Tan sheep scale farm epidemic disease risk prediction method and system
By constructing a dynamic graph model and integrating multi-source data from large-scale Tan sheep farms, and utilizing graph convolutional networks and long short-term memory networks, accurate prediction of disease risks in large-scale Tan sheep farms was achieved. This solved the problem of inaccurate risk assessment in existing technologies and improved early warning accuracy and early identification capabilities.
Patent Information
- Application Number
- CN202511854192.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-03-06
AI Technical Summary
Currently, the risk prediction of diseases in large-scale Tan sheep farms lacks integration of multi-source data and ignores the interconnected networks between farms, resulting in inaccurate risk assessment, weak early identification capabilities, and insufficient accuracy in early warning.
A dynamic graph model is constructed to integrate internal and external data from farms. By using graph convolutional networks and long short-term memory networks, combined with a three-branch output head, quantitative prediction of disease type, onset time, and scale can be achieved.
It has improved the sensitivity of early risk identification, accurately depicted the transmission path of diseases, significantly reduced the probability of major disease outbreaks, and supported the healthy development of the Tan sheep industry.
Smart Images
Figure CN121617657A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Tan sheep breeding technology, and more specifically to a method and system for predicting the risk of disease in large-scale Tan sheep farms. Background Technology
[0003] There are still significant shortcomings in the current disease risk prediction work for large-scale Tan sheep farms. Specifically, these shortcomings are as follows: a large amount of data related to disease risk has been accumulated throughout the entire Tan sheep breeding process, but this data is mostly scattered across different stages such as breeding management, testing and monitoring, and circulation supervision. Insufficient integration of multi-source data means that the risk information contained in the data cannot be fully extracted; existing prevention and control predictions mostly focus on the static risk assessment of a single farm, ignoring the interconnected networks formed between farms through live animal transportation and personnel movement, making it difficult to identify the risk of cross-farm transmission; current risk assessments still rely heavily on traditional experience or single-dimensional data, and a quantitative prediction model adapted to the Tan sheep breeding scenario has not yet been formed, resulting in weak early risk identification capabilities and insufficient accuracy in early warning.
[0004] Therefore, how to integrate multi-source breeding data of Tan sheep and improve the accuracy of disease risk prediction in large-scale Tan sheep farms is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides a method and system for predicting the disease risk of large-scale Tan sheep farms to solve the above problems.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention discloses a method for predicting the disease risk of large-scale Tan sheep farms, including: Collect historical breeding data from various large-scale Tan sheep farms. The historical breeding data includes internal farm data, external farm data, and disease data. Using each large-scale Tan sheep farm as a node, the information of each node is determined based on the internal data of the farm and the disease data, and the relationship between each node is determined based on the external data of the farm, thus constructing a dynamic graph; The constructed dynamic graph is sampled by a time sliding window of a set duration to obtain several training samples, which are then divided into a training set and a test set. The risk prediction model is trained and tested using the training set and the test set. Obtain the current breeding data and organize and update the dynamic graph, and use the trained risk prediction model to predict the disease risk of each large-scale Tan sheep farm.
[0007] Furthermore, the internal data of the farm includes: the number of Tan sheep raised, the month of raising, vaccination data, sampling and testing data of various pathogens, and temperature and humidity data of the farm; the external data of the farm includes: the geographical coordinates of the farm, the data on the transportation of live sheep to the farm, and the data on the movement of farm personnel; the disease data includes the type of disease, the time of onset of the disease, and the scale of the disease.
[0008] Furthermore, the construction of the dynamic graph includes constructing a geographic relationship graph, constructing a dynamic relationship graph, and merging dynamic graphs; The construction of the geographic relationship map specifically involves: determining the node set of large-scale Tan sheep farms. ,in To indicate the first Node information for each scale field; determine the edge set of the geographic relationship graph based on the geographic coordinates of each node. , Nodes in a geographic relationship graph With nodes The weights of the edges between them are determined by the adjacency matrix. Quantification, medium elements Determined by the following formula: ;in For nodes With nodes The straight-line distance between them Distance threshold; Geographic relationship diagram represented as ; The construction of the dynamic relationship graph specifically involves: determining the edge set of the dynamic relationship graph based on live sheep transportation data from large-scale farms and data on the movement of farmers. , Representing the time in the dynamic relationship graph node With nodes The edges between them, with edge weights determined by the temporal adjacency matrix. Quantification; medium elements From time The weighted calculation of transportation frequency and personnel turnover within the previous sliding window is as follows: ;in, These are the weighting coefficients. For a moment Front window inner scale field with scale field Standardized frequency of transportation For a moment Front window inner scale field with scale field Standardized frequency of personnel exchanges; dynamic relationship diagram represented as ; The dynamic graph fusion specifically involves using Hadamard product to fuse the geographic relationship graph and the dynamic relationship graph to obtain the time-based graph. fusion matrix By adding a self-looping term to optimize the propagation properties, a dynamic graph adjacency matrix is obtained. The formula is: ; in, It represents the Hadamah accumulation. for identity matrix For hyperparameters; final time The dynamic graph is represented as .
[0009] Furthermore, the risk prediction model includes a spatial component, a temporal component, and a prediction component; The spatial portion, based on Constructing a layered graph convolutional network, with time step 1 as the input. Node feature matrix Adjacency matrix of the dynamic graph The output is a spatial feature embedding; The time component is constructed based on a long short-term memory network. The input is a time sequence constructed based on the spatial feature embedding, and the output is a node comprehensive embedding that integrates spatiotemporal information. The prediction part consists of a shared hidden layer and a three-branch output head. The input is the node integration embedding of the fused spatiotemporal information, and the output is the disease type, onset time, and scale of each node.
[0010] Furthermore, in the spatial portion, for time... Animated GIF , No. Output of layer graph convolution The formula is: ; in, It is the ReLU activation function; , ; For the first Weight parameters for layer graph convolution.
[0011] Furthermore, the processing procedure for the time portion is as follows: First, construct the input sequence, and take continuous sequences within the time sliding window. The output of the spatial portion at each moment spliced into a time series ,in For a moment The output of the spatial part; Then, the data is fed into a Long Short-Term Memory (LSTM) network for feature extraction. The hidden state update process of the LSTM is as follows: ; ; ; ; ; ; in, , , , and They are respectively The results of input gate, forget gate, output gate, candidate cells, and cell update at each step. for The result of constant cell renewal; , , and These represent the weights of the input gate, forget gate, output gate, and candidate cell state, respectively. , , and For the corresponding hidden state interaction weight, , , and For the corresponding bias term; for middle Spatial characteristics at any given time; and All are activation functions. Represents the element-wise product; This represents the hidden state of the LSTM at the previous time step. This is the final output of the LSTM.
[0012] Furthermore, the shared hidden layer of the prediction part adopts a single fully connected network with ReLU activation function to integrate spatiotemporal features; the three-branch output head of the prediction part includes: The disease type output header uses a single-layer fully connected network with softmax activation function to output the probability distribution of each disease type. The onset time output header uses a single-layer fully connected network with no activation function to directly output the predicted future onset time value. The disease scale output head uses a single-layer fully connected network with the sigmoid activation function to output quantified disease scale values, where 0-0.3 corresponds to mild, 0.3-0.7 corresponds to moderate, and 0.7-1 corresponds to severe.
[0013] Furthermore, the method of using a trained risk prediction model to predict the disease risk of various Tan sheep farms specifically includes: Obtain the aquaculture data for the current time period and use linear interpolation to fill in the missing values; Update the dynamic graph according to the method of claim 3, and determine the updated node feature matrix. Adjacency Matrix of Dynamic Graph ; Will and The spatial portion is input, and spatial features are extracted using a graph convolutional network to obtain a real-time spatial embedding. Embed the historical spatial features extracted from the dynamic graph before the update and spliced into a time series Input the time portion to obtain the node comprehensive embedding of real-time fused spatiotemporal information. ;Will By inputting the prediction part, the prediction results of each node are obtained, and the onset time and scale prediction results of various types of diseases in each Tan sheep farm are determined.
[0014] This invention also discloses a disease risk prediction system for large-scale Tan sheep farms, comprising: Data acquisition module: Collects historical breeding data from various large-scale Tan sheep farms. The historical breeding data includes internal farm data, external farm data, and disease data. Dynamic graph module: Taking each large-scale Tan sheep farm as a node, the information of each node is determined based on the internal data of the farm and the disease data, and the relationship between each node is determined based on the external data of the farm to construct a dynamic graph; Model training module: The constructed dynamic graph is sampled by a time sliding window of a set duration to obtain several training samples, which are divided into training set and test set; the risk prediction model is trained and tested using the training set and test set. Risk prediction module: Acquires breeding data for the current period and organizes and updates dynamic graphs, and uses a trained risk prediction model to predict the disease risk of each large-scale Tan sheep farm.
[0015] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method and system for predicting the disease risk of large-scale Tan sheep farms, which has the following technical effects: (1) This invention integrates the internal breeding conditions of the farm, the external network of connections, and historical data of diseases, which solves the pain points of the traditional prediction of scattered multi-source data and insufficient risk information mining. It allows the disease correlation patterns hidden in the data to be fully released, greatly improves the sensitivity of early risk identification, and can capture potential hidden dangers in advance, leaving sufficient window period for the implementation of prevention and control measures. (2) The dynamic graph of this invention integrates the static geographical correlation with the dynamic information of transportation and personnel movement, accurately depicts the potential path of disease transmission between farms, effectively identifies the risk of cross-farm chain transmission, and makes up for the shortcomings of traditional static assessment that ignores the correlation between farms. (3) The spatiotemporal fusion model of this invention extracts spatial features through graph convolutional networks, captures temporal patterns through LSTM, and is equipped with a three-branch output head to realize the quantitative prediction of disease type, onset time, and disease scale. The accuracy is far superior to traditional experience judgment or single-dimensional models. Moreover, the data missing completion and dynamic graph update mechanism ensures adaptability and real-time performance in complex breeding scenarios. Applying this method can significantly reduce the probability of outbreaks of major diseases such as brucellosis, reduce economic losses in animal husbandry, and provide scientific support for the construction of "disease-free zones" and the large-scale and healthy development of the Tan sheep industry.
[0016] This invention achieves accurate and comprehensive prediction of disease risks in large-scale Tan sheep farms through collaborative innovation of multi-source data integration, dynamic graph modeling, and spatiotemporal fusion prediction. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0018] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] This invention discloses a method for predicting the risk of disease in large-scale Tan sheep farms, such as... Figure 1 As shown, it includes: Collect historical breeding data from various large-scale Tan sheep farms. The historical breeding data includes internal data from the farms, external data from the farms, and disease data. Using each large-scale Tan sheep farm as a node, the information of each node is determined based on internal farm data and disease data, and the relationship between each node is determined based on external farm data, thus constructing a dynamic graph. The constructed dynamic graph is sampled by a time sliding window of a set duration to obtain several training samples, which are then divided into training set and test set; the risk prediction model is trained and tested using the training set and test set. Obtain the current breeding data and organize and update the dynamic graph, and use the trained risk prediction model to predict the disease risk of each large-scale Tan sheep farm.
[0021] In one specific embodiment, the internal data of the farm includes: the number of Tan sheep raised, the month of raising, vaccination data, sampling and testing data of various pathogens, and temperature and humidity data of the farm; the external data of the farm includes: the geographical coordinates of the farm, the data on the transportation of live sheep to the farm, and the data on the movement of farm personnel; the disease data includes the type of disease, the time of onset of the disease, and the scale of the disease.
[0022] In one specific embodiment, constructing a dynamic graph includes constructing a geographic relationship graph, constructing a dynamic relationship graph, and merging dynamic graphs; Constructing the geographic relationship map specifically involves: determining the node set of large-scale Tan sheep farms. ,in To indicate the first Node information for each scale field; determine the edge set of the geographic relationship graph based on the geographic coordinates of each node. , Nodes in a geographic relationship graph With nodes The weights of the edges between them are determined by the adjacency matrix. Quantification, medium elements Determined by the following formula: ;in For nodes With nodes The straight-line distance between them Distance threshold; Geographic relationship diagram represented as .
[0023] Specifically, the number of Tan sheep raised, the months of raising, vaccination data, sampling and testing data for various pathogens, and temperature and humidity data of the farm are all converted into calculable node feature items. The specific processing method is as follows: The number of Tan sheep raised is directly taken as the total number of Tan sheep in the current scale farm. In order to eliminate the impact of the difference in farm size on the subsequent model, Min-Max normalization is used to transform it into a numerical feature in the 0-1 interval. The breeding month is mapped to a numerical label according to the season, or the month value (1-12) is taken directly to reflect the impact of season / month on the occurrence of disease; Vaccination data, including vaccine type and the number of days since the last vaccination; if not vaccinated, the maximum value of the interval is used. For sampling and testing data of various pathogens, for common infectious diseases of Tan sheep such as Brucella, peste des petits ruminants virus, and sheep pox virus, the most recent negative sampling and testing result is recorded as 0, and positive result is recorded as 1. If there are multiple tests, the average of the test results within the past 3 months is used (e.g., if 2 out of 3 tests are negative, it is recorded as 0.33). The temperature and humidity data of the farm are taken as the average daily temperature and average daily relative humidity of the most recent 7 days and directly used as the environmental feature items of the node. If there is a missing data for a single day, linear interpolation of the data of the two adjacent days is used to supplement it.
[0024] Data on disease type, onset time, and disease scale are used to extract historical disease risk characteristics of nodes, serving as an important supplement to node information. The specific processing method is as follows: Historical disease type characteristics are processed using unique thermal coding. For example, if a node has experienced brucellosis, the "brucellosis characteristic item" is recorded as 1; if it has not occurred, it is recorded as 0, thus forming a multi-dimensional disease type label. Historical disease occurrence time characteristics: extract the start and end times of the disease. If no epidemic has occurred, the historical disease frequency is recorded as 0. Historical disease scale characteristics are categorized based on the proportion of affected sheep to the total number of Tan sheep in stock: 0-0.3 for mild, 0.3-0.7 for moderate, and 0.7-1 for severe. If no disease has occurred, it is recorded as 0.
[0025] By integrating the basic features based on internal data with the historical risk features based on disease data, information for each Tan sheep farm node is formed, namely a feature vector with a dimension of 1×M, where M is the total number of feature items. This feature vector fully reflects the individual farm's own breeding conditions, environmental status, and historical disease risks, providing core data support for the subsequent construction of the node feature matrix of the dynamic graph.
[0026] Constructing the dynamic relationship graph involves: determining the edge set of the dynamic relationship graph based on live sheep transportation data from large-scale farms and data on the movement of farmers. , Representing the time in the dynamic relationship graph node With nodes The edges between them, with edge weights determined by the temporal adjacency matrix. Quantification; medium elements From time The weighted calculation of transportation frequency and personnel turnover within the previous sliding window is as follows: ;in, These are the weighting coefficients. For a moment Front window inner scale field with scale field Standardized frequency of transportation For a moment Front window inner scale field with scale field Standardized frequency of personnel exchanges; dynamic relationship diagram represented as .
[0027] Specifically, when there is direct transportation of breeding sheep and lambs between two farms, it is counted as 1 standardized transportation frequency; if it is indirect transportation, such as two farms simultaneously transporting live sheep to quarantine, trading, or slaughtering locations, it is counted as 0.5 standardized transportation frequencies. The summation of all standardized transport frequencies within the previous sliding window is obtained. When there is direct personnel interaction between two farms of different sizes, it is recorded as 1 standardized personnel interaction frequency. If the interaction is indirect, such as when farm workers simultaneously visit the same place (e.g., a feed mill, veterinary drug store, industry conference, or technical training site), it is recorded as 0.5 standardized personnel interaction frequencies. The summation of the standardized frequencies of all people coming and going within the sliding window is obtained. .
[0028] The dynamic graph fusion process involves using Hadamard product to fuse the geographic relationship graph with the dynamic relationship graph to obtain the time-series graph. fusion matrix By adding a self-looping term to optimize the propagation properties, a dynamic graph adjacency matrix is obtained. The formula is: ; in, It represents the Hadamah accumulation. for identity matrix For hyperparameters; final time The dynamic graph is represented as .
[0029] In one specific embodiment, the risk prediction model includes a spatial component, a temporal component, and a prediction component; The spatial component, based on Constructing a layered graph convolutional network, with time step 1 as the input. Node feature matrix Adjacency Matrix of Dynamic Graph The output is a spatial feature embedding; For the temporal component, it is constructed based on a long short-term memory network. The input is a temporal sequence constructed based on spatial feature embedding, and the output is a node comprehensive embedding that integrates spatiotemporal information. The prediction part consists of a shared hidden layer and a three-branch output head. The input is a node comprehensive embedding that integrates spatiotemporal information, and the output is the disease type, onset time, and scale of each node.
[0030] In one specific embodiment, in the spatial portion, for time... Animated GIF , No. Output of layer graph convolution The formula is: ; in, It is the ReLU activation function; , ; For the first Weight parameters for layer graph convolution.
[0031] In one specific embodiment, the time portion of the processing is as follows: First, construct the input sequence, and take continuous sequences within the time sliding window. Output of the spatial part at each time point spliced into a time series ,in For a moment The output of the spatial part; Then, the data is fed into a Long Short-Term Memory (LSTM) network for feature extraction. The hidden state update process of the LSTM is as follows: ; ; ; ; ; ; in, , , , and They are respectively The results of input gate, forget gate, output gate, candidate cells, and cell update at each step. for The result of constant cell renewal; , , and These represent the weights of the input gate, forget gate, output gate, and candidate cell state, respectively. , , and For the corresponding hidden state interaction weight, , , and For the corresponding bias term; for middle Spatial characteristics at any given time; and All are activation functions. Represents the element-wise product; This represents the hidden state of the LSTM at the previous time step. This is the final output of the LSTM.
[0032] In one specific embodiment, the shared hidden layer of the prediction part adopts a single fully connected network with ReLU activation function to integrate spatiotemporal features; the three-branch output head of the prediction part includes: The disease type output header uses a single-layer fully connected network with softmax activation function to output the probability distribution of each disease type. The onset time output header uses a single-layer fully connected network with no activation function to directly output the predicted future onset time value. The disease scale output head uses a single-layer fully connected network with the sigmoid activation function to output quantified disease scale values, where 0-0.3 corresponds to mild, 0.3-0.7 corresponds to moderate, and 0.7-1 corresponds to severe.
[0033] In one specific embodiment, a trained risk prediction model is used to predict the disease risk of each large-scale Tan sheep farm, specifically including: Obtain the aquaculture data for the current time period and use linear interpolation to fill in the missing values; Update the dynamic graph according to the method of claim 3, and determine the updated node feature matrix. Adjacency Matrix of Dynamic Graph ; Will and For the input spatial part, spatial features are extracted using a graph convolutional network to obtain real-time spatial embeddings. Embed the historical spatial features extracted from the dynamic graph before the update and spliced into a time series Input the time portion to obtain the node comprehensive embedding of real-time fused spatiotemporal information. ;Will The prediction results of each node are obtained from the input prediction section, and the onset time and scale prediction results of various types of diseases in each Tan sheep farm are determined.
[0034] This invention also discloses a disease risk prediction system for large-scale Tan sheep farms, comprising: Data acquisition module: Collects historical breeding data from various large-scale Tan sheep farms. The historical breeding data includes internal data, external data, and disease data of the farms. Dynamic graph module: Taking each large-scale Tan sheep farm as a node, the information of each node is determined based on internal farm data and disease data, and the relationship between each node is determined based on external farm data to construct a dynamic graph; Model training module: The constructed dynamic graph is sampled by a time sliding window of a set duration to obtain several training samples, which are divided into training set and test set; the risk prediction model is trained and tested using the training set and test set. Risk prediction module: Acquires breeding data for the current period and organizes and updates dynamic graphs, and uses a trained risk prediction model to predict the disease risk of each large-scale Tan sheep farm.
[0035] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0036] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for predicting the risk of epidemic diseases in a Tan sheep scale farm, characterized by, The method comprises the following steps: Collecting historical breeding data of each Tan sheep scale farm, wherein the historical breeding data comprises internal data of the breeding farm, external data of the breeding farm and epidemic data; Taking each Tan sheep scale farm as a node, determining node information based on the internal data of the breeding farm and the epidemic data, determining the association between nodes based on the external data of the breeding farm, and constructing a dynamic graph; Sampling the constructed dynamic graph through a time sliding window with a set time length to obtain a plurality of training samples, and dividing the training samples into a training set and a test set; training a risk prediction model and testing the risk prediction model by using the training set and the test set; Obtaining breeding data of a current period and updating the dynamic graph, and predicting the epidemic risk of each Tan sheep scale farm by using the trained risk prediction model.
2. The method according to claim 1, wherein, The internal data of the breeding farm comprises the number of Tan sheep, breeding months, vaccination data, sampling inspection data of each pathogen, and temperature and humidity data of the breeding farm; the external data of the breeding farm comprises geographic coordinates of the scale farm, live sheep transportation data of the scale farm and breeding personnel data; and the epidemic data comprises epidemic type, onset time and epidemic scale data.
3. The method according to claim 1, wherein, The dynamic graph comprises a geographic relationship graph, a dynamic relationship graph and a dynamic graph fusion; The constructing the geographical relation graph specifically comprises: determining a node set of the Tan sheep scale field , wherein is node information of a first scale field; according to geographical coordinates of each node, a geographical relation graph edge set is determined , wherein represents an edge between a node and a node in the geographical relation graph, and a weight of each edge is quantified by an adjacency matrix , wherein an element in the matrix is determined by the following formula: , wherein is a straight-line distance between the node and the node , and is a distance threshold value; the geographical relation graph is represented as ; The construction of the dynamic relationship graph specifically involves: determining the edge set of the dynamic relationship graph based on live sheep transportation data from large-scale farms and data on the movement of farmers. , Representing the time in the dynamic relationship graph node With nodes The edges between them, with edge weights determined by the temporal adjacency matrix. Quantification; Middle elements From time The weighted calculation of transportation frequency and personnel turnover within the previous sliding window is as follows: ;in, These are the weighting coefficients. For a moment Front window inner scale field with scale field Standardized frequency of transportation For a moment Front window inner scale field with scale field Standardized frequency of personnel exchanges; dynamic relationship diagram represented as ; The dynamic graph fusion specifically includes: adopting Hadamard product to fuse the geographic relation graph and the dynamic relation graph to obtain a fusion matrix of a time point , and adding a self-loop item to optimize the propagation characteristics to obtain a dynamic graph adjacency matrix , and the formula is: ； wherein, denotes a Hadamard product, is an identity matrix, is a hyperparameter; the dynamic graph representation at the final time is .
4. The method according to claim 3, wherein, The risk prediction model comprises a spatial part, a time part and a prediction part; The space part is based on Layer graph convolution network construction, input is time Node feature matrix And the dynamic graph adjacency matrix , Output is spatial feature embedding; The time part is constructed based on a long short-term memory network, the input is a time sequence constructed based on the spatial feature embedding, and the output is a node comprehensive embedding fused with space-time information; The prediction part comprises a shared hidden layer and a three-branch output head, the input is the node comprehensive embedding fused with space-time information, and the output is the epidemic type, onset time and epidemic scale of each node.
5. The method according to claim 4, wherein, In the spatial part, for time instant of the dynamic graph , the output of the first layer graph convolution is formulated as: ； wherein, is a ReLU activation function; , ; is a weight parameter of the layer graph convolution.
6. The method according to claim 4, wherein, The processing process of the time part is as follows: First, the input sequence is constructed, taking the outputs of the spatial parts at consecutive time instants within a time sliding window are concatenated into a temporal sequence where is the output of the spatial part at time instant Then, the input is subjected to feature extraction by a long short-term memory network, and the hidden state updating process of the LSTM is as follows: ； ； ； ； ； ； in, , , , and They are respectively The results of input gate, forget gate, output gate, candidate cells, and cell update at each step. for The result of constant cell renewal; , , and These represent the weights of the input gate, forget gate, output gate, and candidate cell state, respectively. , , and For the corresponding hidden state interaction weight, , , and For the corresponding bias term; for middle Spatial characteristics at any given time; and All are activation functions. Represents the element-wise product; This represents the hidden state of the LSTM at the previous time step. This is the final output of the LSTM.
7. The method according to claim 4, wherein, The shared hidden layer of the prediction part adopts a one-layer full connection network, and the activation function is ReLU, which is used for integrating space-time features; The three-branch output head of the prediction part comprises: An epidemic type output head adopts a one-layer full connection network, and the activation function is softmax, which is used for outputting the probability distribution of each epidemic type; An onset time output head adopts a one-layer full connection network without an activation function, and directly outputs the predicted future onset day value; An epidemic scale output head adopts a one-layer full connection network, and the activation function is sigmoid, which is used for outputting a quantitative epidemic scale value, wherein 0-0.3 corresponds to mild, 0.3-0.7 corresponds to moderate, and 0.7-1 corresponds to severe.
8. The method according to claim 4, wherein, The prediction of the epidemic risk of each Tan sheep scale farm by using the trained risk prediction model comprises the following steps: Obtaining the breeding data of the current period and supplementing the missing values by using a linear interpolation method; updating the dynamic graph according to the method of claim 3, and determining the updated node feature matrix and the dynamic graph adjacency matrix ; The input space part is input into the graph convolution network to extract spatial features to obtain real-time spatial embedding and The input space part is input into the graph convolution network to extract spatial features to obtain real-time spatial embedding The historical spatial feature embedding extracted by the dynamic graph before updating and are spliced into a time sequence The input time part is input to obtain a node comprehensive embedding fused with real-time spatio-temporal information The input time part is input to obtain a node comprehensive embedding fused with real-time spatio-temporal information The input prediction part obtains the prediction results of each node, and determines the prediction results of the incidence time and scale of each type of epidemic disease of each Tan sheep scale farm.
9. A system for predicting the risk of disease in a Tan sheep scale farm, applying the method for predicting the risk of disease in a Tan sheep scale farm according to any one of claims 1 to 8, characterized in that, The method comprises the following steps: A data collection module is configured to collect historical breeding data of each Tan sheep scale farm, wherein the historical breeding data comprises internal data of the breeding farm, external data of the breeding farm and epidemic data; A dynamic graph module is configured to take each Tan sheep scale farm as a node, determine node information based on the internal data of the breeding farm and the epidemic data, determine the association between nodes based on the external data of the breeding farm, and construct a dynamic graph; and The model training module: a time window of a set length is set to sample the constructed dynamic graph to obtain a plurality of training samples, which are divided into a training set and a test set; the risk prediction model is trained and tested by using the training set and the test set; The risk prediction module: the breeding data of the current period are obtained and the dynamic graph is updated, and the trained risk prediction model is used to predict the disease risk of each Tan sheep scale farm.