risk, etc.), providing accurate health risk assessment for subsequent risk management.

[0152] Then execute step E2, which is to construct a mortality risk assessment sub-model based on the survival analysis method. Mortality risk assessment focuses on the probability of death and expected survival time of livestock under specific conditions, which is crucial for formulating risk management strategies and evaluating economic impacts. The system innovatively uses the survival analysis method to construct this sub-model. Survival analysis is particularly suitable for dealing with time-event problems involving censored data. In specific implementation, the system uses methods such as the Cox Proportional Hazards Model or the Random Survival Forest, using the risk factor combination and survival time of historical death cases as training data. The basic form of the Cox model is h(t|X)=h o (t)exp(βX), where h o(t) is the baseline risk function, X is the risk factor vector, and β is the coefficient vector. The model takes into account a variety of risk factors, including basic characteristics (such as age, breed), health status (such as previous medical history), environmental conditions (such as extreme temperature exposure) and management factors (such as the implementation of epidemic prevention measures). The system controls the complexity of the model through regularization techniques (such as LASSO or elastic network) to prevent overfitting. The output of this sub-model is the death probability curve (survival function) and expected survival time under specific risk conditions, with special attention to early warning signals of high death risk, such as sharp changes in breathing patterns and abnormally high body temperature, providing a scientific basis for timely intervention.

[0153] Then, step E3 is executed, which constructs a production performance risk assessment sub-model based on regression analysis. In livestock farming, production performance (such as milk production in dairy cows and weight gain rate in beef cattle) is directly related to economic benefits. Assessing production performance risk plays an important role in guiding livestock farming decisions. The system uses regression analysis to construct this sub-model, predicting the expected output level and volatility risk of livestock under specific risk conditions. Based on the data characteristics, the system selects an appropriate regression method, including multiple linear regression (suitable for linear relationships), support vector regression (suitable for nonlinear relationships), random forest regression (suitable for handling interactions between features), or deep neural network regression (suitable for complex pattern recognition). Model inputs include individual livestock characteristics (such as days of lactation and parity), health status indicators, nutritional intake data, and environmental conditions. To improve prediction accuracy, the system uses feature selection methods (such as stepwise regression or tree-based feature importance assessment) to select the most relevant predictor variables. Furthermore, the system not only predicts the expected value of output but also estimates the confidence interval of the prediction to quantify the prediction uncertainty. This approach, which includes uncertainty estimation, can more comprehensively characterize production performance risks, helping managers assess risk levels under different scenarios and develop corresponding risk management strategies.

[0154] Next, step E4 is performed, integrating the results of each sub-model using an ensemble learning approach to generate a risk assessment model. While health risk, mortality risk, and production performance risk each focus on their own, they are often interrelated and interdependent in real-world farming. The system integrates the outputs of these three sub-models using an ensemble learning approach to form a comprehensive risk assessment model. The ensemble strategy comprises three levels: First, within each sub-model, multiple base models are integrated using methods such as model averaging or stacking. Second, risk assessment results are integrated across sub-models using a weighted average approach, with weights dynamically adjusted based on the needs of different scenarios. Finally, the system constructs a meta-model to learn the complex mapping between sub-model outputs and the overall risk. This meta-model is implemented using logistic regression, random forests, or deep neural networks. Its input is the predictions of each sub-model, and its output is a standardized overall risk score ranging from 0 to 100. The system also provides interpretable analysis of the score, such as the contribution of various risk factors and the key drivers of risk evolution, to help users understand the sources and structure of risk. This integrated approach fully leverages the strengths of different sub-models, improves the comprehensiveness and accuracy of risk assessment, and provides a solid foundation for differentiated risk management.

[0155] In one embodiment, analyzing the correlation between risk factors and actual events through a Bayesian network to calibrate the risk assessment model and outputting a calibrated risk scoring system includes:

[0156] F1. Based on historical data, collect and organize information on risk event types, loss levels, and timing;

[0157] F2. Based on the risk event information, construct a dataset that correlates risk factors with actual events;

[0158] F3. Based on the correlation data set, apply Bayesian network to perform causal association analysis;

[0159] F4. Based on the results of the causal association analysis, calculate the intensity of the impact of the risk factors on the conditions for the occurrence of the event;

[0160] F5. Based on the intensity of the conditions, set differentiated calibration parameters;

[0161] F6. Using the calibration parameters to modify and adjust the output of the risk assessment model;

[0162] F7. Construct a tiered risk scoring system based on the calibrated assessment results;

[0163] F8. In the hierarchical risk scoring system, weight coefficients are assigned to different risk types and levels to form a comprehensive risk scoring standard.

[0164] In an embodiment of the present invention, risk calibration is performed by analyzing the relationship between risk factors and actual events through Bayesian networks. First, in step F1, based on historical data, information on risk event types, loss levels, and timing is collected and organized. The quality of risk calibration depends largely on the quality and completeness of historical data. The system comprehensively collects and organizes historical risk event data, including disease events (such as respiratory diseases, digestive system diseases, metabolic diseases, etc.), mortality events, and production performance decline events. For each type of event, the system records key information such as the time of occurrence, duration, severity, and economic losses. At the same time, the system also collects background information such as environmental conditions, management measures, and livestock status before and after the event to provide comprehensive data support for subsequent correlation analysis. The data organization process includes data cleaning (handling missing values and outliers), data standardization (standardizing the format and units of data from different sources), and data enhancement (supplementing some missing information through reasonable assumptions). The system pays special attention to ensuring the temporal integrity of the data, ensuring that the evolution of risk factors before the event can be traced, which is crucial for establishing causal association models.

[0165] Step F2, based on the risk event information, constructs a dataset linking risk factors and actual events. In this step, the system pairs the collected risk events with possible risk factors to create a structured, linked dataset. The system first defines a time window (typically one to 30 days before the event, adjusted based on the event type) and extracts risk factor data within this time window, including environmental indicators (such as changes in temperature and humidity), physiological indicators (such as body temperature trends), behavioral indicators (such as changes in feed intake), and management indicators (such as feed adjustments). For each risk event, the system constructs a sample characterized by the time series of risk factors preceding the event, labeled with the event type and severity. To address sample imbalance (typically, there are fewer risk events and more normal samples), the system employs a combination of undersampling and oversampling strategies, such as the SMOTE (Synthetic Minority Over-sampling Technique) algorithm, to generate synthetic minority samples. Furthermore, the system constructs a control group consisting of samples without risk events under similar conditions for comparative analysis of risk factor differences. This carefully designed linked dataset provides high-quality input data for subsequent Bayesian network analysis.

[0166] Step F3, based on the associated data set, applies a Bayesian network to perform causal association analysis. A Bayesian network is a probabilistic graphical model that can intuitively represent the conditional dependencies between variables and is particularly suitable for analyzing the complex relationships between risk factors and risk events. The system constructs the Bayesian network in two stages: structural learning and parameter learning. First, a structural learning algorithm (such as the PC algorithm, K2 algorithm, or MCMC method) is used to determine the network's topology, i.e., which risk factors directly affect risk events and which factors are interdependent. Then, a parameter learning algorithm (such as maximum likelihood estimation or Bayesian estimation) is used to determine the conditional probability table and quantify the dependency strength between nodes. To improve the interpretability of the model, the system also combines domain knowledge to set some structural priors, such as known causal relationships (for example, ambient temperature affects body temperature). The constructed Bayesian network forms a probabilistic graphical model of risk factors → risk events → loss outcomes, which can intuitively display the risk transmission path and key influencing factors, providing a theoretical basis for risk calibration.

[0167] Step F4, based on the results of the causal association analysis, calculates the conditional impact strength of the risk factor on the occurrence of the event. Within the Bayesian network framework, the system calculates the conditional impact strength of each risk factor on the risk event, which is the core basis for risk calibration. The system uses multiple indicators to quantify the impact strength, including conditional probability difference (the difference in the probability of a risk event occurring when a specific factor is present and absent), information gain (the reduction in event uncertainty due to factor observation), and causal strength (the intervention effect calculated based on Pearl's do-calculus). In addition, the system calculates the combined impact strength to assess the effect of multiple risk factors working together, with particular attention to synergistic effects (the combined effect of multiple factors is stronger than the sum of the individual factors) and antagonistic effects (the ability of certain combinations of factors to offset each other). To improve the robustness of the calculations, the system uses sensitivity analysis to assess the impact of parameter changes on the impact strength calculation and constructs confidence intervals to quantify uncertainty. These conditional impact strength analysis results intuitively demonstrate the importance ranking and interaction patterns of different risk factors, providing a quantitative basis for setting calibration parameters in the next step.

[0168] Step F5, i.e., setting differentiated calibration parameters based on the intensity of the conditional impact. Risk calibration requires setting differentiated calibration parameters for different scenarios to adapt to diverse risk environments. Based on the results of the conditional impact intensity analysis, the system designs a multidimensional calibration parameter system, including: factor weight parameters (setting weight coefficients based on the intensity of the impact of each risk factor), threshold adjustment parameters (adjusting the risk threshold based on the risk incidence rate in historical data), and confidence adjustment parameters (adjusting the confidence level of the risk score based on prediction uncertainty). The system considers various situational factors to differentiate parameters, such as livestock breed (different breeds have different sensitivities to risk factors), growth stage (such as differences in risk patterns during the rearing and lactation periods), farming model (such as differences in risk characteristics between intensive farming and free-range farming), and regional characteristics (such as differences in environmental risks in different climatic zones). These calibration parameters are not fixed. The system is designed with a regular update mechanism to adjust the calibration parameters based on newly accumulated data and changes in risk patterns to ensure that risk assessments can adapt to the dynamically changing risk environment.

[0169] Step F6, i.e., using the calibration parameters, corrects and adjusts the output of the risk assessment model. This step applies the previously set calibration parameters to the original output of the risk assessment model for calibration correction. Specifically, the system uses multiple calibration methods, including linear calibration (performing a linear transformation of the original risk score, such as r'=a*r+b, where a and b are calibration coefficients), probability calibration (calibrating the probability output using methods such as Platt scaling or equal bin averaging), and quantile calibration (adjusting the quantile mapping of the predicted distribution). The calibration process considers multiple aspects, such as bias correction (correcting systematic overestimation or underestimation), distribution adjustment (making the predicted risk distribution match the actual risk distribution), and confidence calibration (making the predicted confidence reflect the actual accuracy). The system evaluates the calibration effect through backtesting methods, comparing the prediction performance before and after calibration, such as the calibration plot, Brier score, and expected calibration error. This comprehensive calibration process ensures the accuracy and reliability of the risk assessment results, providing a scientific basis for subsequent risk grading and decision-making.

[0170] Then, step F7 is executed, constructing a tiered risk scoring system based on the calibrated assessment results. This tiered risk scoring system provides a standardized representation of risk assessment results, facilitating understanding and decision-making. Based on the calibrated risk assessment results, the system constructs a multi-tiered risk scoring system: first, an overall risk score, using a standard 0-100 scale to reflect the overall risk level of livestock; second, a categorical risk score, with separate scoring criteria for health risks, mortality risks, and production performance risks; and third, a detailed risk score, further refining specific types of risk (such as mastitis risk, respiratory disease risk, and heat stress risk). The system categorizes risk scores into multiple levels, such as low risk (0-20 points), medium-low risk (21-40 points), medium risk (41-60 points), medium-high risk (61-80 points), and high risk (81-100 points). Each level corresponds to a different risk management strategy and warning level. Furthermore, the system provides a risk trend score, which analyzes risk changes over a continuous time window to assess rising or falling risk trends, helping users predict the direction of risk development. This hierarchical risk scoring system makes complex risk information clear and easy to understand, making it easier for livestock managers and risk decision makers to quickly grasp the risk situation and formulate corresponding countermeasures.

[0171] Finally, step F8 is executed, that is, in the hierarchical risk scoring system, weight coefficients are assigned to different risk types and levels to form a comprehensive risk scoring standard. The importance and impact of different risk types and levels vary, and need to be balanced through reasonable weight distribution. The system assigns weights to risk types based on multiple factors: economic impact (considering the size of economic losses that may be caused by different risk types), risk persistence (considering the length of time the risk impact lasts), intervention feasibility (considering whether the risk can be alleviated through management measures) and risk frequency (considering the prevalence of risk events). The system uses the Analytic Hierarchy Process (AHP) and expert scoring method to determine the initial weights, and then adjusts and verifies them based on statistical analysis of historical data. For different risk levels under the same risk type, the system adopts a nonlinear weight distribution strategy. The weight growth rate of high-risk levels is faster than that of low-risk levels, reflecting the marginal effect of increased risk. The calculation formula for the comprehensive risk score is:

[0172] Risk comprehensive =∑(w i ×Risk i )

[0173] where w i is the weight of risk type i, Risk iThis is the risk score for this type of risk. The system regularly evaluates the effectiveness of the weighting configuration and dynamically adjusts it based on new data and feedback from risk management practices to ensure that the comprehensive risk score accurately reflects the actual risk situation and provides effective guidance for risk prevention and control decisions.

[0174] In one embodiment, the scoring calculation and differentiation threshold setting of the real-time monitoring data include:

[0175] G1. Build a real-time data processing mechanism based on the stream processing architecture;

[0176] G2. Set up a multi-level warning threshold strategy based on risk level classification;

[0177] G3. Calculate the rate of change and acceleration of risk scores based on real-time data;

[0178] G4. Select appropriate warning notification methods based on risk scoring and warning strategies;

[0179] G5. Based on the early warning results, record verification information and continuously optimize the early warning model.

[0180] In one embodiment of the present invention, when implementing risk warnings based on a calibrated risk scoring system, step G1 is first executed, which involves building a real-time data processing mechanism based on a stream processing architecture. Real-time data processing is the technical foundation of the risk warning system, requiring the system to continuously receive, process, and respond to large amounts of sensor data streams in a timely manner. The system utilizes a modern stream processing architecture, comprising a data access layer, a processing layer, and a storage layer. The data access layer supports multiple protocols (such as MQTT and HTTP) to receive data from IoT devices and perform preliminary data validation and format conversion. The processing layer utilizes a stream computing engine (such as Apache Kafka Streams and Apache Flink) to process incoming data streams with millisecond latency. The storage layer employs a tiered storage strategy, storing hot data (such as today's data) in an in-memory database for fast access, while cold data is transferred to a persistent storage system. The system implements a sliding window computing model, capable of calculating risk characteristics and assessing risk at different time scales (such as minutes, hours, and days). For example, for body temperature data, the system simultaneously calculates short-term fluctuations in a 15-minute window, medium-term trends in a 3-hour window, and diurnal cyclical changes in a 24-hour window. For sudden high-risk events (such as a sharp rise in body temperature, sudden abnormal activity, etc.), the system has designed a rapid response channel and uses a dedicated rule engine for real-time monitoring to ensure that key risk signals can be processed and responded to within seconds, greatly improving the system's response speed to acute health risks.

[0181] Next, step G2 is executed, which involves setting a multi-level warning threshold strategy based on the risk level classification. The scientific setting of warning thresholds is key to accurate risk warnings. The system designs a differentiated multi-level warning threshold system based on factors such as livestock breed, growth stage, breeding model, and seasonal environment. For the overall risk score, the system divides the risk level into five levels: low risk (0-20 points), medium-low risk (21-40 points), medium risk (41-60 points), medium-high risk (61-80 points), and high risk (81-100 points); for the categorized risk score and the segmented risk score, the system sets corresponding threshold standards. The system uses multiple threshold strategies: an absolute threshold strategy, which sets the warning threshold based on a fixed score value; a relative threshold strategy, which sets the threshold based on the degree of deviation from the individual historical baseline, such as triggering a warning when the threshold exceeds 2 times the standard deviation of the individual baseline; and a trend threshold strategy, which sets the threshold based on the rate and acceleration of change in the risk score, such as triggering a warning when the risk score increases by more than 10 points in a short period of time. The system also features a threshold combination strategy, combining different thresholds using logical operators to form compound conditions, such as "(body temperature > 39.5°C OR activity level decreased > 30%) AND feed intake decreased > 20%." This improves the accuracy of early warnings. Different warning levels correspond to different response mechanisms, and the system sets appropriate notification methods, response time requirements, and processing procedures for each level of warning to ensure the practicality and effectiveness of warnings.

[0182] Then, step G3 is executed, calculating the rate and acceleration of change of the risk score based on real-time data. The dynamic characteristics of the risk score are important for predicting risk trends. Based on time series analysis, the system calculates various risk score characteristics: the rate of change (first-order derivative), which indicates the speed of change of the risk score, is calculated as Rate(t) = [Score(t) - Score(t - Δt)] / Δt; the acceleration of change (second-order derivative), which indicates the increase or decrease in the risk rate of change, is calculated as Acceleration(t) = [Rate(t) - Rate(t - Δt)] / Δt; and the change pattern, which uses time series pattern matching to identify specific risk change patterns, such as step-like increases, exponential growth, and cyclical fluctuations. The system sets corresponding thresholds for these change characteristics, such as a risk score increase of more than 15 points within 24 hours or a risk acceleration that remains positive for more than 12 hours, to trigger trend warnings. The system pays special attention to unusual patterns of risk score changes. For example, if a chronic risk that typically changes slowly (such as poor nutritional status) rapidly deteriorates over a short period of time, the system will raise the alert level, warning of the possibility of acute complications. By analyzing the dynamic characteristics of risk scores, the system can achieve "predictive early warning," issuing warning signals before the risk actually reaches a high value, creating a valuable window of opportunity for risk intervention.

[0183] Next, proceed to step G4: selecting the appropriate warning notification method based on the risk score and warning strategy. Effective warning notification is crucial for timely disseminating risk information to relevant personnel. The system has designed a multi-channel, multi-tiered warning notification mechanism, selecting appropriate notification methods based on the urgency and importance of the risk. For low-risk warnings, the system primarily displays warning information within the system, color-coding individuals or areas at risk on the management interface. For medium-risk warnings, the system adds mobile app push notifications to alert managers to potential risks. For high-risk warnings, the system initiates multi-channel notifications, including SMS alerts, phone alerts, and even automated voice call systems, ensuring that key personnel are immediately informed of the risk. The system also supports tiered notification of warning information, delivering different levels of content and detail based on user roles and responsibilities. For example, frontline animal keepers receive detailed animal information and preliminary treatment recommendations; veterinarians receive more detailed health indicators and suspected cause analysis; and management receives a risk overview and potential economic impact assessment. The warning information contains key elements such as risk description, risk level, recommended measures and response time limit, and is presented in a clear and concise format to ensure that the recipient can quickly understand the situation and make decisions.

[0184] Finally, step G5 is executed: based on the warning results, verification information is recorded and the warning model is continuously optimized. The value of the warning system depends on its accuracy and reliability, and performance improvement requires continuous verification and optimization. The system has established a comprehensive warning verification and feedback mechanism, recording the actual results after each warning. After receiving the warning, the manager or veterinarian is required to conduct on-site verification and provide feedback, including confirmation of risk events (true positives), exclusion of risks (false positives), or detection of missed risks (false negatives). Based on this feedback, the system calculates multiple warning performance indicators: warning accuracy (the proportion of correct warnings), missed warning rate (the proportion of risk events that were not warned), false alarm rate (the proportion of false alarms), and average lead time (the interval between the warning time and the actual occurrence of the risk). The system uses machine learning methods (such as Bayesian optimization and reinforcement learning) to automatically adjust the parameters and threshold settings of the warning model based on these performance indicators. For example, for specific risk types with excessively high false alarm rates, the system may raise the warning threshold. For missed risk events, the system analyzes case characteristics and adjusts feature weights or adds new risk signal monitoring. The system also incorporates a manual review process, allowing professionals to modify rules for recurring false alarm patterns or adjust warning strategies for specific scenarios. Through this closed-loop optimization mechanism combining human and machine interaction, the warning system continuously learns and improves, adapting to the dynamically changing farming environment and risk patterns, and continuously improving the accuracy and practicality of warnings.

[0185] In one embodiment, generating intervention suggestions for risk quantification assessment, risk prevention and control decision-making, and loss assessment by establishing a risk-loss mapping relationship includes:

[0186] H1. Based on the risk assessment results, construct a correlation model between risk factors and loss extent;

[0187] H2. Calculate risk quantification indicators based on the association model;

[0188] H3. Generate risk prevention and control recommendations and intervention measures based on risk quantification indicators;

[0189] H4. Based on historical data, analyze the consistency between actual events and risk monitoring records to obtain analysis results;

[0190] H5. Based on the analysis results, generate targeted risk intervention recommendations.

[0191] In another embodiment of the present invention, when generating intervention recommendations based on risk assessment results, step H1 is first executed, that is, based on the risk assessment results, a correlation model between risk factors and the degree of loss is constructed. Accurately estimating the potential losses caused by risks is the core of risk quantitative assessment. Based on the risk assessment results and historical risk event data, the system constructs a correlation model between risk factors and the degree of loss. For health risks, the model takes into account the severity, duration, treatment costs and production losses of the disease; for mortality risks, the model analyzes the probability of death and livestock value under different risk levels; for production performance risks, the model evaluates the magnitude, duration and economic impact of output decline. The system implements multiple modeling methods: statistical regression models, which reveal the linear or nonlinear relationship between a single risk factor and loss; generalized linear models, which handle the combined effects of multiple factors; machine learning models (such as random forests, neural networks, etc.), which capture complex nonlinear relationships and interaction effects. These models together constitute a risk-loss mapping system that can predict potential economic losses based on the current risk status. For example, for a dairy cow that is detected with early symptoms of mastitis, the system can predict the possible losses under different intervention strategies: reduced milk production and the risk of worsening disease in the absence of intervention, and the treatment cost and recovery time in the case of timely treatment, thereby providing a quantitative basis for risk management decisions.

[0192] Next, step H2 is executed, which involves calculating risk quantification indicators based on the correlation model. Risk quantification indicators are numerical representations of risk assessment results and are used to support risk pricing, underwriting decisions, and risk management. Based on the risk-loss correlation model, the system calculates various risk quantification indicators: Expected Loss Value (ELV), which calculates the product of the probability of a risk event occurring and the magnitude of the loss, representing the mathematical expectation of risk; Value at Risk (VaR), which represents the maximum possible loss at a given confidence level (e.g., 95%); Conditional Value at Risk (CVaR), which calculates the expected value of losses exceeding VaR to better describe tail risk; and Risk Loss Ratio, which represents the ratio of risk-induced losses to livestock value. The system has designed specialized quantitative indicators for different risk types: For health risks, a health loss index is calculated, which comprehensively considers treatment costs, output losses, and recovery time; for mortality risks, a mortality risk value is calculated, which takes into account the probability of mortality and livestock value; and for production performance risks, an output risk index is calculated, which assesses the economic losses caused by reduced production. These quantitative risk indicators transform abstract risk assessment results into concrete numerical expressions, facilitating risk management decisions and economic analysis. The system presents these indicators in intuitive charts and data dashboards, helping users quickly grasp the economic significance of risks and prioritize them.

[0193] Then, step H3 is executed: generating risk prevention and control recommendations and intervention measures based on risk quantification indicators. Effective risk prevention and control recommendations are crucial for translating risk assessment into practical action. Based on risk quantification indicators and risk profile analysis, the system generates targeted risk prevention and control recommendations and intervention measures. The system includes a rich library of prevention and control strategies, with corresponding responses designed for different risk types. For health risks, it provides preventive measures (such as adjusting feed ratios and increasing exercise) and early intervention recommendations (such as specific examinations and preventive medications); for environmental risks, it provides environmental improvement recommendations (such as adjusting temperature and humidity and increasing ventilation) and emergency measures (such as extreme weather protection); and for management risks, it provides recommendations for optimizing operational procedures and adjusting processes. The system uses a cost-benefit analysis approach to evaluate the costs (such as implementation costs and time costs) and expected benefits (such as risk reduction and economic returns) of different intervention measures, selecting the most cost-effective intervention plan. Intervention recommendations adhere to the SMART principle (specific, measurable, achievable, relevant, and time-bound) to ensure their actionability. For example, to address heat stress risks during hot seasons, the system might generate a comprehensive intervention plan that includes specific cooling measures, implementation schedules, monitoring indicators, and effectiveness evaluation methods. The system can also adjust the priority and implementation difficulty of recommendations based on the farm's actual conditions and resource constraints, ensuring that the recommendations are feasible and can be implemented.

[0194] Next, step H4 is performed: analyzing the consistency between actual events and risk monitoring records based on historical data to obtain analysis results. Verifying the accuracy of risk monitoring and the effectiveness of risk prevention and control measures is key to continuously optimizing risk management. The system analyzes the consistency between historical risk event data and corresponding risk monitoring records to evaluate risk monitoring performance. The system calculates multiple consistency metrics: risk identification rate, which indicates the proportion of risk events successfully identified; risk identification lead time, which indicates the time interval between risk monitoring warnings and event occurrence; risk rating accuracy, which indicates the degree of match between risk level assessments and actual severity; and intervention effectiveness evaluation, which compares changes in risk indicators before and after intervention. The system uses a paired analysis method to match each historical risk event with the risk monitoring data from the corresponding period, identifying patterns in successful and unsuccessful monitoring cases. For successful monitoring cases, the system analyzes valid risk signals and key early warning indicators; for unsuccessful monitoring cases, the system analyzes missing risk signals and potential improvement areas. The system also analyzes the relationship between the implementation records of risk prevention and control measures and their actual results to assess the effectiveness of different intervention strategies. For example, the system analyzes the long-term implementation records of a risk reduction strategy for a particular disease and calculates the change in the disease incidence rate after the intervention to evaluate the strategy's effectiveness. This consistency analysis based on empirical data provides a scientific basis for the optimization of risk monitoring systems and prevention and control strategies.

[0195] Finally, step H5 is executed, generating targeted risk intervention recommendations based on the analysis results. Continuously optimized risk intervention recommendations are the ultimate outcome of risk management. Based on the consistent analysis results and the latest risk assessment data, the system generates more precise risk intervention recommendations. The generation of intervention recommendations follows three principles: targeting, which directly addresses the root causes of risk identified in the analysis; adaptability, which considers the actual conditions and resource constraints of the farm; and efficiency, which prioritizes intervention measures with the highest cost-effectiveness. The system supports multiple intervention recommendation types: emergency intervention recommendations, which provide immediate response measures to current high-risk situations; short-term intervention recommendations, which offer risk mitigation plans that can be implemented within 1-4 weeks; and long-term intervention recommendations, which provide structured risk prevention and control strategies and system improvement suggestions. These intervention recommendations cover multiple areas: livestock management measures (such as adjusting stocking density and optimizing feed formulations); health protection measures (such as vaccination programs and regular inspection schedules); environmental improvement measures (such as ventilation system upgrades and temperature and humidity control); and emergency response plans (such as extreme weather response and disease outbreak management). The system also provides implementation guidelines for intervention measures, including specific steps, critical control points, and effectiveness evaluation methods, to help users translate recommendations into practical action. Through this closed-loop intervention optimization mechanism, the system can continuously accumulate experience, improve the accuracy and effectiveness of risk intervention, and ultimately achieve a deep integration of risk assessment results and risk management processes, providing a comprehensive risk management solution for animal husbandry.

[0196] Based on the above embodiments, the present invention also provides an intelligent assessment device for livestock breeding risks based on multi-source data fusion. Figure 2 , Figure 2 FIG1 is a schematic diagram of the structure of an intelligent livestock breeding risk assessment device based on multi-source data fusion according to an embodiment of the present invention. Figure 2 As shown, the device includes:

[0197] The data collection and preprocessing module 110 is used to collect individual livestock vital sign data, breeding environment parameters, management behavior data, and risk-related historical data through multiple IoT devices, and clean, process outliers, and extract features from the multi-source data to obtain a standardized multi-source data set;

[0198] The risk feature extraction module 120 is used to extract individual-level health risk features, group-level disease risk features, and environmental-level risk features based on the standardized multi-source dataset, and fuse the risk features to form a multidimensional risk feature library;

[0199] The risk assessment model construction module 130 is used to construct differentiated risk assessment models for different livestock breeds and different growth stages based on the multi-dimensional risk feature library and generate model assessment results;

[0200] The risk calibration module 140 is used to analyze the correlation between risk factors and actual events based on risk-related historical data, calibrate the risk assessment model, and output a calibrated risk scoring system;

[0201] The risk classification and early warning module 150 is used to score and calculate the real-time monitoring data based on the calibrated risk scoring system, and set differentiated risk thresholds to achieve dynamic classification of livestock risk levels and early warning triggering, thereby obtaining graded risk early warning results;

[0202] The risk management module 160 is used to generate intervention suggestions for risk quantitative assessment, risk prevention and control decisions and loss assessment based on the graded risk warning results, so as to achieve the integration of risk assessment results and risk management processes.

[0203] In some embodiments, the risk feature extraction module 120 further includes:

[0204] A time series analysis unit is used to extract physiological rhythm characteristics based on livestock body temperature and activity data, behavioral characteristics based on feeding and rumination data, and state characteristics based on activity state data, and to construct a health baseline model;

[0205] Cluster analysis unit, which is used to identify abnormal patterns at the group level and extract group disease risk characteristics based on the collective behavior data of similar livestock groups within the farm;

[0206] A correlation analysis unit is used to extract abnormal characteristics, fluctuation characteristics and combined characteristics of environmental parameters based on the breeding environment monitoring data, and to analyze the lagged correlation between environmental parameters and livestock physiological and behavioral responses;

[0207] A management behavior analysis unit is used to extract characteristics of feeding management regularity, epidemic prevention management timeliness, and treatment response timeliness based on feeding management records and operation logs;

[0208] The feature fusion unit is used to integrate risk features at all levels into a multi-dimensional risk feature library through a combination of feature-level fusion and decision-level fusion.

[0209] In some embodiments, the risk assessment model building module 130 further includes:

[0210] A hierarchical model unit is used to establish a hierarchical model framework based on livestock breed characteristics and growth and development stages, setting differentiated normal ranges of physiological parameters and risk sensitivities for each breed-stage combination;

[0211] A health risk assessment unit is used to construct a health risk assessment sub-model based on a supervised learning method to output the health risk probability and risk type of individual livestock;

[0212] A death risk assessment unit is used to construct a death risk assessment sub-model based on the survival analysis method to predict the death probability and expected survival time of livestock under specific risk conditions;

[0213] Output risk assessment unit, which is used to build an output performance risk assessment sub-model based on regression method to predict the expected output level and fluctuation risk under current conditions;

[0214] The model integration unit is used to integrate the results of each sub-model through the model integration method and output a standardized comprehensive risk score and interpretable analysis.

[0215] In some embodiments, the risk calibration module 140 further includes:

[0216] Data collection unit, used to collect and organize historical risk event data and establish a data set that correlates risk events and risk factors;

[0217] Association analysis unit, used to apply Bayesian networks to analyze the causal relationship between risk factors and actual risk events, and identify key risk factors and their weights;

[0218] A coefficient calculation unit is used to calculate the calibration coefficient of the risk assessment model based on the results of the association analysis, and to set differentiated calibration parameters for different livestock breeds, breeding models and regions;

[0219] The performance evaluation unit is used to evaluate the predictive performance of the calibrated risk model through the backtesting method and optimize the model parameters based on the evaluation results.

[0220] In some embodiments, the risk classification warning module 150 further includes:

[0221] Data stream processing unit, used to build a real-time data processing mechanism based on the stream processing architecture, supporting feature calculation and risk assessment at different time scales;

[0222] The threshold setting unit is used to set a differentiated risk threshold system based on factors such as livestock breed and growth stage, and implement a multi-level early warning threshold strategy;

[0223] A dynamic assessment unit, which is used to calculate risk scores and their changing trends for individual livestock and groups based on real-time data streams and a calibrated risk scoring system;

[0224] An early warning trigger unit is used to intelligently trigger graded early warnings based on risk levels and changing trends, and send risk information through a multi-channel early warning notification mechanism;

[0225] The verification and feedback unit is used to record the actual results after the warning, analyze the warning performance indicators, and continuously optimize the warning model based on manual feedback.

[0226] In some embodiments, the risk management module 160 further includes:

[0227] The pricing support unit is used to build a correlation model between risk factors and expected loss rates based on risk assessment models and historical event analysis, providing data support for risk product pricing;

[0228] The underwriting decision-making unit is used to provide multi-dimensional assessment information such as livestock health status, breeding environment risks and management level for the risk underwriting process, and generate underwriting recommendations;

[0229] Verification support unit, used to analyze the consistency of claim applications with historical risk monitoring data, identify potential abnormal risks, and provide verification support to investigators;

[0230] The intervention recommendation unit is used to generate targeted risk intervention recommendations based on risk assessment results and recommend corresponding management measures and prevention and control strategies;

[0231] The closed-loop management unit is used to build a closed-loop system of risk services and aquaculture risk management, and promote the deep integration of risk management and risk prevention and control.

[0232] In the above-mentioned device, the data acquisition and preprocessing module 110 is used to implement step S101 in the method, the risk feature extraction module 120 is used to implement step S102 in the method, the risk assessment model construction module 130 is used to implement step S103 in the method, the risk calibration module 140 is used to implement step S104 in the method, the risk classification warning module 150 is used to implement step S105 in the method, and the risk management module 160 is used to implement step S106 in the method.

[0233] The intelligent livestock breeding risk assessment method and system based on multi-source data fusion provided by the embodiments of the present invention collects multi-source data through an Internet of Things (IoT) device network, extracts multi-level risk characteristics, and constructs a differentiated risk assessment model. This enables accurate risk assessment and graded early warning, providing scientific decision-making support for livestock breeding risk management. This invention, when applied to risk assessment and management in the livestock breeding industry, can help improve breeding efficiency, reduce economic losses, and promote the sustainable development of the livestock industry.

[0234] The present disclosure also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program executes the steps of the method for intelligent livestock breeding risk assessment based on multi-source data fusion described in the above-described method embodiment. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0235] In addition, an embodiment of the present disclosure also provides a computer program product, which stores a computer program. When the computer program is run by a processor, it executes the steps of the intelligent livestock breeding risk assessment method based on multi-source data fusion provided in any of the above embodiments of the present disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0236] The computer program product may be implemented in hardware, software, or a combination thereof. In one embodiment, the computer program product is implemented as a computer storage medium, which may be a volatile or non-volatile computer-readable storage medium. In another embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0237] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment and devices can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed equipment, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0238] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0239] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0240] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0241] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The scope of protection of the present disclosure is not limited thereto. Although the present disclosure has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-mentioned embodiments within the technical scope disclosed in the present disclosure, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. An intelligent assessment method for livestock breeding risk based on multi-source data fusion, characterized in that: include: Obtain livestock individual vital signs data, breeding environment parameters, management behavior data, and risk-related historical data collected by multiple IoT devices, and obtain a standardized multi-source data set through data cleaning, outlier processing, and feature extraction; Based on the standardized multi-source dataset, individual-level health risk characteristics are extracted through time series analysis, group-level disease risk characteristics are extracted through cluster analysis, and environmental-level risk characteristics are extracted through correlation analysis. A feature fusion algorithm is then used to form a multidimensional risk feature library. Based on the multidimensional risk signature database, a machine learning method is used to construct differentiated risk assessment models for different livestock breeds and different growth stages, and generate model assessment results; Based on the model assessment results and risk-related historical data, the correlation between risk factors and actual events is analyzed through Bayesian networks to calibrate the risk assessment model and output a calibrated risk scoring system; Based on the calibrated risk scoring system, by calculating scores and setting differentiated thresholds for real-time monitoring data, dynamic classification of livestock risk levels and early warning triggering are achieved, thereby obtaining graded risk early warning results; Based on the hierarchical risk warning results, by establishing a risk-loss mapping relationship, intervention suggestions are generated for risk quantification assessment, risk prevention and control decisions, and loss assessment, thereby realizing the integration of risk assessment results and risk management processes.

2. The method according to claim 1, characterized in that The risk characteristics at the environmental level are extracted through correlation analysis based on the standardized multi-source data set, including: Based on the standardized multi-source data set, a discriminant probability graphical model of the conditional neural field is constructed, and conditional probability calculation is performed on the multi-source sensor data to obtain an initial health state distribution; Based on the initial health state distribution, local spatiotemporal features are extracted through a 3D convolutional neural network, and the spatiotemporal features are analyzed for long-range dependencies using a self-attention mechanism to generate a spatiotemporal feature expression; Based on the spatiotemporal feature expression, the discriminant probability graphical model is optimized and trained by introducing the heat diffusion equation constraint and the energy conservation constraint as physical constraint regularization terms to obtain the optimized spatiotemporal health field; The optimized spatiotemporal health field is sampled multiple times by a Monte Carlo dropout method to construct a posterior probability distribution of health status and obtain a health risk representation with uncertainty estimation; Based on the health risk representation with uncertainty estimation, feature fusion is performed through an attention weighting mechanism to generate an enhanced risk feature representation.

3. The method according to claim 1, characterized in that Based on the hierarchical risk warning results, by establishing a risk-loss mapping relationship, intervention suggestions are generated for risk quantitative assessment, risk prevention and control decision-making, and loss assessment, including: Based on the graded risk warning results, a context vector is constructed containing basic information of the cattle, physiological indicators, nutritional parameters and environmental factors, and an evaluation function is constructed according to treatment effects, production performance changes and resource consumption to generate basic decision-making data; Based on the decision-making basic data, a Q function network is constructed through a four-layer neural network, and a time difference learning method and a prediction model are used to deal with the delayed feedback problem to generate intervention evaluation indicators; Based on the intervention evaluation metric, a variant of Thompson sampling is used to maintain the posterior distribution of the action-context combination, and a historical influence attenuation factor is introduced to adjust the exploration strategy and generate dynamic decision parameters. Based on the dynamic decision parameters, a hierarchical decision structure is constructed through a two-stage decision method to classify and select intervention measures and generate a specific intervention plan, wherein the specific intervention plan includes a number of intervention suggestions; Based on the specific intervention plan, a hybrid decision-making model is achieved by combining rule guidance and data-driven through a progressive weight adjustment mechanism.

4. The method according to claim 1, wherein The method of extracting disease risk characteristics at the population level through cluster analysis based on the standardized multi-source dataset includes: Based on the standardized multi-source dataset, a multi-layer heterogeneous risk association graph including livestock nodes, environmental nodes, and management behavior nodes is constructed, and edge connections between nodes are established by analyzing spatial proximity and association relationships to generate an initial graph structure; Based on the initial graph structure, by designing node feature reconstruction, edge prediction and subgraph attribute prediction tasks, a self-supervised analysis of the data structure is performed to generate a sample set for training livestock breeding risks; Based on the training sample set, the mutual information of node representation is decomposed into feature transfer information, structural common information and structural independent information through information theory method, and a balanced loss function is constructed; Based on the loss function, information is transferred between different types of nodes and edges through the heterogeneous graph attention network, and the weight parameters are dynamically adjusted according to the characteristics of the graph structure to obtain an optimized graph representation; Based on the optimized graph representation, a classifier is used to identify disease transmission paths and generate risk assessment results.

5. The method according to claim 1, wherein The multiple IoT devices are used to collect the individual vital signs data of the livestock, including biochip ear tags and NB chip ear tags for collecting body temperature, activity level, number of ruminations and lying time, as well as environmental monitoring equipment for collecting temperature, humidity, ammonia concentration and light, wherein the management behavior data includes feeding frequency, feed ratio, immunization records and treatment records.

6. The method according to claim 1, characterized in that The individual-level health risk characteristics are extracted through time series analysis based on the standardized multi-source dataset, including: Based on body temperature and activity data, physiological rhythm characteristics are extracted through time series analysis; Based on feeding and rumination data, behavioral characteristics are extracted through behavioral pattern analysis; Based on the activity state data, state features are extracted through state transition analysis; A health baseline model is constructed based on the physiological rhythm characteristics, behavioral characteristics and state characteristics.

7. The method according to claim 1, characterized in that Based on the multi-dimensional risk signature library, a differentiated risk assessment model is constructed for different livestock breeds and different growth stages using a machine learning method, including: Construct a health risk assessment sub-model based on supervised learning algorithms; Construct a mortality risk assessment sub-model based on survival analysis method; Construct a production performance risk assessment sub-model based on regression analysis method; Based on the ensemble learning method, the results of each sub-model are integrated to generate a risk assessment model.

8. The method according to claim 1, characterized in that The risk assessment model is calibrated by analyzing the correlation between risk factors and actual events through the Bayesian network, and the calibrated risk scoring system is outputted, including: Based on historical data, collect and organize information on risk event types, loss levels, and timing; Based on the risk event information, construct a correlation data set between risk factors and actual events; Based on the association data set, applying Bayesian network to perform causal association analysis; Based on the results of causal association analysis, calculate the intensity of the risk factor's influence on the occurrence of the event; Setting differentiated calibration parameters based on the intensity of the conditional influence; Using the calibration parameters to correct and adjust the output results of the risk assessment model; Based on the calibrated assessment results, a tiered risk scoring system is constructed; In the hierarchical risk scoring system, weight coefficients are assigned to different risk types and levels to form a comprehensive risk scoring standard.

9. The method according to claim 1, characterized in that The scoring calculation and differentiated threshold setting of real-time monitoring data include: Build a real-time data processing mechanism based on stream processing architecture; Based on risk level classification, set up multi-level warning threshold strategies; Calculate the rate of change and acceleration of risk scores based on real-time data; Select the appropriate warning notification method based on risk scoring and warning strategy; Based on the early warning results, record verification information and continuously optimize the early warning model.

10. The method according to claim 1, characterized in that The aforementioned generation of intervention suggestions for risk quantification assessment, risk prevention and control decision-making, and loss assessment by establishing a risk-loss mapping relationship includes: Based on the risk assessment results, a correlation model between risk factors and loss extent is constructed; Calculating risk quantification indicators based on the association model; Generate risk prevention and control suggestions and intervention measures based on risk quantification indicators; Based on historical data, analyze the consistency between actual events and risk monitoring records to obtain analysis results; Based on the analysis results, targeted risk intervention recommendations are generated.

11. An intelligent livestock breeding risk assessment device based on multi-source data fusion, characterized in that: include: A data collection and preprocessing module is used to collect livestock individual vital sign data, breeding environment parameters, management behavior data, and risk-related historical data through multiple IoT devices, and clean, process outliers, and extract features from the multi-source data to obtain a standardized multi-source data set; A risk feature extraction module is used to extract individual-level health risk features, group-level disease risk features, and environmental-level risk features based on the standardized multi-source data set, and to fuse the risk features to form a multidimensional risk feature library; A risk assessment model construction module is used to construct differentiated risk assessment models for different livestock breeds and different growth stages based on the multidimensional risk feature library and generate model assessment results; A risk calibration module is used to analyze the correlation between risk factors and actual events based on risk-related historical data, calibrate the risk assessment model, and output a calibrated risk scoring system; A risk grading and early warning module is used to score and calculate real-time monitoring data based on the calibrated risk scoring system, and set differentiated risk thresholds to achieve dynamic classification and early warning triggering of livestock risk levels, thereby obtaining graded risk early warning results; The risk management module is used to generate intervention suggestions for risk quantification assessment, risk prevention and control decisions and loss assessment based on the hierarchical risk warning results, so as to achieve the integration of risk assessment results and risk management processes.

Images