A dynamic prevention and control method and system for chicken coccidiosis

Abstract

The application relates to the technical field of animal disease prevention and control, and discloses a dynamic prevention and control method and system for chicken coccidiosis, which comprises the following steps: step 1: acquiring environmental parameter data, excrement hyperspectral data and chicken group behavior image data; step 2: processing the environmental parameter data, the excrement hyperspectral data and the chicken group behavior image data to obtain a micro-ecological state tensor; step 3: performing dynamic embedding and reasoning by using a graph attention network to obtain a pathogenicity coefficient of a pathogen, a host susceptibility index and a drug resistance evolution factor; step 4: inputting the pathogenicity coefficient of the pathogen, the host susceptibility index and the drug resistance evolution factor into a digital twin model to output an intestinal lesion score value, a drug resistance risk value and a predicted coccidial oocyst density value; and step 5: inputting the intestinal lesion score value, the drug resistance risk value and the predicted coccidial oocyst density value into a machine learning algorithm to generate a prevention and control strategy.

Description

Technical Field

[0001] This application belongs to the field of animal disease prevention and control technology, and more specifically, relates to a dynamic prevention and control method and system for chicken coccidiosis. Background Technology

[0002] Coccidiosis is a highly destructive parasitic disease affecting the global poultry industry, caused by protozoa of the genus Eimeria that parasitize the intestinal epithelial cells of chickens. It has a high morbidity rate, primarily causing bloody diarrhea, stunted growth, and even death in chicks. Even surviving chickens may experience a significant decrease in feed conversion rate due to intestinal damage.

[0003] Current technologies for the prevention and control of coccidiosis in chickens mainly suffer from the following challenges:

[0004] 1. Monitoring lag: Traditional monitoring methods rely on manually collecting fresh feces for microscopic oocyst counting (OPG) or observing intestinal lesion scores (Lesion Score) by necropsy of dead chickens. These methods are all "post-hoc" diagnoses, and often by the time they are discovered, coccidia have already completed multiple reproductive cycles in the flock, causing irreversible intestinal damage.

[0005] 2. Drug resistance problem: Long-term, indiscriminate use of ionophore-based or chemically synthesized anticoccidial drugs (such as monensin and diclazuril) makes coccidia highly susceptible to drug resistance. The lack of scientifically data-driven guidance on drug rotation or shuttle therapy means that the speed of new drug development lags far behind the rate of drug resistance development.

[0006] 3. Neglecting microecological interactions: Existing prevention and control systems often focus only on killing coccidia themselves, while neglecting the synergistic pathogenic effects of gut microbiota (such as Clostridium perfringens) and coccidia (e.g., coccidiosis-induced necrotizing enterocolitis). There is a lack of a systematic perspective to comprehensively consider the complex dynamic relationships between the environment, host immunity, microbiota, and pathogens.

[0007] 4. Lack of pre-planning for decision-making: Before using a certain prevention and control plan, farms cannot predict the specific effects of the plan under the current flock status and environmental conditions, and often can only rely on experience to "trial and error".

[0008] Therefore, the technical problem addressed by this application is: how to achieve precise prevention and control of coccidiosis in chickens. Summary of the Invention

[0009] The main purpose of this application is to provide a dynamic prevention and control method for chicken coccidiosis. It constructs a microecological state tensor of multi-source data, then uses a graph attention network to embed and reason about the pathological knowledge graph, and combines digital twin technology to realize the quantitative simulation of the dynamic game process among coccidia, microbiota, and immunity, thereby achieving precise prevention and control of chicken coccidiosis.

[0010] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0011] A dynamic control method for coccidiosis in chickens includes the following steps:

[0012] Step 1: Acquire environmental parameter data, excrement hyperspectral data, and flock behavior image data;

[0013] Step 2: Analyze and stitch together the environmental parameter data, excrement hyperspectral data and chicken behavior image data to obtain the microecological state tensor, and then perform time series expansion on the microecological state tensor to obtain the time series tensor;

[0014] Step 3: Construct a pathological knowledge graph. Based on the microecological state tensor, use a graph attention network to dynamically embed and reason about the pathological knowledge graph to obtain pathogen pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability, and drug resistance evolution factor.

[0015] Step 4: Input the microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor into the digital twin model. The digital twin model outputs the intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value.

[0016] Step 5: Input the intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value into the machine learning algorithm, and the machine learning algorithm generates a prevention and control strategy.

[0017] Preferably, step 2 includes the following sub-steps:

[0018] Step A1: Perform Min-Max normalization on the environmental parameter data to obtain the environmental feature vector;

[0019] Step A2: Based on the excrement hyperspectral data, the coccidia oocyst index and hemorrhage index are calculated using the formulas for the coccidia oocyst index and hemorrhage index. The coccidia oocyst index and hemorrhage index are then input into a pre-trained convolutional neural network regression model to obtain the current coccidia oocyst density value. The current coccidia oocyst density value is then log-normalized to obtain an estimated coccidia density value. Based on the coccidia oocyst index, hemorrhage index, and estimated coccidia density value, a spectral feature vector is constructed.

[0020] Step A3: Analyze the chicken flock behavior image data using a target detection algorithm and perform ratio normalization to obtain the proportions of feeding activity, movement activity, and resting time. Based on the proportions of feeding activity, movement activity, and resting time, construct a behavioral feature vector.

[0021] Step A4: Concatenate the environmental feature vector, spectral feature vector, and behavioral feature vector to obtain the micro-ecological state tensor;

[0022] Step A5: Perform time series expansion on the microecological state tensor, and construct a time series tensor based on the microecological state tensor within 24 hours.

[0023] Preferably, in step A2, the formula for the coccidia oocyst index is:

[0024] ;

[0025] in, , These are the characteristic reflectance wavelengths of lipids and proteins in the coccidia oocyst wall. The background interference wavelength is R, and the reflectivity is R.

[0026] The formula for the bleeding index is:

[0027] ;

[0028] in, For wavelength, wavelength The first derivative of the reflectivity at that point, For wavelength Accumulate points. , These are the start and end wavelengths of the characteristic absorption band of oxyhemoglobin, respectively.

[0029] Preferably, in step 3, the use of graph attention networks to dynamically embed and reason about the pathological knowledge graph to obtain the pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability, and drug resistance evolution factor includes the following sub-steps:

[0030] Step B1: Construct a heterogeneous graph, which includes environmental entity nodes, pathogen entity nodes, host entity nodes, microbial community entity nodes, and drug entity nodes;

[0031] Step B2: Calculate the attention coefficients between environmental entity nodes, pathogen entity nodes, host entity nodes, microbial entity nodes, and drug entity nodes using a graph attention network, and then normalize them using the Softmax function to obtain the attention weights.

[0032] Step B3: Update the node feature vectors by weighted summation based on attention weights;

[0033] Step B4: The graph attention network maps the updated node feature vectors through a fully connected layer to obtain the pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability, and drug resistance evolution factor.

[0034] Preferably, the formula for calculating the attention coefficient is:

[0035] ;

[0036] in, and These are feature vectors for environmental entity nodes, pathogen entity nodes, host entity nodes, microbial community entity nodes, or drug entity nodes, respectively. It is a non-linear activation function. To share the weight matrix, Let || be the attention weight vector, and || denotes the concatenation operation;

[0037] The formula for calculating attention weights is:

[0038] ;

[0039] in, For nodes The set of neighboring nodes, For the natural constant An exponential function with base 0;

[0040] The formula for updating the node feature vector using weighted summation is as follows:

[0041] ;

[0042] in, This is the updated node feature vector. It is a non-linear activation function. For attention weights, To share the weight matrix, For nodes eigenvectors, For nodes The set of neighboring nodes.

[0043] Preferably, the digital twin model includes a coccidia-microbiota-immune cell interaction sub-model and an intestinal substance transport sub-model based on computational fluid dynamics;

[0044] Step 4 includes the following sub-steps:

[0045] Step C1: Use the intestinal substance transport submodel to determine the real-time concentration distribution of the drug in the intestine;

[0046] Step C2: Based on the real-time concentration distribution of drugs in the gut, the microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor are input into the coccidia-microbe-immune cell interaction sub-model to obtain the coccidia biomass evolution curve, immune intensity evolution curve and microbe abundance evolution curve.

[0047] Step C3: Based on the evolution curves of coccidia biomass, immune intensity, and microbial abundance, the evolution curve of intestinal lesion score is calculated using the intestinal lesion score mapping formula.

[0048] Step C4: Based on the coccidia biomass evolution curve and historical drug use records, the drug resistance risk calculation formula is used to calculate the drug resistance development trend;

[0049] Step C5: Based on the coccidia biomass evolution curve, intestinal lesion score evolution curve, and drug resistance development trend, obtain the intestinal lesion score, drug resistance risk value, and predicted coccidia oocyst density value.

[0050] Preferably, the coccidia-microbiota-immune cell interaction sub-model includes:

[0051] The coccidia reproduction equation is as follows:

[0052] ;

[0053] in, For the first The endophytic reproduction rate of coccidia For the first Biomass of coccidia in the gut The environmental carrying capacity of the gut to coccidia. The first intestinal Biomass of coccidia. For the immune system to the first The lethality coefficient of coccidia, For the strength of the host's local immunity, For the first species of bacteria on the first Antagonistic or synergistic coefficient of coccidia, For the first The abundance of dominant gut microbiota This represents the killing or inhibition coefficient of the drug against coccidia. This represents the real-time concentration distribution of drugs within the intestines.

[0054] The immune response equation is as follows:

[0055] ;

[0056] in, This represents the response coefficient of the immune system to coccidia stimulation. For the first Biomass of coccidia in the gut This is the natural decay coefficient of immune strength. This refers to the strength of the host's local immune response;

[0057] The microbial community succession equation is as follows:

[0058] ;

[0059] in, For the first The endophytic proliferation rate of the microbial community, For the first The abundance of dominant gut microbiota For the gut microbiota's environmental carrying capacity, For the first Coccidia against the first Influence coefficient of microbial community For the first Biomass of coccidia in the gut This represents the coefficient of drug's killing or inhibition of bacterial flora. This represents the real-time concentration distribution of drugs within the intestines.

[0060] The intestinal lesion scoring mapping formula is:

[0061] ;

[0062] in, This is the Sigmoid function, with coefficient 4 being a calibration factor. This is the normalized value of coccidia biomass relative to environmental carrying capacity. For immune-pathogen interaction terms, The normalized value for the abundance of Clostridium perfringens is given. , , All are weighting coefficients;

[0063] The formula for calculating drug resistance risk is:

[0064] ;

[0065] in, As a factor in the evolution of drug resistance, The resistance variation coefficient driven by population size. For reference population size, The pressure-driven cumulative resistance coefficient for drug selection. The time integral of drug concentration, This represents the heritability coefficient of drug resistance based on historical drug use records. This is the historical drug use intensity index. For time The first time Coccidia biomass.

[0066] Preferably, in step 5, the machine learning algorithm is a proximal policy optimization (PPO) algorithm, a genetic algorithm, a particle swarm optimization (PSO) algorithm, or a simulated annealing algorithm.

[0067] Preferably, in step 5, a multi-objective optimization function is used to optimize and train the machine learning algorithm;

[0068] The multi-objective optimization function is:

[0069] ;

[0070] in, To predict daily weight gain, The intestinal lesion score, For economic costs, This represents the risk value for drug resistance. , , and All are weighting coefficients.

[0071] It should be noted that:

[0072] Pathogenicity coefficient: refers to the relative intensity coefficient of intestinal lesions caused by coccidia pathogens under the current microecological conditions, with a baseline value of 1.0.

[0073] Host susceptibility index: refers to the susceptibility of a flock of chickens to coccidiosis infection under their current physiological state. It is dimensionless and ranges from [0, 1]. This index comprehensively reflects the host's immune function, intestinal barrier integrity, and nutritional status. This indicates complete resistance (theoretical state). It indicates extreme susceptibility.

[0074] Intestinal lesion score: This is an index used to quantify the degree of pathological damage caused by coccidiosis in the chicken intestine, based on the criteria proposed by Johnson and Reid (1970). The score range is [0, 4], from 0 (no obvious lesions) to 4 (severe, extensive lesions).

[0075] In addition, a dynamic control system for chicken coccidiosis is also provided to implement the above-mentioned dynamic control method for chicken coccidiosis, comprising the following units:

[0076] Data acquisition module: used to acquire environmental parameter data, excrement hyperspectral data, and flock behavior image data;

[0077] State Tensor Construction Module: Used to analyze, process, and stitch together environmental parameter data, excrement hyperspectral data, and chicken flock behavior image data to obtain a microecological state tensor. Then, the microecological state tensor is extended by time series to obtain a time series tensor.

[0078] Knowledge Computation Module: Used to construct a pathological knowledge graph. Based on the microecological state tensor, a graph attention network is used to dynamically embed and reason about the pathological knowledge graph to obtain pathogen pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability and drug resistance evolution factor.

[0079] Digital twin simulation module: used to input microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor into digital twin model. The digital twin model outputs intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value.

[0080] Intelligent decision-making module: This module is used to input intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value into machine learning algorithm, which generates prevention and control strategy.

[0081] One of the above-mentioned technical solutions in this application has at least one of the following advantages or beneficial effects:

[0082] The prevention and control method proposed in this application constructs a microecological state tensor, enabling subsequent calculations to comprehensively consider the complex dynamic relationships between the environment, host immunity, microbiota, and pathogens. By combining a pathological knowledge graph with a graph attention network, it achieves dynamic weight allocation and interpretable reasoning for pathogenic risk factors, overcoming the logical bias problem of pure data-driven methods under small sample conditions. Using digital twin technology, it realizes quantitative simulation of the dynamic game process among coccidia, microbiota, and immunity, breaking through the limitations of existing technologies that separate coccidia control from intestinal microbiota management. The microecological state tensor combined with a convolutional neural network regression model enables coccidia species classification and quantification, shifting the traditional "post-infection" diagnosis to the subclinical infection stage. Attached Figure Description

[0083] The present application will be further described below with reference to the accompanying drawings and embodiments;

[0084] Figure 1 This is a flowchart of the dynamic prevention and control method for coccidiosis in chickens in Example 1;

[0085] Figure 2 This is a block diagram of the dynamic prevention and control system for chicken coccidiosis in Example 2. Detailed Implementation

[0086] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0087] The following disclosure provides many different implementation methods or examples for different schemes of implementing this application.

[0088] Example 1

[0089] refer to Figure 1 A dynamic control method for chicken coccidiosis includes the following steps:

[0090] Step 1: Acquire environmental parameter data, excrement hyperspectral data, and flock behavior image data;

[0091] In this embodiment, environmental parameters (temperature, humidity, ammonia concentration, and litter moisture content) within the chicken coop are acquired using appropriate sensors. Assume that in a particular monitoring session, the sensors collected the following environmental parameters: temperature... (Within a suitable range), humidity (Above 70% threshold indicates high humidity stress), ammonia concentration (Within normal range), bedding moisture content (The level has increased significantly due to continuous rain; the normal range is 20%-30%).

[0092] Hyperspectral data of excrement, i.e., the hyperspectral reflectance curve of excrement, is obtained by a hyperspectral imager. .

[0093] High-definition cameras deployed on the top of the chicken coop collect image data of the chickens' behavior.

[0094] Step 2: Analyze and stitch together the environmental parameter data, excrement hyperspectral data and chicken behavior image data to obtain the microecological state tensor, and then perform time series expansion on the microecological state tensor to obtain the time series tensor;

[0095] Preferably, step 2 includes the following sub-steps:

[0096] Step A1: Perform Min-Max normalization on the environmental parameter data to obtain the environmental feature vector;

[0097] In this embodiment, the formula for Min-Max normalization is: ,in, The normalized values ​​are temperature, humidity, ammonia concentration, or bedding moisture content. The factors include temperature, humidity, ammonia concentration, or moisture content of the bedding material. This represents the minimum value of temperature, humidity, ammonia concentration, or bedding moisture content. This represents the maximum value of temperature, humidity, ammonia concentration, or moisture content of the bedding material.

[0098] Temperature Input into the Min-Max normalization formula, assuming the normal temperature range is... The normalized temperature is ;

[0099] Humidity Inputting this into the Min-Max normalization formula, assuming the normal humidity range is [40%, 90%], the normalized humidity will be: ;

[0100] ammonia concentration Inputting this into the Min-Max normalization formula, assuming the normal range for ammonia concentration is [0 ppm, 25 ppm], the normalized ammonia concentration will be... ;

[0101] Moisture content of bedding material Inputting this into the Min-Max normalization formula, assuming the normal range for bedding moisture content is [10%, 60%], the normalized bedding moisture content will be... ;

[0102] Based on the normalized environmental parameter data, an environmental feature vector is constructed. .

[0103] Step A2: Based on the excrement hyperspectral data, the coccidia oocyst index and hemorrhage index are calculated using the formulas for the coccidia oocyst index and hemorrhage index. The coccidia oocyst index and hemorrhage index are then input into a pre-trained convolutional neural network regression model to obtain the current coccidia oocyst density value. The current coccidia oocyst density value is then log-normalized to obtain an estimated coccidia density value. Based on the coccidia oocyst index, hemorrhage index, and estimated coccidia density value, a spectral feature vector is constructed.

[0104] In this embodiment, a characteristic wavelength was selected based on the lipid characteristics of the coccidia oocyst wall. and and background wavelength The hyperspectral data of excrement is the hyperspectral reflectance curve of excrement. Assuming the collected real-time reflectivity data for the corresponding characteristic wavelength is The real-time reflectance data is input into the coccidial oocyst index formula to calculate the coccidial oocyst index, specifically:

[0105] ;

[0106] in, , These are the characteristic reflectance wavelengths of lipids and proteins in the coccidia oocyst wall. The background interference wavelength is R, and the reflectivity is R.

[0107] Hyperspectral reflectance curves in the 540nm-580nm wavelength range (characteristic absorption region of hemoglobin) The first derivative is calculated to obtain the derivative spectrum. Then, the absolute value of the derivative is integrated over this band to calculate the hemorrhage index, specifically:

[0108] ;

[0109] in, For wavelength, wavelength The first derivative of the reflectivity at that point, For wavelength Accumulate points. , These are the start and end wavelengths of the characteristic absorption band of oxyhemoglobin, respectively.

[0110] The bleeding index can sensitively reflect the level of early-stage microbleeding in the intestine. Assuming the example value is calculated through numerical integration... (This indicates the presence of mild intestinal damage).

[0111] Then count the coccidia oocysts. and bleeding index Input a pre-trained convolutional neural network (CNN) regression model, output the current coccidia oocyst density value at the current time step t, and then perform logarithmic normalization on the current coccidia oocyst density value to obtain the estimated coccidia density value; the logarithmic normalization formula is:

[0112] ;

[0113] in, This represents the current coccidia oocyst density value. This represents the upper limit of normal values ​​for coccidia oocyst density.

[0114] Assume the current coccidia oocyst density output by the convolutional neural network regression model is... CFU / g (at early infection levels), this value is the sum of the current coccidia oocyst density values ​​for the three coccidia species. The normal upper limit for coccidia oocyst density is [missing value]. The specific calculation process for the estimated coccidia density is as follows: (number per gram)

[0115] ;

[0116] Based on the coccidia oocyst index, hemorrhage index, and estimated coccidia density, a spectral feature vector is constructed. .

[0117] It should be noted that the specific construction and training process of the Convolutional Neural Network (CNN) regression model is as follows:

[0118] In terms of model architecture, the convolutional neural network regression model adopts a one-dimensional convolutional neural network structure, including an input layer, a feature extraction layer, a feature fusion layer, and a regression output layer. The input layer receives two inputs: the first is the hyperspectral reflectance curve. The first path is the spectral vector obtained by equal-interval sampling within the 400nm-1000nm band (200-dimensional, corresponding to one sampling point every 3nm); the second path is the calculated physical characteristic index vector. (2D). The feature extraction layer consists of three cascaded one-dimensional convolutional blocks. Each convolutional block contains a one-dimensional convolutional layer (kernel sizes of 7, 5, and 3, with output channels of 32, 64, and 128 respectively), a batch normalization layer (BatchNorm1d), and a ReLU activation function. Each convolutional block is followed by a max pooling layer (kernel size of 2). The output of the feature extraction layer is then subjected to global average pooling to obtain a 128-dimensional depth feature vector. The feature fusion layer concatenates the 128-dimensional depth feature vector with a 2-dimensional physical feature exponent vector to obtain a 130-dimensional fused feature vector. The regression output layer consists of two fully connected layers. The first fully connected layer maps 130 dimensions to 64 dimensions (followed by ReLU activation and Dropout, with a dropout rate of 0.3). The second fully connected layer maps 64 dimensions to a 3-dimensional output, corresponding to the current coccidia oocyst density values ​​of the three main coccidia species (Eimeria tenella, Eimeria toxicae, and Eimeria smotherum).

[0119] For the construction of the training dataset, the training data came from actual excrement samples. Each sample underwent both hyperspectral imaging and laboratory microscopic oocyst counting (McMaster method). Hyperspectral imaging obtained the reflectance curves of the excrement in the 400nm-1000nm band as model input, while the OPG values ​​of each coccidia species obtained by the laboratory McMaster method served as supervision labels. The training dataset contained 5000 labeled samples, divided into a training set (3500 samples), a validation set (750 samples), and a test set (750 samples) in a 7:1.5:1.5 ratio. To enhance the model's generalization ability, data augmentation was performed on the training data, including adding Gaussian noise (mean 0, standard deviation 0.02), random band shift (range ±2nm), and random baseline drift to the spectral curves.

[0120] In the training process, the Adam optimizer was used with an initial learning rate of 0.001. A cosine annealing strategy was employed to dynamically adjust the learning rate, with a minimum learning rate of 0.00001. The batch size was set to 64, and the number of training epochs was set to 200. The mean squared error loss (MSE Loss) was used as the loss function. Where N is the number of samples and c is the coccidia species index. This represents the oocyst density value of the c-th type of coccidia in the n-th sample predicted by the model. This represents the actual oocyst density value of the c-th type of coccidia in the n-th sample obtained from actual detection (i.e., laboratory microscopic counting). An early stopping strategy is employed, stopping training when the validation set loss does not decrease for 20 consecutive epochs. On the test set, the model's determination coefficient... The mean square error (RMSLE) was 0.89, and the root mean square logarithmic error (RMSLE) was 0.35, indicating that the model has good prediction accuracy.

[0121] OPG represents the density of coccidia oocysts.

[0122] This method allows for the acquisition of hyperspectral reflectance curves, capturing microscopic biochemical features invisible to the human eye. It also enables the construction of a coccidia oocyst index. The weak absorption signals of lipids and proteins in the ovum sac wall were amplified by using logarithmic difference ratio calculations, eliminating background interference from uneven illumination. A hemorrhage index was constructed. The level of trace occult blood was quantified using derivative spectral integrals, which can sensitively reflect early pathological damage in the intestine. A convolutional neural network regression model was used to fuse physical feature indices and depth features, enabling accurate typing and quantification of different coccidia species, providing data support for subsequent precision medication.

[0123] Step A3: Analyze the chicken flock behavior image data using a target detection algorithm and perform ratio normalization to obtain the proportions of feeding activity, movement activity, and resting time. Based on the proportions of feeding activity, movement activity, and resting time, construct a behavioral feature vector.

[0124] In this embodiment, a target detection algorithm (such as YOLO v5) is used to identify the position and posture of a single chicken, and the number of chickens and their dwell time in the feeding area, drinking area, and resting area are counted to calculate the behavioral feature vector. All three behavioral characteristic indicators were normalized using a ratio normalization method, as detailed below:

[0125] feeding activity Defined as the ratio of the number of chickens exhibiting foraging behavior within a unit time window (e.g., per hour) to the total number of chickens in the flock. ,in, This represents the number of chickens identified by the target detection algorithm that are in the foraging area and have their heads down in a pecking posture. The total number of chickens is represented by the index, which ranges from [0,1].

[0126] Exercise activity Defined as the ratio of the average displacement distance of the flock within a unit time window to the normal movement reference distance of the flock at that age, i.e. ,in, The average displacement distance of all tracked chickens within the time window (calculated through target association between consecutive frames). This is the average reference distance for movement of a healthy flock of chickens of this age under normal conditions (obtained from historical data statistics). The value range of this indicator is [0, 1]. If it exceeds 1, it is truncated to 1.

[0127] percentage of rest time Defined as the ratio of the cumulative duration of the flock remaining stationary (continuously stationary for more than 5 minutes) within a unit time window to the total length of the observation time window, i.e. ,in, To accumulate the duration of stillness, The value of this indicator is the length of the time window, and its range is [0, 1].

[0128] Suppose that in a certain monitoring session, the behavioral characteristic value calculated by the system is: feeding activity level. (Indicating that 72% of the chickens are in a normal feeding state), activity level (This indicates that the average activity level of the flock is 48% of the normal reference value, which is at a moderate level), and the percentage of rest time. (This indicates that the flock was at rest for 25% of the observation period).

[0129] A behavioral feature vector is constructed based on the proportions of feeding activity, movement activity, and rest duration. .

[0130] Step A4: Concatenate the environmental feature vector, spectral feature vector, and behavioral feature vector to obtain the microecological state tensor. ;

[0131] Step A5: Perform time series expansion on the microecological state tensor, and construct a time series tensor based on the microecological state tensor within 24 hours.

[0132] To capture the dynamic evolution trend of the micro-ecological state, the system also stores the historical micro-ecological state tensor of the past 24 hours. , constitute time series tensor .

[0133] Step 3: Construct a pathological knowledge graph containing inherent physiological and pathological mechanisms. Based on the currently acquired microecological state tensor, extract local correlation subgraphs. Use graph attention network to dynamically embed and reason about the pathological knowledge graph to obtain pathogen pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability and drug resistance evolution factor.

[0134] When constructing a pathological knowledge graph, the first step is to set up an "ontology layer (Schema)" and a "static knowledge base". Specifically, the following needs to be pre-inputted: (1) Inherent concepts and entities: For example, all immune factors (sIgA, cytokines), anatomical structures (intestinal barrier, intestinal epithelial cells), and pathogen species database (various coccidia) are entered into the system as basic nodes. (2) Inherent biological relationships (edges): For example, pre-defined rules: "host -> secretion -> sIgA", "sIgA -> inhibition -> Eimeria tenella", "Clostridium perfringens -> destruction -> intestinal barrier". (3) Prior parameters: For example, the baseline immune maturity parameters corresponding to different ages.

[0135] Preferably, in step 3, the use of graph attention networks to dynamically embed and reason about the pathological knowledge graph to obtain the pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability, and drug resistance evolution factor includes the following sub-steps:

[0136] Step B1: Construct a dynamic heterogeneous graph representing the pathological knowledge graph. The heterogeneous graph is based on the inherent physiological nodes (including immune factor entity nodes and intestinal barrier state nodes) in the pathological knowledge graph. Environmental entity nodes, pathogen entity nodes, host entity nodes, microbiome entity nodes and drug entity nodes are dynamically instantiated according to the current real-time state. Each entity node is accompanied by a feature vector representing its current attribute state as the initial embedding of the node.

[0137] In this embodiment, a heterogeneous graph is constructed. In this heterogeneous graph, each node Represents an entity object, each edge This represents the semantic relationships between entities. Each node is accompanied by a feature vector. The feature vector is used to characterize the current attribute state of the entity. Nodes are structural elements of the graph, defining the type and identity of the entity; feature vectors are numerical attribute descriptions of the nodes, updated at each time step based on real-time monitoring data. Graph Attention Networks (GAT) infer the degree of influence between different entities by calculating the attention coefficients between node feature vectors.

[0138] Node set V contains the following five types of entity nodes:

[0139] The first type is environmental entity nodes. Create an entity node representing the current state of the chicken coop environment. The feature vector of this node The environmental feature vector obtained in step 2 Assignment, i.e. Among them, humidity value At a high level (corresponding to an actual humidity of 78%), this information will receive higher weight in inference through GAT's attention mechanism.

[0140] The second category is pathogen entity nodes. Pathogen entity nodes are created for each type of coccidia currently detected. In this embodiment, *Eimeria tenella* nodes are created based on the coccidia density estimates derived from the typing output of a convolutional neural network (CNN) regression model. Eimeria coccidia lesions and heap-type Eimeria nodes Each pathogen entity node's feature vector contains an estimated coccidia density for that species, a developmental stage encoding inferred from spectral indices (merozoite stage encoded as [1,0,0], gametophyte stage encoded as [0,1,0], and oocyst stage encoded as [0,0,1]), and biological characteristic parameters of that species of coccidia (such as baseline proliferation rate). (Host affinity site encoding). For example, the feature vector of the Eimeria tenella node is: .

[0141] In addition, each pathogen entity node is connected to the Coccidia genus general node through a "belongs" relationship edge. , eigenvectors It includes estimated coccidia density values ​​for all coccidia species and an overall infection intensity index; the overall infection intensity index is the aggregated weighted value of the estimated coccidia density values ​​for each currently detected coccidia species, used to characterize the total macroscopic pathogen load currently borne by the host.

[0142] The third type is the host entity node. Creating a host entity node. Its eigenvectors Normalized values ​​including flock age (assuming an age range of [1, 56] days, then 14 days normalized to...). ), behavioral feature vector The system also includes parameters for immune maturity at that age, obtained from a pathological knowledge graph based on age (e.g., an immune maturity of approximately 0.35 for a 14-day-old, indicating that the immune system is not yet fully developed). This node is connected to immune factor entity nodes (e.g., sIgA nodes, T lymphocyte nodes) via "secretion" relationships and to intestinal barrier status nodes via "possess" relationships.

[0143] The fourth category is microbial entity nodes. Activating microbial entity nodes from the pathological knowledge graph includes beneficial bacteria nodes (such as lactobacilli). ) and opportunistic pathogenic nodes (such as Clostridium perfringens) The feature vector of a microbial community entity node includes the current abundance estimate of the microbial community (or a prior distribution value based on age and environmental conditions if direct detection data is lacking at the current time step) and a proliferation rate parameter. Microbial community entity nodes and pathogen entity nodes are connected by edges representing "antagonistic" or "cooperative" relationships; the edge attributes include the antagonism / cooperative coefficient. The prior values ​​are as follows: The proliferation rate parameter is derived from a standard veterinary microbiology literature database or standard experimental data; the prior distribution value is calculated and output by a Bayesian conditional probability model, which is pre-constructed using historical environmental parameters of the farm, age data, and excrement sequencing samples from the corresponding period as a training set. When direct detection data is lacking at the current time step, the prior distribution value of the current microbial community abundance can be obtained by inputting the current environmental parameters and age data into the Bayesian conditional probability model; the prior value of the antagonistic / synergistic coefficient is determined by inputting historical case data into the coccidia-microbial-immune cell interaction sub-model and using a parameter fitting algorithm (such as least squares) to back-calibrate the coefficients. All the above parameters are pre-stored in a static basic pathological knowledge graph.

[0144] The fifth category is drug entity nodes. This involves creating an anticoccidial drug entity node (such as diclazuril). sulfachlorpyrifos sodium Each drug entity node's feature vector contains the drug's mechanism of action encoding and kill coefficient. The baseline values ​​and half-life parameters are defined. Drug entity nodes are connected to their targeted coccidia species nodes via "acting on" relationships and to microbial entity nodes via "inhibitory" or "promoting" relationships. Furthermore, the historical medication records of the farm are queried, and historical medication information is encoded as part of the drug entity node feature vector (such as the number of days since the last use and the cumulative number of uses) for subsequent resistance inference.

[0145] Step B2: Calculate the attention coefficients between environmental entity nodes, pathogen entity nodes, host entity nodes, microbial entity nodes, and drug entity nodes using a graph attention network, and then normalize them using the Softmax function to obtain the attention weights.

[0146] In this embodiment, relational reasoning and attention calculation are performed. A graph attention network is used to calculate the attention coefficients between each entity node. and entity nodes Calculate the attention coefficient ,

[0147] The formula for calculating the attention coefficient is:

[0148] ;

[0149] in, and These are feature vectors for environmental entity nodes, pathogen entity nodes, host entity nodes, microbial community entity nodes, or drug entity nodes, respectively. It is a non-linear activation function. To share the weight matrix, Let || be the attention weight vector, and || denotes the concatenation operation;

[0150] The attention weights are then obtained by normalization using the Softmax function. The formula for calculating attention weights is:

[0151] ;

[0152] in, For nodes The set of neighboring nodes, For the natural constant An exponential function with base 0;

[0153] Step B3: Based on the attention weights, update the node feature vectors using a weighted summation method. The calculation formula is as follows:

[0154] ;

[0155] in, This is the updated node feature vector. It is a non-linear activation function (such as ReLU or ELU). For attention weights, To share the weight matrix, For nodes eigenvectors, For nodes The set of neighboring nodes.

[0156] Assuming that in this inference, the system calculates the attention weights for each factor as follows: Example values ​​for the attention weight of the environmental factor (high humidity): (Significantly higher than other factors), the weight of the etiological factor Weight of the age factor (Relatively low) indicates that the current risk is mainly driven by environmental factors.

[0157] Step B4: The graph attention network maps the updated node feature vectors through a multi-task fully connected layer to obtain the pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability, and drug resistance evolution factor.

[0158] In this embodiment, after multiple layers of attention propagation, the graph attention network obtains the updated feature vectors (such as environmental features) for each entity node. Pathogen characteristics Host characteristics (etc.). The system concatenates these feature vectors to obtain a globally fused feature vector. Subsequently, a multi-task learning framework was adopted to... The inputs are respectively fed into four parallel task-specific multilayer perceptron (MLP) branches for mapping and output, specifically as follows:

[0159] Regarding the pathogenicity coefficient, the pathogenicity branch captures the interaction between environmental factors and pathogen characteristics through a fully connected layer. During the joint training phase of the graph attention network, the supervision labels for this branch are derived from a historical database: the ratio of the actual average intestinal lesion score of the same coccidia species under corresponding stress conditions to the baseline lesion score under standard conditions. Assuming that in this inference, the system determines that the main driving factor is high environmental humidity, the example value of the output pathogenicity coefficient is... (A value higher than the baseline of 1.0 indicates a high-risk state). This coefficient will be passed as a correction parameter to the subsequent digital twin model to adjust for coccidia proliferation rate. .

[0160] Regarding the host susceptibility index, the host susceptibility branch comprehensively assesses the host's immune function, intestinal barrier integrity, and nutritional status, outputting the host susceptibility index. The supervision labels for this branch during the training phase are derived from: the relative mortality rate of chicken flocks of corresponding age and under stress conditions in historical real-world farming data, and the normalized values ​​of mucosal immune barrier (e.g., sIgA) damage scores measured in the laboratory. Assuming that in this inference, due to the young age of the chicken flock (14 days) and its stress state (high humidity), the example value of the output host susceptibility index is... (Values ​​range from [0, 1], with 0.65 indicating moderate to high susceptibility). This index will affect the immune kill coefficient in the digital twin model, specifically the following relationship: .

[0161] Regarding the probability of coccidiosis outbreak risk, this probability index serves as a macro-level early warning trigger, intuitively indicating to managers the overall disease outbreak risk level of the flock and determining whether to initiate subsequent intelligent decision generation. It does not directly participate in the micro-dynamic calculations of the subsequent digital twin model. The specific calculation process is as follows: After multiple layers of attention propagation, the graph attention network obtains the updated feature vectors of each entity node. The risk probability branch concatenates the updated feature vectors of environmental entity nodes, pathogen entity nodes, and host entity nodes to obtain a fused feature vector. ,in This indicates a concatenation operation. Then... Input a two-layer fully connected network. The first layer maps the fused features to a 64-dimensional hidden layer (followed by ReLU activation). The second layer maps the 64-dimensional features to a single scalar output. Finally, the output value is constrained to the [0, 1] interval by a Sigmoid activation function.

[0162] ;

[0163] in, , The weights and biases of the first fully connected layer. , The weights and biases of the second fully connected layer. The parameter is the Sigmoid function. The parameters of the fully connected network are obtained by joint training with the parameters of the attention layer during the training phase of the graph attention network. The training label is a binary label indicating whether a coccidiosis outbreak actually occurred in the subsequent observation window in historical cases (0 indicates no outbreak, 1 indicates an outbreak).

[0164] It should be noted that, It is not a simple weighted sum of the attention weights of each factor, but rather a comprehensive inference result of the graph attention network after nonlinear transformation and multi-layer feature propagation. The aforementioned attention weights ( , , This reflects the relative importance of each factor in the information dissemination process, while The final risk probability value is obtained by further nonlinear mapping through a fully connected layer based on the attention-weighted updated node feature vectors. Assume that in this inference, the example value of the coccidiosis outbreak risk probability output by the system is... (The value range is [0, 1], 0.82 indicates high risk, and it is recommended to start prevention and control measures immediately).

[0165] Regarding the drug resistance evolution factor, it should be noted that when the system constructs the initial static pathological knowledge graph in step B1, it has pre-constructed "drug resistance gene nodes" (representing the inherent resistance potential of a specific coccidia to a certain class of drugs) and linked them to drug entity nodes. After extracting and fusing features from the drug resistance evolution branch, the features of the gene node are activated by combining historical drug use records. The training labels for this branch are derived from the normalized values ​​of the drug efficacy decay slopes of historical batch anticoccidial drug sensitivity tests (such as the Anticoccidial Index, ACI) from the farm. Assuming that no anticoccidial drugs have been used recently in this inference, the example value of the output drug resistance evolution factor is... (The value ranges from [0, 1], and 0.15 indicates a low risk of drug resistance, meaning the drug is still effective).

[0166] Step 4: Input the microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor into the digital twin model. The digital twin model outputs the intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value.

[0167] The digital twin model includes a coccidia-microbiota-immune cell interaction sub-model and an intestinal substance transport sub-model based on computational fluid dynamics;

[0168] Step 4 includes the following sub-steps:

[0169] Step C1: Use the intestinal substance transport submodel to determine the real-time concentration distribution of the drug in the intestine;

[0170] In terms of drug distribution simulation, computational fluid dynamics (CFD) is used. The intestinal substance transport sub-model specifically introduces a non-Newtonian fluid model to describe the rheological properties of chyme, with the governing equations being: ,in, The density of the chyme, in units of , This represents the rate of change of the chyme flow velocity vector over time (local acceleration). This represents the flow velocity vector of chyme within the intestine. Let be the velocity gradient tensor. This represents the pressure gradient vector within the intestine. For spatial differential operators, For the shear stress tensor, This is the momentum sink caused by water absorption by the intestinal mucosa.

[0171] Power-law model Describing the shear stress tensor ,in, This is the consistency coefficient. As a liquidity behavior index, Shear rate, in units of ;

[0172] Furthermore, based on the acquired flow velocity vector The real-time concentration distribution of drugs in the gut was calculated by coupling convection-diffusion equations. The formula is: .in, This represents the rate of change of drug concentration over time. This represents the real-time concentration of the drug in the intestine. This is the convection term (characterizing the effect of chyme flow on drug transport). The diffusion coefficient of the drug in chyme. For diffusion term, This refers to the local absorption of drugs by the intestinal mucosa.

[0173] By introducing a non-Newtonian fluid model, the shear thinning or thickening characteristics of chicken intestinal chyme are accurately characterized, eliminating errors introduced by the traditional Newtonian fluid assumption. The local shear stress and viscosity calculated based on the power-law model directly affect the flow velocity in the convection-diffusion equations. and diffusion coefficient This allows for precise simulation of the diffusion resistance and distribution gradient of drug molecules in the intervillous space and deep crypts of the intestine, thus determining the real-time concentration distribution of the drug within the intestine. This provides a precise local concentration input for the drug killing term in the kinetic equation, thereby enabling targeted optimization of the dosage and avoiding incomplete insecticidal effects due to insufficient drug concentration or toxic side effects due to excessively high concentration.

[0174] Step C2: Based on the real-time concentration distribution of drugs in the gut, the microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor are input into the coccidia-microbe-immune cell interaction sub-model to obtain the coccidia biomass evolution curve, immune intensity evolution curve and microbe abundance evolution curve.

[0175] Preferably, the coccidia-microbiota-immune cell interaction sub-model includes: a coccidia proliferation equation, an immune response equation, and a microbiota succession equation.

[0176] The coccidia reproduction equation is as follows:

[0177] ;

[0178] in, For the first The endophytic reproduction rate of coccidia is expressed in units of 1 / h (per hour), with a typical range of values. , For the first The biomass of coccidia in the intestine (including merozoites and gametophytes) is expressed as the number of coccidia per unit volume of the intestine (cities / mL). The intestinal environment's carrying capacity for coccidia, expressed in cells / mL, represents the upper limit of the total coccidia biomass that the intestinal environment can accommodate. A typical value is [value missing]. to The number of cells / mL is affected by intestinal volume, nutrient supply and host immune status. The first intestinal Biomass of coccidia; It represents the total biomass of all coccidia and is used to characterize resource competition between and within species. For the immune system to the first The lethality coefficient of coccidia, in units of The typical range of values ​​is This coefficient reflects the clearance efficiency of immune cells (such as T cells and macrophages) against coccidia. For the strength of the host's local immunity, For the first species of bacteria on the first Antagonistic or synergistic coefficient of coccidia, in units of ,when The time indicates that the bacterial flora has an antagonistic effect on coccidia (such as the organic acids produced by lactic acid bacteria inhibiting coccidia). The time indicates a synergistic pathogenic effect (such as Clostridium perfringens and Coccidia jointly damaging the intestinal barrier), with a typical value range of [value missing]. , For the first Abundance of dominant gut microbiota (such as lactobacilli and Clostridium perfringens), expressed in colony forming units per gram of intestinal contents (CFU / g). The drug's killing or inhibition coefficient against coccidia is expressed in units of... The typical range of values ​​is Different drugs Different values, for example, dickelzul. sulfachlorpyrimethamine sodium , This represents the real-time concentration distribution of drugs within the intestines.

[0179] Initial coccidia biomass Set to a value corresponding to the current coccidia oocyst density value output by the pre-trained convolutional neural network (CNN) regression model (assuming...). = (Units / gram, at an early infection level).

[0180] In actual simulation, Instead of directly using benchmark values ​​from literature or experience Instead, it is based on the pathogenicity coefficient output in step 2. Use after making corrections.

[0181] Specifically, the effective proliferation rate actually substituted into the coccidia proliferation equation The calculation formula is:

[0182] ;

[0183] in, The baseline endophytic proliferation rate of the i-th coccidia species, measured under standard experimental conditions or obtained from the literature, is given. The pathogenicity coefficient, output from the pathological knowledge graph reasoning in step 2, reflects the comprehensive corrective effect of multiple factors such as current environmental stress and host age on coccidia proliferation. At that time, the effective proliferation rate equals the baseline value; when This indicates that the current conditions are favorable for coccidia proliferation, and the effective proliferation rate is higher than the baseline value.

[0184] In this embodiment, the output pathogenicity coefficient Substitute into the above formula. Assume the baseline reproduction rate of *Eimeria tenella*. (Typical value range is [0.05, 0.3]), then the corrected effective proliferation rate is In the numerical integration of the coccidia reproduction equation, all occurrences The effective proliferation rate is used in all locations. Perform the calculation.

[0185] Real-time concentration distribution of drugs in the gut under non-interventional conditions ,therefore, Assuming environmental carrying capacity cells / mL (based on typical values).

[0186] Based on the numerical integration of the kinetic equations, without intervention, the coccidia biomass will decrease in the next 48 hours. It will show an exponential growth trend. Simulation results show that: in 24 hours, Increase to approximately cells / mL; at 48 hours near Cells / mL, close to the upper limit of environmental carrying capacity. .

[0187] In setting the parameters of the coccidia reproduction equation, besides correcting the reproduction rate... In addition, the immune kill coefficient also needs to be corrected. . The effective value is affected by the host susceptibility index. The adjustment, specifically the correction relationship is as follows: The host susceptibility index output in step 3. Substitute the values ​​and assume the baseline immune kill coefficient. 1 / (h·unit of immune strength) (based on the typical value range [0.01, 0.1]), then the corrected effective immune killing coefficient is: 1 / (h·unit of immune strength). This correction value will be substituted into the coccidia proliferation equation. Item, that is, the actual calculation is This means that the higher the host's susceptibility ( The larger the coccidia, the weaker the immune system's ability to kill them, and the less immunosuppression the coccidia's proliferation suffers.

[0188] This outputs the coccidia biomass evolution curve. For example: at hour 0, per gram; in 24 hours per gram; at 48 hours pcs / gram;

[0189] The immune response equation is as follows:

[0190] ;

[0191] in, The response coefficient of the immune system to coccidia stimulation, expressed in units of immune strength. Typical value This indicates the intensity of the immune response induced by a unit of coccidia biomass. For the first Biomass of coccidia in the gut The natural decay coefficient of immune strength, in units of h typically takes the value of It reflects the apoptosis rate of immune cells and the degradation rate of immune factors. The strength of the host's local immunity (such as sIgA and T lymphocyte levels) is dimensionless and ranges from [0, 10].

[0192] This equation describes the strength of the host immune system. Dynamic changes: Coccidia biomass Stimulating the immune system to produce a response (by (Item-driven), while immune strength naturally declines over time (by...) (Item description). It should be noted that the host susceptibility index... The impact on the immune system is reflected in the immune kill coefficient in the coccidia proliferation equation. The efficiency of the immune system in actually clearing coccidia is affected. The immune response coefficient is adjusted (in terms of its modulation) without directly altering the parameters of the immune response equation itself. Simultaneously, to more comprehensively reflect the impact of host susceptibility on immune kinetics, the immune response coefficient is... Also make corrections, the correction relationship is as follows: ,in, As the baseline immune response coefficient, This is the host susceptibility index, meaning that the higher the host susceptibility, the weaker the immune system's response to coccidia stimulation.

[0193] Assuming initial immune strength (Dimensionless, range of values ​​is) Based on typical values, the baseline immune response coefficient Immune strength unit / ( The corrected effective response coefficient is Immunological intensity unit / (h·cells / mL). Immunological attenuation coefficient. h.

[0194] Based on the numerical integration results, the immune strength evolves over time as follows: at 0 hours, ; within 24 hours, due to coccidia stimulation, It rose to approximately 3.5; within 48 hours, The value further increased to approximately 5.2, indicating that the host's immune system was actively responding to the infection, but due to the host's high susceptibility, the speed and intensity of the immune response were lower than those in a healthy state.

[0195] This results in the output of the immune strength evolution curve. For example: at hour 0, ; In 24 hours, time ; at 48 hours, ;

[0196] The microbial community succession equation is as follows:

[0197] ;

[0198] in, For the first Endogenous proliferation rate of the microbial community, in units of The typical range of values ​​is The proliferation rates of different bacterial groups vary considerably. For the first Abundance of dominant gut microbiota (such as lactobacilli and Clostridium perfringens), expressed in colony forming units per gram of intestinal contents (CFU / g). The gut microbiota's environmental carrying capacity, expressed in CFU / g, with a typical value of to CFU / g, For the first Coccidia against the first The influence coefficient of the microbial community, in units of ,when This indicates that after coccidia disrupt the intestinal barrier, they provide a growth substrate for the gut microbiota. The time represents the competitive inhibition of bacterial communities by coccidia, with a typical value range of [value missing]. , For the first Biomass of coccidia in the gut The coefficient of drug killing or inhibiting bacterial flora, in units of 1. The typical range of values ​​is ,when This indicates that the drug has an inhibitory effect on the bacterial flora. This indicates that the drug promotes the growth of gut microbiota (such as some prebiotic drugs). This represents the real-time concentration distribution of drugs within the intestines.

[0199] Assume the system focuses on monitoring two key bacterial groups: lactobacilli ( ) and Clostridium perfringens ( ).

[0200] For Lactobacillus, assuming initial abundance CFU / g (within normal levels). Based on typical values, the proliferation rate of lactobacilli... Environmental carrying capacity CFU / g, the effect coefficient of coccidia on lactobacilli 1 / ( • cells / mL (negative values ​​indicate that coccidia inhibit the growth of lactobacilli after disrupting the intestinal barrier). In a no-intervention scenario, ,therefore .

[0201] Based on the numerical integration results, the abundance of Lactobacillus evolved over time as follows: at 0 hours... CFU / g; at 24 hours, due to coccidia disrupting the intestinal barrier, Decrease to approximately CFU / g; at 48 hours Further decline to approximately A CFU / g level indicates that the gut microbiota balance is being disrupted.

[0202] For Clostridium perfringens, assuming an initial abundance CFU / g (low level). Based on typical values, the proliferation rate of *Clostridium perfringens*. The influence coefficient of coccidia on Clostridium perfringens 1 / (h·c_c_mL) (A positive value indicates that after coccidia destroy the intestinal barrier, they provide a growth substrate for Clostridium perfringens).

[0203] Based on the numerical integration results, the abundance of *Clostridium perfringens* evolved over time as follows: at 0 hours... CFU / g; at 24 hours, due to coccidia disrupting the intestinal barrier, Rise to approximately CFU / g; at 48 hours Further rise to approximately A CFU / g level indicates that pathogenic bacteria are multiplying and there is a risk of secondary infection.

[0204] This outputs the evolution curve of bacterial community abundance. For example: Lactobacillus from CFU / g decreased to CFU / g, Clostridium perfringens from CFU / g rose to CFU / g;

[0205] Step C3: Based on the evolution curves of coccidia biomass, immune intensity, and microbial abundance, the evolution curve of intestinal lesion score is calculated using the intestinal lesion score mapping formula.

[0206] The intestinal lesion scoring mapping formula is:

[0207] ;

[0208] in, Sigmoid function This is used to constrain the output within the interval [0, 1], with coefficient 4 as a calibration factor, mapping the output to the Johnson-Reid scoring range of [0, 4]. This is a normalized value of coccidia biomass relative to environmental carrying capacity, reflecting the degree of direct damage coccidia cause to intestinal epithelial cells, and environmental carrying capacity. cells / mL; This is an immune-pathogen interaction term, reflecting secondary damage to intestinal tissue caused by inflammatory responses during the immune response. (This is a reference normalized value for immune strength, with a value of 10). The abundance of Clostridium perfringens is a normalized value, reflecting the additional damage to the intestine caused by secondary bacterial infection following coccidiosis, and the bacterial carrying capacity. CFU / g; , , These are the weighting coefficients for each item, with typical values ​​as follows: , , This reflects that direct damage from coccidia is the main pathogenic factor, while secondary bacterial infection and immune inflammation are auxiliary factors; This is the threshold offset, typically 1.5, used to control the sensitivity of LS score increases.

[0209] Predicted intestinal lesion score ( The calculation is based on a comprehensive assessment of three dimensions: coccidia biomass, immune damage, and dysbiosis.

[0210] Substitute the simulation values ​​at each time step into the above formula. At 24 hours, the coccidia biomass... cells / mL , CFU / g, substitute into the formula to calculate the three dimensions:

[0211] Direct damage caused by coccidia: ;

[0212] Immunopathogen interaction: ;

[0213] Secondary bacterial infections: .

[0214] Sum the above three items and subtract the threshold offset. To obtain the input variables of the Sigmoid function : After Sigmoid nonlinear activation and mapping with calibration factor (4), the final score is obtained: The score is 1.0 (meaning the intestines show early, mild lesions).

[0215] Based on the aforementioned simulation results, the condition worsened after 48 hours: coccidia biomass cells / mL, immune strength Clostridium perfringens CFU / g. Similarly, substitute into the formula to calculate:

[0216] Direct damage caused by coccidia: ;

[0217] Immunopathogen interaction: ;

[0218] Secondary bacterial infections: .

[0219] Sum the above three items and subtract the threshold offset. ,get: After Sigmoid mapping, the final score is obtained: The score (i.e., due to the pathogenic factors exceeding the tolerance threshold, the lesion rapidly deteriorates, and the score rises to approximately 3.3 points).

[0220] The weighting coefficients in the above formula , , and threshold The following method was used to determine the optimal values ​​of each coefficient by using the LS score obtained from actual autopsies as a supervision label, and the coccidia biomass, immune intensity, and bacterial abundance at the corresponding time points as inputs, and by fitting the data using the least squares method.

[0221] Step C4: Based on the coccidia biomass evolution curve and historical drug use records, the drug resistance risk calculation formula is used to calculate the drug resistance development trend;

[0222] The formula for calculating drug resistance risk is:

[0223] ;

[0224] in, As a factor in the evolution of drug resistance, in this embodiment ; The resistance variation coefficient driven by population size reflects the biological principle that the larger the coccidia population, the higher the probability of natural mutations producing drug-resistant individuals; a typical value is 0.05. For reference population size, the value is taken as... cells / mL; The cumulative resistance coefficient driven by drug selection pressure reflects the principle that the greater the intensity of anticoccidial drug use, the higher the resistance selection pressure. A typical value is 0.011 / (h·μg / mL). The time integral of drug concentration (drug exposure), in h·μg / mL; This is the genetic coefficient of drug resistance based on historical drug use records, reflecting the impact of past drug use history on the frequency of drug resistance genes in the current coccidia population. A typical value is 0.02. For time The first time Coccidia biomass; Indicates time The total biomass of all coccidia at that time; This is the historical drug use intensity index (dimensionless, calculated from a knowledge graph based on the frequency and dosage of various anticoccidial drugs used over the past 6 months). Final The value is restricted to the interval [0, 1] by the truncation function.

[0225] In the non-interventional scenario of this embodiment, since anticoccidial drugs have not yet been used, Therefore, drug selection stress item Assuming the farm has not used anticoccidial drugs recently, the historical drug use intensity index is... Therefore, historical genetic items The risk value of drug resistance is mainly determined by baseline values ​​and population size drivers:

[0226] At hour 0, pcs / gram (Slightly above the 0.15 baseline level);

[0227] In 24 hours, pcs / gram (Slight increase; the expansion of the coccidia population increases the probability of natural variation.)

[0228] At 48 hours, pcs / gram (It remains at a low-risk level because no drug has been used to exert selection pressure).

[0229] The above results indicate that, in the absence of drug use, the risk of drug resistance is primarily influenced by the size of the coccidia population, and its growth is slow. Once anticoccidial drugs are started, This will significantly increase the rate at which the risk of drug resistance accumulates, which is also an important basis for the present invention's emphasis on precision medicine and drug rotation strategies.

[0230] Step C5: Based on the coccidia biomass evolution curve, intestinal lesion score evolution curve, and drug resistance development trend, obtain the intestinal lesion score, drug resistance risk value, and predicted coccidia oocyst density value.

[0231] The digital twin model outputs the complete evolution trajectory over the next 48 hours, providing a forward-looking basis for subsequent decision-making. Specific output examples include:

[0232] (1) Coccidia biomass evolution curve At hour 0, per gram, over 24 hours per gram, at 48 hours pcs / gram;

[0233] (2) Evolution curve of immune intensity At hour 0, In 24 hours, At 48 hours, ;

[0234] (3) Curve of bacterial community abundance evolution Lactobacillus from CFU / g decreased to CFU / g, Clostridium perfringens from CFU / g rose to CFU / g;

[0235] (4) Evolution curve of intestinal lesion score At hour 0, In 24 hours, At 48 hours, ;

[0236] (5) Trends in drug resistance At hour 0, In 24 hours, At 48 hours, (Low risk level, because no medication has been used yet);

[0237] Therefore, intestinal lesion score Drug resistance risk value Predicted coccidia oocyst density value pcs / gram.

[0238] This approach constructs a sub-model of coccidia-microbiota-immune cell interactions, reconstructing the dynamic game-theoretic process of the micro-ecosystem in a virtual space. Using the coccidia proliferation equation, the combined effects of four factors—endogenous reproduction, immune killing, microbiota antagonism, and drug inhibition—are quantified, accurately simulating the rise and fall of the coccidia population. Immune response equations and microbiota succession equations are used to simulate the dynamic changes in host resistance and the micro-ecological environment. This model can extrapolate different intervention methods (such as the administration of probiotics). Or use medication The disturbance effect on the entire system.

[0239] Step 5: Input the intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value into the machine learning algorithm, and the machine learning algorithm generates a prevention and control strategy.

[0240] Preferably, in step 5, the machine learning algorithm is a proximal policy optimization (PPO) algorithm, a genetic algorithm, a particle swarm optimization (PSO) algorithm, or a simulated annealing algorithm.

[0241] In this embodiment, machine learning optimizes the PPO algorithm for near-end policies.

[0242] First, construct the state space, using the predicted values ​​as the current state. Then, a strategy search is performed to optimize the agent's attempts at different action combinations in the virtual environment. .

[0243] In terms of optimizing objective calculation, a multi-objective optimization function is designed:

[0244] ;

[0245] in, To predict daily weight gain, The intestinal lesion score, For economic costs, This represents the risk value for drug resistance. , , and All are weighting coefficients.

[0246] The specific calculation methods for each indicator in the multi-objective reward function are as follows:

[0247] ① Predicted daily weight gain The calculation method: The digital twin model also includes a growth performance sub-model to predict daily weight gain. It was calculated using the growth performance sub-model within the digital twin model. This sub-model is based on the following empirical formula:

[0248] ;

[0249] in, The baseline daily weight gain of healthy chickens in this age group under standard feeding conditions (obtained by referring to the breed growth curve table). The data for the breed growth curve table comes from publicly available existing technologies, specifically including production performance target manuals for specific breeds published by commercial poultry breeding companies (such as the production performance management manual for Ross 308 or Cobb 500 broiler chickens), or growth reference curves in relevant agricultural industry feeding standards. The intestinal lesion is a factor that diminishes growth performance, and is correlated with the predicted intestinal lesion score. Related, the calculation formula is as follows (That is, for every 1 point increase in LS, the daily weight gain decreases by about 15%, and this coefficient is determined based on literature and historical breeding data), with a value range of [0.4, 1.0]; This is a factor influencing feed intake, related to feed intake activity in the behavioral feature vector. Related, the calculation formula is as follows (in, (This is a reference value for normal feeding activity), with a value range of [0.5, 1.0]. Environmental stress factors are considered when environmental parameters are within a suitable range. When the temperature or humidity exceeds the suitable range, it decreases linearly according to the degree of deviation. The calculation formula is as follows: (in, For temperature, H represents the optimal temperature, and H represents the humidity. This is the upper limit threshold for humidity. , These are the stress coefficients for temperature and humidity, typically taking values ​​of 0.02 / ℃ and 0.005 / % respectively.

[0250] Taking strategy A as an example: after using 2 times the dose of diclazuril, the digital twin model predicts... ,but Since high doses of medication may cause decreased appetite, it is assumed that... If it drops to 0.65, then Environmental factors Assume a value of 0.92 (due to high humidity stress). Substitute into the calculation: .

[0251] Taking strategy B as an example: After ventilation and dehumidification, the environment improves, and predictions are made. ,but Probiotics and low-dose medications have little effect on feed intake, assuming If it is 0.80, then After ventilation and dehumidification Increased to 0.98. Substituted into the calculation: .

[0252] It should be noted that in practical systems, probiotic preparations can also improve nutrient absorption efficiency by improving the balance of gut microbiota. This effect can be achieved by adding a probiotic synergist to the formula. To reflect (when probiotics are added) The value is set to 1.1-1.3, thereby further improving the predicted daily weight gain of Strategy B. Substitute, .

[0253] ② Economic Costs Calculation method:

[0254] Economic costs of prevention and control strategies The normalized dimensionless value is calculated using the following formula:

[0255] ;

[0256] in, The actual total cost of the prevention and control strategy within a decision-making cycle (e.g., 24 hours) (unit: yuan / thousand chickens). This is a cost normalization reference value (unit: yuan / thousand chickens), used to map the cost to a range comparable to the values ​​of other items in the optimization function (daily weight gain in grams and intestinal lesion score in 0-4 points). The value is determined based on the actual cost level of the farm, and the determination principle is: to make the typical prevention and control strategy... The values ​​fall within the same order of magnitude as daily weight gain (approximately 30-55g) and intestinal lesion score (0-4 points), thus ensuring that each reward item can effectively play a regulatory role when weighted and summed. It is obtained by summing the following costs:

[0257] ;

[0258] in, The formula for calculating the cost of chemical drugs is as follows: , Let be the unit price (yuan / mg) of the m-th chemical drug. This refers to the dosage of the drug (mg / kg body weight). The average weight of the flock (kg). A flock of chickens is defined as a unit of one thousand chickens; The cost of probiotic preparations is calculated using the following formula: , Price per unit of probiotic preparation (RMB / g preparation). The dosage is calculated per kilogram of feed (g preparation / kg feed). The average daily feed intake of the flock (kg feed / thousand chickens). This refers to a flock size of 1,000 chickens. The energy consumption cost for environmental regulation is calculated using the following formula: , The power of the ventilation equipment (kW). Runtime (h) The price is the electricity price (yuan / kWh).

[0259] Taking Strategy A as an example: using twice the dose of diclazuril, assuming the unit price of diclazuril is 0.8 yuan / mg, the dosage is 2 mg / kg body weight (twice the conventional dose), and the average weight of the flock is 0.35 kg. Based on a unit of 1000 chickens, then... Yuan per thousand chickens; cost without probiotics No additional environmental control . Yuan per thousand chickens.

[0260] For example, Strategy B involves using a low dose of diclazuril (50% of the standard dose, i.e., 1 mg / kg × 50% = 0.5 mg / kg), along with the addition of Bacillus subtilis preparation and activation of the ventilation system. (Chemical drug cost) Yuan / thousand chickens; assuming the unit price of probiotic preparations is... Yuan / g, amount added per kilogram of feed g / kg, average daily feed intake kg feed / 1000 chickens Yuan / thousand chickens (Note: Probiotic preparations are biological products; the unit cost is low, but the dosage is large, so the total cost is higher than that of chemical drugs); assuming ventilation equipment power... kW, runtime h, electricity price Yuan / kWh, then Yuan per thousand chickens (Note: Ventilation and dehumidification require continuous operation of the fan, and energy consumption costs are the main expense). Yuan per thousand chickens.

[0261] Based on the above calculations, the actual total costs of the two typical strategies are 0.56 yuan / thousand chickens and 11.54 yuan / thousand chickens, respectively. To make... The numerical magnitudes are comparable to the daily weight gain (approximately 30-45 g) and lesion score (approximately 1.5-2.0 points) in the reward function, indicating that the calibration... The value of . Considering the reward function Item weight coefficient In order to make cost penalty items It should occupy a reasonable proportion of the total reward function value (approximately 10%-20% of the total reward value), and The price is set at 0.77 yuan per thousand chickens. Based on this, the following calculations are made:

[0262] Strategy A: Cost penalty items ;

[0263] Strategy B: Cost penalty items .

[0264] The system compares the effects of different strategies in a virtual environment:

[0265] Strategy A (Increased Dosage): Use diclazuril at twice the usual dose. Virtual simulation results show: (Daily weight gain decreased due to medication side effects) (The lesion score dropped rapidly). (The cost of chemical drugs alone is extremely low, making the economic cost very low.) (Due to the use of twice the dose of chemical drugs, the drug selection stress item) (A significant increase leads to a substantial rise in the risk of drug resistance). Substituting into the formula for calculation:

[0266]

[0267] Strategy B (Ventilation and Dehumidification + Probiotics + Low-Dose Medication): The ventilation system was activated to reduce indoor humidity to below 60%, Bacillus subtilis preparation was added, and a low dose of diclazuril (50% of the standard dose) was used. Virtual simulation results show: (Daily weight gain remains good) (The lesion score steadily declined). (Involves the addition of probiotics and continuous ventilation, which has a high economic cost.) (The risk of drug resistance remains low). Substitute into the formula to calculate:

[0268] .

[0269] The comparison results show that Strategy B has the best overall score. Significantly higher than strategy A Although the direct economic cost of strategy B is higher than that of strategy A, its advantages in daily weight gain (45g vs 30g) and drug resistance risk control (0.2 vs 0.8) more than offset the difference in economic cost. This result fully demonstrates the decision-making value of multi-objective optimization functions—not simply choosing the lowest-cost option, but seeking a globally optimal balance between production efficiency, disease control, economic cost, and long-term drug resistance risk.

[0270] Finally, the system issues commands, selecting strategy B and automatically sending control instructions: activate the ventilation system to reduce indoor humidity to a suitable range (target humidity 60%), and activate the automatic dosing machine to add probiotic preparation (Bacillus subtilis, dosage...). CFU / g feed) and low-dose anticoccidial drug (dicrazol, 0.5 mg / kg body weight).

[0271] Furthermore, it also includes virtual virus attack experiments based on digital twins, specifically including:

[0272] This digital twin model sets up a virtual "highly virulent strain invasion" scenario, and simulates extreme pathogenic conditions by adjusting model parameters: the endophytic proliferation rate of coccidia. Increase the viral load by 15%-30% to simulate the high reproductive capacity of virulent strains, thereby reducing the host susceptibility index. Increase the concentration to 0.7-0.9 to simulate an immunosuppressive state (such as mycotoxin poisoning, vaccine failure, and other stress conditions), and set environmental stress parameters (such as ammonia concentration and humidity) to unfavorable levels;

[0273] The complete life cycle of coccidia in intestinal epithelial cells was simulated in a virtual environment, including: (1) asexual reproduction cycle: after sporozoites invade intestinal epithelial cells, they form merozoites through schizogony. This process lasts 48-72 hours. The model calculates the number of merozoites and their degree of damage to the intestinal epithelium in real time by solving the coccidia proliferation equation; (2) sexual reproduction cycle: merozoites differentiate into macrogametoids and microgametoids, form zygotes through gametogony, and finally develop into oocysts and are released from the body. This process lasts 72-120 hours. The formation rate of gametoids and the amount of oocysts produced are simulated by the coccidia-microbe-immune cell interaction model.

[0274] In a virtual challenge scenario, the effects of different prevention and control strategies were compared: multiple virtual experiments were set up to test single chemical drugs (such as diclazuril and sulfonamides), drug combinations (such as ion carriers + chemically synthesized shuttle drugs), biological control (probiotics + traditional Chinese medicine extracts), and comprehensive control (drugs + microecological regulation + environmental management). The evolution curves of intestinal lesion scores, coccidia clearance rates, drug resistance development trends, and aquaculture economic benefits under each scheme were predicted by digital twin models, so as to screen out the optimal prevention and control plan that can still maintain effectiveness under extreme conditions.

[0275] The results of the virtual virus attack experiment are stored in this knowledge graph as historical experience for subsequent decision-making. When the actual monitoring data matches the initial conditions of the virtual virus attack scenario, the system can directly call the corresponding optimal plan to achieve rapid response and avoid economic losses caused by on-site trial and error.

[0276] In this way, virtual virus challenge experiments simulate extreme pathogenic conditions in digital space, providing a low-cost, zero-risk testing platform for verifying the robustness of prevention and control strategies. By comparing the performance of different schemes in a virtual environment, potential risks of prevention and control failure can be identified in advance, and contingency plans can be developed. The knowledge storage of virtual experiment results enables the accumulation and reuse of experience, possessing "foresight" and "learning ability," significantly improving the scientific and forward-looking nature of prevention and control decisions.

[0277] Example 2

[0278] refer to Figure 2 A dynamic control system for chicken coccidiosis, used to implement the above-mentioned dynamic control method for chicken coccidiosis, includes the following units:

[0279] Data acquisition module: used to acquire environmental parameter data, excrement hyperspectral data, and flock behavior image data;

[0280] State Tensor Construction Module: Used to analyze, process, and stitch together environmental parameter data, excrement hyperspectral data, and chicken flock behavior image data to obtain a microecological state tensor. Then, the microecological state tensor is extended by time series to obtain a time series tensor.

[0281] Knowledge Computation Module: Used to construct a pathological knowledge graph. Based on the microecological state tensor, a graph attention network is used to dynamically embed and reason about the pathological knowledge graph to obtain pathogen pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability and drug resistance evolution factor.

[0282] Digital twin simulation module: used to input microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor into digital twin model. The digital twin model outputs intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value.

[0283] Intelligent decision-making module: This module is used to input intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value into machine learning algorithm, which generates prevention and control strategy.

[0284] In this embodiment, the data acquisition module acquires environmental parameter data, excrement hyperspectral data, and chicken behavior image data from sensors, hyperspectral imagers, and cameras, and then sends the environmental parameter data, excrement hyperspectral data, and chicken behavior image data to the state tensor construction module.

[0285] The state tensor construction module processes environmental parameter data, excrement hyperspectral data, and chicken flock behavior image data to obtain environmental feature vectors, spectral feature vectors, and behavioral feature vectors. Then, the environmental feature vectors, spectral feature vectors, and behavioral feature vectors are concatenated to obtain the microecological state tensor. Combined with the historical microecological state tensor of the past 24 hours, a time series tensor is formed. Finally, the microecological state tensor and the time series tensor are sent to the knowledge computing module and the digital twin simulation module.

[0286] The knowledge computing module uses graph attention network to dynamically embed and reason about the pathological knowledge graph based on the microecological state tensor, and obtains pathogen pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability and drug resistance evolution factor. Then, the pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor are sent to the digital twin simulation module.

[0287] The digital twin simulation module includes a gut substance transport sub-model and a coccidia-microbe-immune cell interaction sub-model. The gut substance transport sub-model obtains the real-time concentration distribution of drugs in the gut. Then, the microecological state tensor, time series tensor, real-time concentration distribution of drugs in the gut, pathogen pathogenicity coefficient, host susceptibility index, and drug resistance evolution factors are input into the coccidia-microbe-immune cell interaction sub-model to obtain intestinal lesion score, drug resistance risk value, and predicted coccidia oocyst density value. Finally, the intestinal lesion score, drug resistance risk value, and predicted coccidia oocyst density value are sent to the intelligent decision-making module.

[0288] The intelligent decision-making module uses machine learning algorithms to calculate the prevention and control strategy based on a multi-objective optimization function. The prevention and control strategy is then parsed into PLC control instructions and sent to the field controller via the Modbus protocol. The field controller then executes the prevention and control strategy.

[0289] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for dynamic prevention and control of coccidiosis in chickens, characterized in that, Includes the following steps: Step 1: Acquire environmental parameter data, excrement hyperspectral data, and flock behavior image data; Step 2: Analyze and stitch together the environmental parameter data, excrement hyperspectral data and chicken behavior image data to obtain the microecological state tensor, and then perform time series expansion on the microecological state tensor to obtain the time series tensor; Step 3: Construct a pathological knowledge graph. Based on the microecological state tensor, use a graph attention network to dynamically embed and reason about the pathological knowledge graph to obtain pathogen pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability, and drug resistance evolution factor. Step 4: Input the microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor into the digital twin model. The digital twin model outputs the intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value. Step 5: Input the intestinal lesion score, drug resistance risk value, and predicted coccidia oocyst density value into the machine learning algorithm, which will then generate a prevention and control strategy. The digital twin model includes a coccidia-microbiota-immune cell interaction sub-model and an intestinal substance transport sub-model based on computational fluid dynamics; The coccidia-microbiota-immune cell interaction sub-model includes: The coccidia reproduction equation is as follows: in, For the first The endophytic reproduction rate of coccidia For the first Biomass of coccidia in the gut The environmental carrying capacity of the gut to coccidia. For the first Biomass of coccidia in the gut For the immune system to the first The lethality coefficient of coccidia, For the strength of the host's local immunity, For the first species of bacteria on the first Antagonistic or synergistic coefficient of coccidia, For the first The abundance of dominant gut microbiota This represents the killing or inhibition coefficient of the drug against coccidia. This represents the real-time concentration distribution of drugs within the intestines. The immune response equation is as follows: in, This represents the response coefficient of the immune system to coccidia stimulation. For the first Biomass of coccidia in the gut This is the natural decay coefficient of immune strength. This refers to the strength of the host's local immune response; The microbial community succession equation is as follows: in, For the first The endophytic proliferation rate of the microbial community, For the first The abundance of dominant gut microbiota For the gut microbiota's environmental carrying capacity, For the first Coccidia against the first Influence coefficient of microbial community For the first Biomass of coccidia in the gut This represents the coefficient of drug's killing or inhibition of bacterial flora. This represents the real-time concentration distribution of drugs within the intestines. The intestinal lesion scoring mapping formula is: in, This is the Sigmoid function, with coefficient 4 being a calibration factor. This is the normalized value of coccidia biomass relative to environmental carrying capacity. For immune-pathogen interaction, The normalized value for the abundance of Clostridium perfringens is given. , , All are weighting coefficients. This is the threshold offset; The formula for calculating drug resistance risk is: in, As a factor in the evolution of drug resistance, The resistance variation coefficient driven by population size. For reference population size, The pressure-driven cumulative resistance coefficient for drug selection. The time integral of drug concentration, This represents the heritability coefficient of drug resistance based on historical drug use records. This is a historical drug use intensity index. For time The first time Coccidia biomass.

2. The method for dynamic prevention and control of coccidiosis in chickens according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step A1: Perform Min-Max normalization on the environmental parameter data to obtain the environmental feature vector; Step A2: Based on the excrement hyperspectral data, the coccidia oocyst index and hemorrhage index are calculated using the formulas for the coccidia oocyst index and hemorrhage index. The coccidia oocyst index and hemorrhage index are then input into a pre-trained convolutional neural network regression model to obtain the current coccidia oocyst density value. The current coccidia oocyst density value is then log-normalized to obtain an estimated coccidia density value. Based on the coccidia oocyst index, hemorrhage index, and estimated coccidia density value, a spectral feature vector is constructed. Step A3: Analyze the chicken flock behavior image data using a target detection algorithm and perform ratio normalization to obtain the proportions of feeding activity, movement activity, and resting time. Based on the proportions of feeding activity, movement activity, and resting time, construct a behavioral feature vector. Step A4: Concatenate the environmental feature vector, spectral feature vector, and behavioral feature vector to obtain the micro-ecological state tensor; Step A5: Perform time series expansion on the microecological state tensor, and construct a time series tensor based on the microecological state tensor within 24 hours.

3. The method for dynamic prevention and control of coccidiosis in chickens according to claim 2, characterized in that, In step A2, the formula for the coccidia oocyst index is: in, , These are the characteristic reflectance wavelengths of lipids and proteins in the coccidia oocyst wall. Where R is the background interference wavelength, and R is the reflectivity; The formula for the bleeding index is: in, For wavelength, wavelength The first derivative of the reflectivity at that point, For wavelength Accumulate points. , These are the start and end wavelengths of the characteristic absorption band of oxyhemoglobin, respectively.

4. The method for dynamic prevention and control of coccidiosis in chickens according to claim 1, characterized in that, In step 3, a graph attention network is used to dynamically embed and reason about the pathological knowledge graph to obtain the pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability, and drug resistance evolution factor, including the following sub-steps: Step B1: Construct a heterogeneous graph, which includes environmental entity nodes, pathogen entity nodes, host entity nodes, microbial community entity nodes, and drug entity nodes; Step B2: Calculate the attention coefficients between environmental entity nodes, pathogen entity nodes, host entity nodes, microbial entity nodes, and drug entity nodes using a graph attention network, and then normalize them using the Softmax function to obtain the attention weights. Step B3: Update the node feature vectors by weighted summation based on attention weights; Step B4: The graph attention network maps the updated node feature vectors through a fully connected layer to obtain the pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability, and drug resistance evolution factor.

5. The method for dynamic control of coccidiosis in chickens according to claim 4, characterized in that, The formula for calculating the attention coefficient is: in, and These are feature vectors for environmental entity nodes, pathogen entity nodes, host entity nodes, microbial community entity nodes, or drug entity nodes, respectively. It is a non-linear activation function. To share the weight matrix, Let || be the attention weight vector, and || denote the concatenation operation; The formula for calculating attention weights is: in, For nodes The set of neighboring nodes, For the natural constant An exponential function with base 0; The formula for updating the node feature vector using weighted summation is as follows: in, This is the updated node feature vector. It is a non-linear activation function. For attention weights, To share the weight matrix, For nodes eigenvectors, For nodes The set of neighboring nodes.

6. The method for dynamic prevention and control of coccidiosis in chickens according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step C1: Use the intestinal substance transport submodel to determine the real-time concentration distribution of the drug in the intestine; Step C2: Based on the real-time concentration distribution of drugs in the gut, the microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor are input into the coccidia-microbe-immune cell interaction sub-model to obtain the coccidia biomass evolution curve, immune intensity evolution curve and microbe abundance evolution curve. Step C3: Based on the evolution curves of coccidia biomass, immune intensity, and microbial abundance, the evolution curve of intestinal lesion score is calculated using the intestinal lesion score mapping formula. Step C4: Based on the coccidia biomass evolution curve and historical drug use records, the drug resistance risk calculation formula is used to calculate the drug resistance development trend; Step C5: Based on the coccidia biomass evolution curve, intestinal lesion score evolution curve, and drug resistance development trend, obtain the intestinal lesion score, drug resistance risk value, and predicted coccidia oocyst density value.

7. The method for dynamic control of coccidiosis in chickens according to claim 1, characterized in that, In step 5, the machine learning algorithm is the Proximal Policy Optimization (PPO) algorithm, the Genetic Algorithm, the Particle Swarm Optimization (PSO) algorithm, or the Simulated Annealing (SAU) algorithm.

8. The method for dynamic prevention and control of coccidiosis in chickens according to claim 1, characterized in that, In step 5, a multi-objective optimization function is used to optimize and train the machine learning algorithm. The multi-objective optimization function is: in, To predict daily weight gain, The intestinal lesion score, For economic costs, This represents the risk value for drug resistance. , , and All are weighting coefficients.

9. A dynamic control system for coccidiosis in chickens, characterized in that, The method for dynamic prevention and control of coccidiosis in chickens according to any one of claims 1-8 includes the following units: Data acquisition module: used to acquire environmental parameter data, excrement hyperspectral data, and flock behavior image data; State Tensor Construction Module: Used to analyze, process, and stitch together environmental parameter data, excrement hyperspectral data, and chicken flock behavior image data to obtain a microecological state tensor. Then, the microecological state tensor is extended by time series to obtain a time series tensor. Knowledge Computation Module: Used to construct a pathological knowledge graph. Based on the microecological state tensor, a graph attention network is used to dynamically embed and reason about the pathological knowledge graph to obtain pathogen pathogenicity coefficient, host susceptibility index, coccidiosis outbreak risk probability and drug resistance evolution factor. Digital twin simulation module: used to input microecological state tensor, time series tensor, pathogen pathogenicity coefficient, host susceptibility index and drug resistance evolution factor into digital twin model. The digital twin model outputs intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value. Intelligent decision-making module: This module is used to input intestinal lesion score, drug resistance risk value and predicted coccidia oocyst density value into machine learning algorithm, which generates prevention and control strategy.

Images