A method for predicting the risk of pathogen transmission in a farm

Abstract

The application discloses a farm pathogen transmission risk prediction method, comprising the following steps: acquiring field area object data collected by a terminal, constructing a candidate transmission topology according to field area layout, ventilation relationship and traffic transfer route; screening data at a prediction reference time, generating pathogen detection, trajectory and environmental evidence confidence, and determining pathogen evidence state; activating candidate transmission edges according to personnel, vehicles, materials, airflow and disinfection data, calculating edge activation coefficients, and constructing a multi-relation spatiotemporal directed transmission graph of a continuous window; inputting the transmission graph into a spatiotemporal transmission graph deep neural network, processing through multi-relation directed graph convolution, confidence gating and timing gating, and outputting pathogen transmission risk values of the next prediction period; and sending control instructions to sampling, environmental monitoring, access control or ventilation control terminals according to the risk values and pathogen evidence state. The application improves the time consistency of multi-source data and the matching degree of risk judgment and field area transmission process, and realizes linkage between risk prediction results and prevention and control terminals.

Description

Technical Field

[0001] This invention relates to the field of biosafety monitoring technology in livestock farms, and in particular to a method for predicting the risk of pathogen transmission in livestock farms. Background Technology

[0002] With the construction of large-scale, intensive livestock farms, there are cross-connections among pens, isolation areas, feed and water supply areas, manure treatment areas, and entrance / exit areas, involving personnel activities, transportation operations, air exchange, and material flow. Pathogen-related risks are usually not formed in a single location or at a single moment, but are affected by the farm layout, work sequence, environmental changes, and the implementation of disease prevention procedures, exhibiting cross-regional, temporal, and uncertain characteristics.

[0003] Current biosecurity management in livestock farms relies heavily on fragmented management based on periodic inspections, environmental instrument readings, entry and exit records, and manual patrols. Different sampling devices vary in sampling cycles, data formats, coverage, and reliability; some data may lose value due to equipment offline, collection delays, fluctuations in test quality, or missing records. Existing processing methods typically display various data separately or provide alerts based on single indicators, fixed rules, or static thresholds, making it difficult to combine the data's formation time, spatial location, sequence of activities, and reliability to make continuous judgments about the area, path, and timing of risk expansion. Furthermore, current early warning results are mostly limited to monitoring interface displays or manual notifications, failing to establish targeted linkages with sampling, entry and exit points, environmental inspections, and ventilation control equipment, resulting in delays in verification and response.

[0004] Therefore, how to transform farm monitoring information with heterogeneous sources, different time scales, and varying degrees of reliability into spatiotemporal risk assessment results that conform to the actual transmission and evolution process, and drive on-site equipment to perform corresponding prevention and control actions accordingly, has become an urgent technical problem to be solved. Summary of the Invention

[0005] In view of this, the present invention aims to provide a method for predicting the risk of pathogen transmission in livestock farms, so as to solve or alleviate the technical problems existing in the prior art.

[0006] The technical solution of this invention is implemented as follows: A method for predicting the risk of pathogen transmission in livestock farms, comprising the following steps: S1. Acquire the field object data collected by the terminal; construct a candidate propagation topology with field objects as graph nodes, including three types of candidate propagation edges: contact, airflow, and material transfer, according to the field layout, ventilation, and passage and transfer routes; S2. Determine the prediction baseline time, select data with timestamps no later than that time, generate evidence confidence, and generate positive anchoring, negative constraint, or invalid or missing status based on pathogen detection data; S3. Activate candidate propagation edges based on personnel, vehicles, materials, airflow and disinfection verification data and calculate the edge activation coefficient as the edge weight of the corresponding dynamic directed propagation edge; update the target graph nodes that are in a negative constraint, invalid or missing state and whose dynamic directed incoming edges come from positive anchor graph nodes and meet the time association conditions to a conflict pending re-inspection state, and construct a multi-relationship spatiotemporal directed propagation graph with continuous windows. S4. Input the propagation graph into the spatiotemporal propagation graph deep neural network, and perform directed aggregation according to edge type with different parameters and edge weights through the multi-relation directed graph convolutional layer; through the confidence gating layer, set the pathogen detection features corresponding to positive anchoring, negative constraint, conflict pending re-examination and invalid or missing states as high cross-node gating, pathogen detection only at this node gating, low cross-node gating with retaining conflict markers, and zero cross-node gating, respectively. Among them, negative constraint does not adjust the cross-node gating of personnel, vehicles, materials, airflow and environmental features; through the temporal gating neural network layer, output the risk value of each graph node in the next prediction period. S5. Based on the pathogen transmission risk value and pathogen evidence status, send control commands to sampling, environmental monitoring, access control, or ventilation control terminals.

[0007] Compared with the prior art, the present invention has the following advantages: This invention associates terminal data from different sources, such as pathogen detection, personnel and vehicle activities, material transportation, ventilation status, environmental monitoring, and disinfection verification, with field objects under a unified time benchmark. It also generates corresponding confidence levels by combining detection quality, terminal status, data completeness, and evidence validity. This allows data with different collection frequencies, data formats, and reliability levels to participate in risk inference according to their validity, avoiding the direct use of outdated, missing, or low-quality data as equivalent risk basis, and improving the temporal consistency and availability of risk input data.

[0008] This invention forms candidate propagation topologies based on site layout, ventilation relationships, personnel and vehicle traffic relationships, and material transfer relationships. It then activates corresponding propagation relationships by combining actual traffic, transfer, airflow, and disinfection verification data. Information such as dwell time, transfer interval, wind speed, pressure difference, filtration status, and disinfection verification status are converted into dynamic propagation relationships with direction and weight. This enables risk reasoning to reflect the actual activity status of different propagation carriers in different time segments, improving the matching degree between risk assessment and the actual propagation process in the farm.

[0009] This invention employs differentiated cross-node transmission control methods for pathogen detection characteristics in different pathogen evidence states, such as positive anchoring, negative constraint, conflict pending retesting, and invalid or missing data. Among them, the negative constraint state only restricts the pathogen detection characteristics of this node, without blocking the cross-node transmission of other characteristics such as personnel and vehicle activities, material transportation, airflow connectivity, and environmental conditions. This reduces the obscuring effect of a single negative test, insufficient test coverage, or missing data on risk assessment, and can promptly form a basis for retesting for objects in the field with transmission conflict conditions.

[0010] This invention obtains the pathogen transmission risk value for the next prediction period through multi-relation directed propagation reasoning, and determines the main risk contribution edge and the transmission path to be verified by calculating the risk contribution of dynamic propagation edges. Then, it sends corresponding control commands to automatic sampling, air sampling, access control, environmental monitoring and ventilation control terminals, so that the risk prediction results can be converted into prevention and control actions with target area objects, control terminals and control parameters. At the same time, through online model freezing, offline candidate model verification and performance rollback mechanism, the impact of the model update process on the continuity and stability of on-site risk prediction is reduced. Attached Figure Description

[0011] The accompanying drawings are provided to further understand the present invention and form part of the specification. They are used together with the embodiments of the present invention to explain the invention and do not constitute a limitation thereof.

[0012] Figure 1 This is a schematic diagram of the overall process of a method for predicting the risk of pathogen transmission in a livestock farm according to the present invention; Figure 2 This is a schematic diagram illustrating the process of generating candidate propagation topology, dynamic directed propagation edges, and conflict-pending-re-inspection states in this invention. Figure 3 This is a schematic diagram illustrating the feature gating, directed graph convolution, and temporal inference process of the spatiotemporal propagation graph deep neural network of the present invention. Figure 4 This diagram illustrates the process of determining the main risk contribution edge, generating terminal control instructions, and updating the double-buffered model, which is the main risk contribution edge of this invention. Detailed Implementation

[0013] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0014] Example 1: This embodiment provides a method for predicting the risk of pathogen transmission in livestock farms, including the following steps.

[0015] Step S1: Obtain the field object data collected by the terminal and construct the candidate propagation topology.

[0016] In this embodiment, the risk prediction server serves as the execution device for data aggregation, propagation graph construction, risk reasoning, determination of major risk contribution edges, and generation of control instructions. The risk prediction server is communicatively connected to pathogen detection terminals, environmental monitoring terminals, personnel and vehicle positioning terminals, material identification terminals, ventilation status acquisition terminals, disinfection status acquisition terminals, access control terminals, sampling terminals, and ventilation control terminals.

[0017] The risk prediction server acquires static configuration data and dynamic monitoring data corresponding to the objects in the site area. Static configuration data includes the location, adjacency relationships, functional areas, access relationships, ventilation duct connection relationships, and material transfer relationships of the objects in the site area. Dynamic monitoring data includes pathogen detection data, environmental monitoring data, personnel and vehicle trajectory data, material transfer data, ventilation status data, disinfection status data, and access control records.

[0018] The risk prediction server uses site objects as graph nodes, establishes candidate contact propagation edges based on the personnel and vehicle permission access relationship, establishes candidate airflow propagation edges based on the ventilation connectivity relationship between exhaust-side site objects and air-intake-side site objects, and establishes candidate material transfer propagation edges based on the material transfer sequence, thus obtaining the candidate propagation topology.

[0019] Candidate propagation edges only indicate that there is a potential propagation condition between two field objects, and do not indicate that propagation has already occurred in the current time period. Each candidate propagation edge is associated with and stored as source graph node identifier, target graph node identifier, edge type, physical path length, preset propagation time range, corresponding terminal type, and edge status fields.

[0020] Step S2: Determine the prediction baseline time, filter and time-align multi-source terminal data, and generate evidence confidence and pathogen evidence status.

[0021] The risk prediction server uses the prediction baseline time as the data cutoff time for the current prediction task, and only calls data packets with timestamps no later than the prediction baseline time to participate in the current propagation graph construction and risk inference. Data packets with timestamps later than the prediction baseline time are written to the data buffer of subsequent prediction tasks.

[0022] The risk prediction server aligns continuous monitoring data and event-based data according to a preset time granularity, associating data within the same time slice with corresponding graph nodes. For continuous measurement data such as temperature, humidity, ammonia concentration, pressure difference, wind speed, valve opening, and fan speed, environmental state characteristics are generated for corresponding graph nodes according to time slices. For event-based data such as personnel and vehicle passage records, material transfer records, disinfection records, and access control records, the data is written to the corresponding time slice according to the event occurrence time; for stay events, ventilation events, or disinfection events spanning multiple time slices, they are split according to duration and written to the corresponding time slices.

[0023] The risk prediction server generates the confidence level of pathogen detection evidence based on sample retention time, detection terminal self-test status, positive and negative control status, consistency of repeated tests, detection signal status, and sampling coverage; it generates the confidence level of trajectory evidence based on the integrity of positioning signals, number of positioning jumps, reasonableness of dwell time, trajectory continuity, and identity recognition matching status; and it generates the confidence level of environmental evidence based on sensor online status, calibration status, continuous acquisition integrity, and abnormal fluctuation status.

[0024] For graph nodes that meet the testing quality requirements, are within the valid evidence period, and have a positive test result, the risk prediction server marks the graph node as a positive anchoring state; for graph nodes that meet the testing quality requirements, have sufficient sampling coverage, are within the valid evidence period, and have a negative test result, they are marked as a negative constraint state; for graph nodes that do not have available pathogen detection data, whose pathogen detection data exceeds the valid evidence period, or whose testing quality does not meet the requirements, they are marked as invalid or missing.

[0025] Step S3: Activate candidate propagation edges, construct a multi-relationship spatiotemporal directed propagation graph, and update the conflict pending review status.

[0026] The risk prediction server determines whether candidate contact propagation edges, candidate airflow propagation edges, and candidate material transfer propagation edges meet the activation conditions within each time slice. Activated candidate propagation edges form dynamic directed propagation edges, and the propagation direction, edge type, edge activation coefficient, propagation arrival time, and the confidence level of the evidence corresponding to the edge are recorded.

[0027] For a target graph node that is in a negative constraint state, invalid state, or missing state, if there is a dynamic directed propagation edge pointing from a source graph node in a positive anchor state to the target graph node, and the edge activation coefficient of the dynamic directed propagation edge exceeds the preset edge activation threshold and the propagation arrival time meets the corresponding time association condition, then the target graph node will be updated to a conflict pending re-examination state, and the original negative pathogen detection result or invalid or missing marker will be retained.

[0028] The risk prediction server combines graph nodes, dynamic directed propagation edges, node features, and edge features from multiple consecutive time slices in chronological order to form a multi-relationship spatiotemporal directed propagation graph with continuous windows.

[0029] Step S4: Input the multi-relationship spatiotemporal directed propagation graph into the spatiotemporal propagation graph deep neural network, and output the pathogen transmission risk value of each graph node in the next prediction period.

[0030] The spatiotemporal propagation graph deep neural network includes a feature encoding layer, a confidence gating layer, a multi-relation directed graph convolutional layer, a temporal gating neural network layer, and a risk output layer.

[0031] The feature coding layer encodes the node features within each time slice, enabling pathogen detection features, personnel and vehicle activity features, material transfer features, airflow connectivity features, environmental state features, and state coding features to form a unified dimension of node representation.

[0032] The confidence gating layer generates gating coefficients corresponding to node features and edge features based on the status of pathogen evidence, the confidence level of pathogen detection evidence, the confidence level of trajectory evidence, the confidence level of environmental evidence, and the effective time period of evidence.

[0033] For pathogen detection features corresponding to positive anchor states, a cross-node gating coefficient of at least the second gating threshold is set, allowing these features to participate in cross-node aggregation with a high transmission strength. For pathogen detection features corresponding to conflict-pending-re-examination states, a cross-node gating coefficient greater than zero and lower than the first gating threshold is set, and conflict markers are retained. For pathogen detection features corresponding to invalid or missing states, the cross-node gating coefficient is set to zero. For pathogen detection features corresponding to negative constraint states, only their moderating effect on the current graph node's own features is retained, and they do not participate in cross-node pathogen detection information transmission.

[0034] The negative constraint state does not change the cross-node gating coefficient of personnel and vehicle activity characteristics, material transfer characteristics, airflow connectivity characteristics, and environmental state characteristics, so that the above non-pathogen detection characteristics still participate in cross-node propagation aggregation according to the confidence level of corresponding trajectory evidence, confidence level of environmental evidence, terminal online status, and edge activation status.

[0035] The multi-relation directed graph convolutional layer uses corresponding relational parameters for contact propagation edges, airflow propagation edges, and material transfer propagation edges, respectively, to process the incoming edge information, outgoing edge information, and the characteristics of the node itself. The edge activation coefficient, as the basic edge weight of the dynamic directed propagation edge, participates in message passing; the edge feature gating result is used to adjust the actual information transmission intensity of the dynamic directed propagation edge without changing the edge's activation state, propagation direction, or edge type.

[0036] The temporal gating neural network layer reads the graph convolutional outputs from multiple consecutive time slices in chronological order, extracting changes in pathogen detection, personnel and vehicle activity, material transfer, airflow, and environmental conditions for each graph node within a continuous window. The risk output layer outputs the pathogen transmission risk value for each graph node in the next prediction period based on the temporal propagation characteristics.

[0037] Step S5: Based on the pathogen transmission risk value and pathogen evidence status, generate a risk management task and send control instructions.

[0038] When the pathogen transmission risk value of a graph node exceeds a preset risk threshold, or when a graph node is in a conflict pending re-inspection state, the risk prediction server generates a risk handling task for the corresponding field object. The control instruction includes at least the target field object identifier, control terminal identifier, control type, control parameters, effective time, effective duration, and execution confirmation field.

[0039] After executing the control command, the terminal returns the execution status to the risk prediction server. The risk prediction server only writes the terminal execution status with a timestamp no later than the subsequent prediction baseline into the node features, edge features, or training samples of the corresponding prediction task, avoiding writing future execution status into the current prediction task.

[0040] Example 2: This embodiment, based on Embodiment 1, explains the construction method of field objects, terminal configuration, terminal data transmission relationship and candidate propagation topology.

[0041] The site includes at least one of the following: breeding pens, isolation area, feed storage area, drinking water supply area, manure transfer area, personnel changing and disinfection area, vehicle disinfection area, material handover area, ventilation intake area, and ventilation exhaust area.

[0042] Each field object is configured with a unique field object identifier, and a mapping relationship is established between the field object identifier and the terminal identifier. When one field object corresponds to multiple terminals, the data collected by each terminal is associated with that field object; when one terminal serves multiple field objects, the terminal data packet carries the sampling location, direction of travel, source of materials, destination of materials, or location of the ventilation duct, so that the risk prediction server can associate the data with the corresponding field object.

[0043] The sampling terminals include automatic sampling terminals and air sampling terminals. Automatic sampling terminals collect pathogen detection samples from the corresponding site area and upload the sample number, sampling time, sampling point identifier, sampling terminal status, and sampling execution results to the risk prediction server. Air sampling terminals collect air samples from inside the site area, air inlets, air outlets, or ventilation path nodes, and upload the air sample number, sampling flow rate, sampling duration, sampling location, and sampling terminal status.

[0044] The pathogen detection terminal receives samples with sample numbers and, after completing the detection, sends the detection results, detection completion time, detection signal status, positive and negative control status, repeat detection results, and terminal self-test status to the risk prediction server. The risk prediction server associates the sampling records and detection results with the same sampling event based on the sample number, field object identifier, and sampling time.

[0045] For two site objects connected by a permitted route for personnel and vehicles, the risk prediction server establishes a candidate contact propagation edge with the direction consistent with the permitted route. For situations where there is a duct connection, wind direction connection, or air transport path between an exhaust-side site object and an intake-side site object, a candidate airflow propagation edge is established pointing from the exhaust-side site object to the intake-side site object. For two site objects with a forward and backward transfer relationship in a material transfer route, a candidate material transfer propagation edge with the direction consistent with the material transfer sequence is established.

[0046] Example 3: This embodiment, based on Embodiment 2, explains the methods for determining a unified clock, time alignment, evidence confidence level, and the effective period of evidence.

[0047] Each terminal generates data packets based on a unified clock reference. Each data packet includes at least a timestamp, terminal identification information, site object identification information, and validity identifier. Continuous measurement data packets also include the measured value, measurement unit, equipment self-test status, and calibration status; event-type data packets also include the event type, object identification, event occurrence time, and event duration.

[0048] In this embodiment, the time granularity is five minutes, the continuous window contains twelve time slices, and the next prediction period is thirty minutes. The risk prediction server uses the prediction baseline time as the data cutoff time for the current prediction task and divides the data within the sixty minutes before the prediction baseline time into continuous time slices. At each time granularity, the risk prediction server discards the earliest time slice and adds the latest time slice to the continuous window.

[0049] For numerical terminal data, the risk prediction server normalizes the data based on preset low-risk and high-risk reference values, converting data from different units and dimensions into a unified comparison range. For indicators where larger values ​​indicate higher risk, the closer the data value is to the high-risk reference value, the higher the corresponding risk characteristic value. For indicators where smaller values ​​indicate higher risk, risk characteristic values ​​are generated according to a reverse rule. For discrete data such as terminal online status, identity recognition matching status, positive / negative comparison status, and filtering status, corresponding status characteristics are generated according to whether the requirements are met, requirements are not met, or pending review.

[0050] The confidence level of pathogen detection evidence is determined comprehensively by the suitability of sample retention time, the self-test status of the detection terminal, the status of positive and negative controls, the consistency of repeated tests, the status of the detection signal, and the sampling coverage. The confidence level of trajectory evidence is determined comprehensively by the integrity of the positioning signal, location jumps, the reasonableness of the dwell time, the continuity of the trajectory, and the status of identity recognition matching. The confidence level of environmental evidence is determined comprehensively by the online status of the sensor, its calibration status, the completeness of continuous data acquisition, and the status of abnormal fluctuations.

[0051] The effective period of pathogen detection evidence is determined jointly based on the sample collection time, the test completion time, the actual sample storage time, the preset allowable storage time, and the preset effective period of evidence. The sample collection time serves as the starting time for the validity period of the evidence, and the test completion time is used to confirm that the test results are available before the predicted baseline time.

[0052] When the sample collection time and the detection completion time are both no later than the prediction baseline time, the actual sample storage time does not exceed the allowed storage time, and the interval between the sample collection time and the prediction baseline time does not exceed the preset valid evidence duration, the corresponding pathogen detection result is used as valid pathogen detection evidence in the current prediction task; in other cases, the corresponding pathogen detection result does not participate in the determination of the pathogen evidence status in the current prediction task.

[0053] Example 4: This embodiment, based on Embodiment 3, explains the activation conditions of candidate propagation edges and the method for determining edge activation coefficients.

[0054] Within each time slice, the risk prediction server performs activation judgments on candidate contact propagation edges, candidate airflow propagation edges, and candidate material transfer propagation edges, respectively.

[0055] When the same person, vehicle, or reusable tool passes through the first and second map nodes sequentially within a preset time interval, the risk prediction server activates the candidate contact propagation edge pointing from the first map node to the second map node, forming a dynamic contact propagation edge.

[0056] The edge activation coefficient of dynamic contact transmission edges is jointly determined by dwell time, cross-regional time interval, passage frequency, disinfection verification status, and trajectory evidence confidence level. The risk prediction server normalizes the dwell time, cross-regional time interval, and passage frequency, respectively. Longer dwell time, shorter cross-regional time interval, and higher passage frequency correspond to higher transmission risk factors; when the disinfection verification status does not meet the requirements, the corresponding risk factor is increased. The risk prediction server integrates each risk factor and trajectory evidence confidence level according to preset weight rules to obtain edge activation coefficients limited to the preset edge weight range.

[0057] When the exhaust direction of the first node and the intake direction of the second node form an airflow connection, and the airflow arrival time is within the preset propagation time range, the risk prediction server activates the candidate airflow propagation edge from the first node to the second node, forming a dynamic airflow propagation edge.

[0058] When a wind speed measurement terminal is set up on the ventilation path, the risk prediction server determines the median or mean of the wind speed measurement values ​​within the current time slice as the effective wind speed; when no wind speed measurement terminal is set up on the ventilation path, the ventilation flow rate is determined based on the number of operating fans, fan speed, fan performance curve, damper opening, duct cross-sectional area and pipeline resistance parameters, and the effective wind speed is determined in combination with the effective ventilation cross-sectional area.

[0059] When the effective wind speed is less than the preset minimum wind speed threshold, the risk prediction server does not activate the corresponding candidate airflow propagation edge; when the effective wind speed is not less than the preset minimum wind speed threshold, the risk prediction server determines the airflow arrival time based on the airflow path length, and activates the corresponding candidate airflow propagation edge when the airflow arrival time is within the preset propagation time range.

[0060] The edge activation coefficient of dynamic airflow propagation edges is jointly determined by wind speed, wind direction stability, fan operating status, valve opening, pressure differential, path length, filtration status, and environmental evidence confidence level. Higher wind speeds, more stable wind directions, more stable fan operation, larger valve openings, pressure differentials more favorable to transport conditions, shorter path lengths, and less favorable filtration status all correspond to higher airflow propagation risk factors. The risk prediction server integrates each risk factor and the environmental evidence confidence level according to preset weighting rules to obtain the edge activation coefficient of dynamic airflow propagation edges.

[0061] When the same material identified by the material identification information passes through multiple graph nodes in the order of transfer, the risk prediction server activates the candidate material transfer propagation edge in the corresponding direction, forming a dynamic material transfer propagation edge.

[0062] The edge activation coefficient of the dynamic material transfer propagation edge is jointly determined by the number of transfers, transfer intervals, material exposure duration, container cleanliness, material disinfection verification status, and trajectory evidence confidence level. The higher the number of transfers, the shorter the transfer intervals, the longer the material exposure duration, the worse the container cleanliness, and the less the material disinfection verification status meets the requirements, the higher the corresponding propagation risk factor. The risk prediction server integrates each risk factor and the trajectory evidence confidence level according to preset weighting rules to obtain the edge activation coefficient of the dynamic material transfer propagation edge.

[0063] Activated dynamic directed propagation edges are associated with and stored as source graph node identifiers, target graph node identifiers, edge type, edge activation coefficient, propagation arrival time, and the confidence level of the evidence corresponding to the edge. Inactive candidate propagation edges do not participate in the graph convolution calculation for the current time slice.

[0064] Example 5: This embodiment, based on embodiment four, explains the disinfection verification status and its impact on dynamic contact transmission and dynamic material transfer transmission.

[0065] The disinfection verification status is determined based on the identity binding status of personnel, vehicles, tools or materials, the working status of the disinfection terminal, the parameters of the disinfection medium, the duration of disinfection, and the time of disinfection completion.

[0066] When the identity is correctly bound, the disinfection terminal is working properly, the disinfection medium parameters meet the preset requirements, the disinfection duration is not less than the preset duration, and the disinfection completion time is within the preset valid time period, the disinfection verification status is deemed to meet the requirements. If some of the above conditions are pending verification, the status is deemed to be pending verification. When the identity is not bound, the disinfection terminal is not working, the disinfection medium parameters are unqualified, or the disinfection duration is insufficient, the disinfection verification status is deemed to not meet the requirements.

[0067] When the disinfection verification status meets the requirements, the risk prediction server reduces the edge activation coefficient of the corresponding dynamic contact transmission edge or dynamic material transfer transmission edge; when the disinfection verification status does not meet the requirements, it increases the edge activation coefficient of the corresponding dynamic contact transmission edge or dynamic material transfer transmission edge.

[0068] The disinfection verification status is only used as a factor in determining the edge activation coefficient of dynamic contact transmission edge and dynamic material transfer transmission edge in one calculation, and will not be repeatedly corrected for edge activation coefficients that have already been generated.

[0069] Example 6: This embodiment, based on Embodiments 4 and 5, explains the determination of pathogen evidence status, priority, and the updating method for conflict-pending re-examination status.

[0070] When a graph node obtains a pathogen detection result that meets the preset detection quality conditions, has a pathogen detection evidence confidence level not lower than the detection confidence threshold, is within the valid evidence period, and has a positive detection result, the risk prediction server marks the graph node as a positive anchor state.

[0071] When a graph node does not obtain a positive pathogen detection result that meets the positive anchoring conditions, but obtains a pathogen detection result that meets the preset detection quality conditions, the sampling coverage meets the requirements, the confidence level of the pathogen detection evidence is not lower than the detection confidence threshold, it is within the valid evidence period, and the detection result is negative, the risk prediction server marks the graph node as a negative constraint state.

[0072] When a graph node does not have available pathogen detection data, or the pathogen detection data exceeds the valid evidence period, or the detection quality does not meet the requirements, the risk prediction server marks the graph node as invalid or missing.

[0073] Within each time slice, the risk prediction server updates the status according to the priority of positive anchor status, conflict pending re-examination status, negative constraint status, and invalid or missing status. Among them, the positive anchor status has the highest priority; when a graph node has a valid positive pathogen detection result, it will not change to the conflict pending re-examination status due to the conflict condition of the transmission edge.

[0074] When the target graph node is in a negative constraint state, invalid state, or missing state, and there is a source graph node in a positive anchor state, and the edge activation coefficient exceeds the preset edge activation threshold and points to the target graph node, the risk prediction server performs time association judgment according to the edge type of the dynamic directed propagation edge.

[0075] When the conditions for dynamic contact propagation or dynamic material transfer propagation are met, the cross-regional time interval or transfer interval should be less than the corresponding preset time threshold. When the conditions for dynamic airflow propagation are met, the airflow arrival time should be within a preset time window determined based on the effective wind speed and the length of the airflow path.

[0076] When both the edge activation coefficient and the time correlation condition are met, the risk prediction server updates the target graph node to a conflict pending re-inspection state, while retaining the original negative pathogen detection result or invalid or missing marker. The updated node attributes also include the confidence level of pathogen detection evidence, the confidence level of trajectory evidence, the confidence level of environmental evidence, the node identifier of the conflict source graph, and the identifier of the dynamic directed propagation edge of the conflict source.

[0077] Example 7: This embodiment, based on Embodiment Six, explains the multi-relation spatiotemporal directed propagation graph, confidence gating layer, multi-relation directed graph convolutional layer, and temporal risk reasoning process.

[0078] A multi-relational spatiotemporal directed propagation graph formed by consecutive time slices comprises multiple graph snapshots arranged in chronological order. Each graph snapshot includes a set of graph nodes, a set of dynamically directed propagation edges, a set of node features, and a set of edge features.

[0079] Node characteristics include pathogen detection characteristics, confidence level of pathogen detection evidence, confidence level of trajectory evidence, confidence level of environmental evidence, pathogen evidence status coding, conflict markers, personnel and vehicle activity characteristics, material transfer characteristics, airflow connectivity characteristics, and environmental status characteristics.

[0080] Edge features include edge type, edge activation coefficient, source graph node identifier, target graph node identifier, propagation arrival time, edge-corresponding evidence confidence, and temporal correlation.

[0081] The risk prediction server first performs node feature gating on node features, and then performs feature gating on the dynamic directed propagation process.

[0082] For source graph nodes, node feature gating generates corresponding cross-node gating coefficients for pathogen detection features, personnel and vehicle activity features, material transfer features, airflow connectivity features, and environmental state features. The gated source graph node features are used for subsequent inbound and outbound edge aggregation.

[0083] For the current graph node, the risk prediction server constructs the node's features after gating. These node features include pathogen detection features after gating, as well as personnel and vehicle activity features, material transfer features, airflow connectivity features, environmental status features, and status coding features that retain their original values ​​or have been gated according to the corresponding node.

[0084] When a graph node is under negative constraint, pathogen detection features only participate in the node's own feature transformation and do not participate in incoming edge aggregation and outgoing edge aggregation; personnel and vehicle activity features, material transfer features, airflow connectivity features, and environmental status features still participate in cross-node propagation according to their respective cross-node gating coefficients.

[0085] For edge feature gating, the risk prediction server constructs edge features corresponding to pathogen detection information and edge features corresponding to non-pathogen detection information respectively.

[0086] The edge features corresponding to pathogen detection information include at least the edge activation coefficient, the confidence level of trajectory evidence associated with the edge, the confidence level of environmental evidence associated with the edge, the temporal correlation, the pathogen evidence status of the source graph node, the pathogen detection evidence confidence level of the source graph node, and the validity status of the source graph node evidence. The pathogen evidence status is represented by a status coding method, indicating positive anchoring status, negative constraint status, conflict pending re-examination status, and invalid or missing status.

[0087] For dynamic contact propagation edges and dynamic material transfer propagation edges, the confidence level of trajectory evidence corresponding to the edge is determined by the trajectory data of personnel, vehicles, tools, or materials, while the confidence level of environmental evidence corresponding to the edge is determined by the online status and continuous data collection completeness of the environmental monitoring terminals corresponding to the source and target map nodes. For dynamic airflow propagation edges, the confidence level of trajectory evidence corresponding to the edge is determined by the completeness of the trajectory data of personnel and vehicles related to the airflow path, while the confidence level of environmental evidence corresponding to the edge is determined by the environmental data corresponding to wind speed, wind direction, pressure difference, fan operating status, and valve opening.

[0088] The temporal correlation is determined based on the degree of matching between the propagation arrival time and the preset propagation time range. The temporal correlation of dynamic contact propagation edges and dynamic material transfer propagation edges is determined based on the degree of matching between the cross-regional time interval or transfer interval and the corresponding preset time threshold. The temporal correlation of dynamic airflow propagation edges is determined based on the degree of matching between the airflow arrival time and the preset propagation time range.

[0089] The risk prediction server generates edge feature gating coefficients corresponding to pathogen detection information based on the edge features and edge types, and uses these gating coefficients and edge activation coefficients together for cross-node message transmission of pathogen detection information.

[0090] The edge features corresponding to non-pathogen detection information include at least the edge activation coefficient, the confidence level of the trajectory evidence corresponding to the edge, the confidence level of the environmental evidence corresponding to the edge, and the temporal correlation. The risk prediction server generates edge feature gating coefficients corresponding to the non-pathogen detection information based on the edge features and edge types, and uses these gating coefficients, along with the edge activation coefficients, for cross-node message transmission of personnel and vehicle activity features, material transfer features, airflow connectivity features, and environmental state features.

[0091] The gating parameters for edge features corresponding to pathogen detection information and edge features corresponding to non-pathogen detection information are configured separately. Negative constraint states are not used to change the edge feature gating results corresponding to non-pathogen detection information, thereby ensuring that negative pathogen detection results do not block the cross-node propagation of non-pathogen detection features such as personnel and vehicle activities, material transfer, airflow connectivity, and environmental conditions.

[0092] The multi-relation directed graph convolutional layer employs different incoming edge relation parameters, outgoing edge relation parameters, and edge feature gating parameters for contact propagation edges, airflow propagation edges, and material transfer propagation edges, respectively. For each graph node, the multi-relation directed graph convolutional layer aggregates the incoming edge information pointing to that node and the outgoing edge information pointing from that node to adjacent graph nodes, and combines the aggregation result with the feature transformation result of the node to form the updated node representation.

[0093] In this embodiment, the multi-relation directed graph convolutional layer is configured as two layers. The first layer is used to aggregate the propagation information of directly related graph nodes, and the second layer is used to aggregate the propagation information transmitted through intermediate graph nodes.

[0094] The risk prediction server inputs the updated node representations corresponding to each time slice into the temporal gated neural network layer in chronological order. The temporal gated neural network layer uses gated recurrent units to perform temporal processing on the node representations within a continuous window, obtaining temporal propagation characteristics that reflect the propagation and evolution process of each graph node.

[0095] The risk output layer generates pathogen transmission risk values ​​for each graph node in the next prediction period based on temporal propagation characteristics. These pathogen transmission risk values ​​represent the probability that an object in the corresponding field area will meet preset pathogen risk conditions in the next prediction period.

[0096] Example 8: This embodiment, based on Embodiment Seven, explains the method for determining the main risk contribution edge and the propagation path to be verified.

[0097] For a target graph node whose pathogen transmission risk value exceeds a preset risk threshold, the risk prediction server keeps the neural network parameters, the features of the remaining graph nodes, and the features of the remaining dynamic directed propagation edges unchanged, sequentially selects each dynamic directed propagation edge associated with the target graph node, and sets the edge activation coefficient of the selected dynamic directed propagation edge to zero.

[0098] The risk prediction server re-executes graph convolution processing, temporal inference, and risk output processing on the propagation graph after setting the edge activation coefficients to zero, to obtain the pathogen propagation risk value after setting the edge activation coefficients to zero.

[0099] The risk prediction server compares the pathogen transmission risk values ​​before and after the edge activation coefficient is set to zero, and determines the change between the two as the risk contribution value of the corresponding dynamic directed transmission edge.

[0100] When the risk contribution value of a dynamic directed propagation edge exceeds a preset contribution threshold, and the propagation direction, propagation arrival time, pathogen evidence status of the source graph node, and corresponding evidence confidence meet preset consistency conditions, the risk prediction server identifies the dynamic directed propagation edge as the main risk contribution edge.

[0101] When dynamic contact propagation edges and dynamic material transfer propagation edges meet the consistency condition, the travel time interval or transfer time interval from the source map node to the target map node should be less than the corresponding preset time threshold, and the edge direction should be consistent with the actual movement direction of personnel, vehicles, tools, or materials. When dynamic airflow propagation edges meet the consistency condition, the airflow arrival time should be within the preset propagation time window, and the edge direction should be consistent with the airflow direction from exhaust to intake.

[0102] The source graph node corresponding to the main risk contribution edge should be in a positive anchoring state, or in a conflict pending re-examination state and the pathogen detection characteristics should still have non-zero cross-node transmission capability; the confidence level of the pathogen detection evidence corresponding to the source graph node should not be lower than the preset confidence threshold.

[0103] The risk prediction server sorts the edges according to their arrival time and direction of propagation. When the target node of a major risk contributing edge is the same as the source node of the next major risk contributing edge, the two major risk contributing edges are connected to form a propagation path segment. Multiple propagation path segments are connected in chronological order to form a propagation path to be verified.

[0104] In one implementation, for a target graph node that is in a conflict-pending re-inspection state but whose current risk value does not exceed a preset risk threshold, the risk prediction server can perform risk contribution determination processing on the dynamic directed incoming edge that triggers the conflict-pending re-inspection state and the dynamic directed propagation edge adjacent to the target graph node, so as to determine the main risk contribution edge corresponding to the directional re-inspection position.

[0105] When no major risk contributing edge exceeding the contribution threshold is found in the target graph node in the conflict pending re-inspection state, the risk prediction server will trigger the dynamic directed propagation edge of the conflict pending re-inspection state to be determined as the re-inspection positioning edge, and take the sampling points inside the target graph node, the sampling points inside the source graph node, and the intersection, access control position, air inlet, air outlet or ventilation path node corresponding to the re-inspection positioning edge as the directional re-inspection position.

[0106] Example 9: This embodiment, based on Embodiment 8, describes the generation, transmission, and terminal execution process of control commands.

[0107] When a target graph node is in a conflict-pending re-inspection state, the risk prediction server sends a targeted re-inspection instruction to the corresponding automatic sampling terminal. The targeted re-inspection instruction includes the shortened sampling interval, target sampling point, adjacent sampling points, number of samplings, sampling start time, and sampling end time.

[0108] Adjacent sampling points are determined based on the location corresponding to the main risk contribution edge, including sampling points inside source map nodes, sampling points inside target map nodes, material handover locations, access control locations, exhaust vents, air inlets, or ventilation path nodes. When the main risk contribution edge is not determined, adjacent sampling points are determined based on the re-inspection positioning edge.

[0109] When the primary risk contributing edge is a dynamic contact transmission edge or a dynamic material transfer transmission edge, the risk prediction server sends a secondary disinfection verification command to the corresponding access control terminal. Upon receiving the secondary disinfection verification command, the access control terminal performs identity verification, disinfection medium parameter verification, and disinfection duration verification on personnel, vehicles, tools, or materials intending to enter the target area. If any verification fails, the access control terminal restricts the corresponding object from entering the target area and sends the restriction result, object identifier, and verification failure reason to the risk prediction server.

[0110] When the primary risk contributing edge is a dynamic airflow propagation edge, the risk prediction server sends high-frequency monitoring commands to the corresponding air sampling terminal and environmental monitoring terminal. These high-frequency monitoring commands are used to shorten the air sample collection interval and the environmental data upload interval.

[0111] The risk prediction server determines whether preset ventilation safety parameters are met based on temperature, humidity, ammonia concentration, pressure difference, and minimum air exchange rate uploaded by the environmental monitoring terminal. The ventilation parameter combination includes at least the fan start / stop status, the number of fans in operation, fan speed or frequency converter frequency, inlet air valve opening, exhaust air valve opening, target pressure difference, and minimum air exchange rate.

[0112] When the main risk contributing edge is a dynamic airflow propagation edge, and the temperature, humidity, ammonia concentration, pressure difference, and minimum air exchange rate all meet the preset ventilation safety parameters, the risk prediction server sends a ventilation parameter switching command to the ventilation control terminal to adjust the fan operating status, fan speed, air inlet valve opening, air outlet valve opening, and target pressure difference.

[0113] When the main risk contributing edge is not a dynamic airflow propagation edge, or when any ventilation safety parameter corresponding to the dynamic airflow propagation edge does not meet the requirements, the risk prediction server does not send a ventilation parameter switching instruction, but only maintains a high-frequency monitoring instruction and writes the environmental parameters that do not meet the requirements into the risk event record.

[0114] After receiving the control command, the automatic sampling terminal, air sampling terminal, access control terminal, and ventilation control terminal return a receipt confirmation message to the risk prediction server; after completing the corresponding task, they respectively return the sampling execution status, monitoring execution status, access control execution status, and ventilation execution status.

[0115] Example 10: This embodiment describes the double-buffered model update process based on embodiments one through nine.

[0116] The risk prediction server sets the spatiotemporal propagation graph deep neural network currently used for pathogen transmission risk prediction as the online model. The parameters of the online model are frozen during operation, and no incremental parameter updates are performed. The risk prediction server sets the spatiotemporal propagation graph deep neural network, trained and validated based on the training sample library, as the offline candidate model.

[0117] For any given prediction baseline time, the risk prediction server acquires the re-inspection data, sampling terminal execution status, access control terminal execution status, ventilation control terminal execution status, and subsequent pathogen detection results obtained in the next prediction cycle, all prior to the prediction baseline time.

[0118] When the re-inspection data meets the requirements of normal sampling terminal status, valid detection control, sample retention time, completeness of positioning trajectory, and timestamp synchronization, the risk prediction server will use the multi-relationship spatiotemporal directed propagation graph, corresponding re-inspection data, and terminal execution status within the continuous window before the prediction benchmark time as input samples.

[0119] When a graph node obtains a positive pathogen detection result in the next prediction period that meets the detection quality requirements, the sampling coverage meets the requirements, the evidence is valid, and the confidence level of the pathogen detection evidence is not lower than the detection confidence level threshold, the risk prediction server determines the pathogen risk label corresponding to the graph node as a positive risk label.

[0120] When a graph node obtains a negative pathogen detection result that meets the detection quality requirements, the sampling coverage meets the requirements, the evidence is valid, and the confidence level of the pathogen detection evidence is not lower than the detection confidence threshold in the next prediction period, and no valid positive pathogen detection result is obtained, the risk prediction server determines the pathogen risk label corresponding to the graph node as a negative risk label.

[0121] If no available pathogen detection data is obtained in the next prediction period, the detection quality does not meet the requirements, or the detection data is insufficient to determine a negative result, the risk prediction server will not generate training labels for the corresponding graph nodes.

[0122] The offline candidate model uses a multi-relationship spatiotemporal directed propagation graph within a continuous window as input samples and pathogen risk labels as supervision labels for training. During training, sample weights are set according to the proportion of positive and negative risk labels in the training samples to reduce the impact of class imbalance on model training.

[0123] The offline candidate model uses the parameter version of the current online model or the most recently validated historical model version as its initial parameters. The parameter version of the online model remains unchanged until the candidate model has completed training and validation and meets the switching conditions.

[0124] The risk prediction server selects the confidence level of pathogen detection evidence, the confidence level of trajectory evidence, the confidence level of environmental evidence, the edge activation coefficient, the proportion of pathogen evidence status, the number of dynamic contact transmission edges, the number of dynamic airflow transmission edges, the number of dynamic material transfer transmission edges, and the terminal execution status to form a feature set.

[0125] For each feature dimension in the feature set, the risk prediction server divides the data into multiple statistical intervals based on the distribution of historical training samples. It then calculates the proportion of historical and new training samples within each interval and determines the data distribution offset for that feature dimension based on the difference between these two proportions. The risk prediction server then performs a weighted aggregation of the data distribution offsets for each feature dimension to obtain the overall data distribution offset between new and historical training samples.

[0126] When the overall data distribution offset exceeds a preset distribution offset threshold, the risk prediction server initiates the offline candidate model training and validation process. The validation period is determined based on the incubation period or detection window of the target pathogen.

[0127] During the validation period, the risk prediction server generates prediction categories based on the risk values ​​output by the offline candidate models and preset risk thresholds. Based on graph nodes that predict high risk and subsequently obtain effective positive pathogen detection results, graph nodes that predict high risk but subsequently do not obtain effective positive pathogen detection results, graph nodes that predict low risk but subsequently obtain effective positive pathogen detection results, and graph nodes that predict low risk but subsequently obtain effective negative pathogen detection results, the prediction performance, false alarm rate, false negative rate, and stability index are calculated.

[0128] The stability index is determined based on the prediction performance fluctuations across multiple rolling sub-windows within the validation period. Smaller prediction performance fluctuations indicate higher output stability of the candidate model across different time sub-windows.

[0129] When the overall data distribution offset exceeds the preset distribution offset threshold, and the offline candidate model meets the preset prediction performance threshold, false positive rate threshold, false negative rate threshold, and stability threshold within the verification period, the risk prediction server will switch the offline candidate model to a new online model and save the parameter version of the online model before the switch.

[0130] After a model switch, the risk prediction server records the new online model prediction performance based on the same rolling verification window. When the new online model prediction performance decreases by more than a preset rollback threshold compared to the verification performance at the time of the switch, the risk prediction server reverts to the online model parameter version saved before the switch and writes the model switch event, model version number, and rollback reason to the model version management database.

[0131] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the risk of pathogen transmission in livestock farms, characterized in that, Includes the following steps: S1. Acquire the field object data collected by the terminal; construct a candidate propagation topology with three types of candidate propagation edges, namely contact, airflow and material transfer, based on the field layout, ventilation and passage and transfer routes; S2. Determine the prediction baseline time, select data with timestamps no later than that time, generate evidence confidence, and generate positive anchoring, negative constraint, or invalid or missing status based on pathogen detection data; S3. Activate candidate propagation edges based on personnel, vehicles, materials, airflow and disinfection verification data and calculate the edge activation coefficient as the edge weight of the corresponding dynamic directed propagation edge; update the target graph nodes that are in a negative constraint, invalid or missing state and whose dynamic directed incoming edges come from positive anchor graph nodes and meet the time association conditions to a conflict pending re-inspection state, and construct a multi-relationship spatiotemporal directed propagation graph with continuous windows. S4. Input the propagation graph into the spatiotemporal propagation graph deep neural network, and perform directed aggregation according to edge type with different parameters and edge weights through the multi-relation directed graph convolutional layer; through the confidence gating layer, set the pathogen detection features corresponding to positive anchoring, negative constraint, conflict pending re-examination and invalid or missing states as high cross-node gating, pathogen detection only at this node gating, low cross-node gating with retaining conflict markers, and zero cross-node gating, respectively. Among them, negative constraint does not adjust the cross-node gating of personnel, vehicles, materials, airflow and environmental features; through the temporal gating neural network layer, output the risk value of each graph node in the next prediction period. S5. Based on the pathogen transmission risk value and pathogen evidence status, send control commands to sampling, environmental monitoring, access control, or ventilation control terminals.

2. The method for predicting the risk of pathogen transmission in livestock farms according to claim 1, characterized in that, The site includes at least one of the following: breeding pens, isolation area, feed storage area, drinking water supply area, manure transfer area, personnel changing and disinfection area, vehicle disinfection area, material handover area, ventilation intake area, and ventilation exhaust area. Configure unique identification information for each field object and establish a mapping relationship between each field object and its corresponding terminal; The terminals include pathogen detection terminals, environmental monitoring terminals, personnel and vehicle positioning terminals, material identification terminals, ventilation status acquisition terminals, disinfection status acquisition terminals, access control terminals, sampling terminals, and ventilation control terminals; The sampling terminal includes an automatic sampling terminal and an air sampling terminal. The automatic sampling terminal is used to collect pathogen detection samples, and the air sampling terminal is used to collect air samples within the site or along the ventilation path. The candidate propagation topology is constructed based on the physical layout of objects in the field, the connection relationship of ventilation ducts, the correspondence between exhaust vents and air inlets, the permitted passage routes for personnel and vehicles, and the material transfer routes. The candidate propagation edges only represent the potential propagation conditions between objects in the field.

3. The method for predicting the risk of pathogen transmission in livestock farms according to claim 2, characterized in that, Each terminal generates a data packet based on a unified clock reference. The data packet includes a timestamp, terminal identification information, and field object identification information. According to the preset time window, pathogen detection data, environmental monitoring data, personnel and vehicle trajectory data, material transfer data, ventilation status data, disinfection status data and access control records are time-aligned and associated with the corresponding graph nodes. Data with timestamps later than the prediction baseline time will not participate in the graph construction and model inference of the current prediction task. The confidence level and effective period of pathogen detection evidence are generated based on the sample preservation time, the self-test status of the detection terminal, the status of positive and negative controls, the consistency of repeated tests, the status of the detection signal, and the sampling coverage. The confidence level of trajectory evidence is generated based on the integrity of the positioning signal, the number of positioning jumps, the duration of dwell time, the continuity of the trajectory, and the identity recognition matching status. The confidence level of environmental evidence is generated based on the sensor's online status, calibration status, continuous data acquisition integrity, and abnormal fluctuation status.

4. The method for predicting the risk of pathogen transmission in livestock farms according to claim 3, characterized in that, The activation of candidate propagation edges and the calculation of edge activation coefficients include: When the same person, vehicle or reusable tool passes through the first map node and the second map node in sequence within a preset time interval, the candidate contact propagation edge pointing from the first map node to the second map node is activated, forming a dynamic contact propagation edge. The activation coefficient of the corresponding edge is calculated based on the dwell time, cross-area time interval, passage frequency, disinfection verification status and trajectory evidence confidence. When the exhaust direction of the first node and the intake direction of the second node form an airflow connection, and the airflow arrival time is within the preset propagation time range, the candidate airflow propagation edge pointing from the first node to the second node is activated to form a dynamic airflow propagation edge. The activation coefficient of the corresponding edge is calculated based on wind speed, wind direction stability, fan operating status, valve opening, pressure difference status, path length, filtration status and environmental evidence confidence. When the same material identified by the material identification information passes through multiple graph nodes in the order of transfer, the candidate material transfer propagation edge in the corresponding direction is activated to form a dynamic material transfer propagation edge. The activation coefficient of the corresponding edge is calculated based on the number of transfers, transfer interval, material exposure time, container cleanliness, material disinfection verification status, and trajectory evidence confidence.

5. The method for predicting the risk of pathogen transmission in livestock farms according to claim 4, characterized in that, The disinfection verification status is determined based on the identity binding status of personnel, vehicles, tools or materials, the working status of the disinfection terminal, the parameters of the disinfection medium, the duration of disinfection and the time of disinfection completion. When the disinfection verification status does not meet the preset requirements, increase the edge activation coefficient of the corresponding dynamic contact transmission edge or dynamic material transfer transmission edge. When the disinfection verification status meets the preset requirements, the edge activation coefficient of the corresponding dynamic contact transmission edge or dynamic material transfer transmission edge is reduced.

6. The method for predicting the risk of pathogen transmission in livestock farms according to claim 4, characterized in that, The status of pathogen evidence is determined in the following ways: When a graph node obtains a positive pathogen detection result that meets the preset detection quality conditions and is within the corresponding valid evidence period, the graph node is marked as a positive anchoring state. When a graph node obtains a negative pathogen detection result that meets the preset detection quality conditions, the sampling coverage meets the requirements, and it is within the corresponding valid evidence period, the graph node is marked as a negative constraint state. When pathogen detection data exceeds the valid period of evidence, the detection quality does not meet the requirements, or no usable pathogen detection data is obtained, the corresponding graph node will be marked as invalid or missing. When the target graph node is in a negative constraint state, invalid state, or missing state, and there is a source graph node in a positive anchor state, the edge activation coefficient exceeds the preset threshold, and a dynamic directed propagation edge points to the target graph node, if the dynamic directed propagation edge is a dynamic contact propagation edge or a dynamic material transfer propagation edge, then it is determined whether the cross-regional time interval or transfer interval is less than the preset time threshold. If the dynamic directional propagation edge is a dynamic airflow propagation edge, then determine whether the airflow arrival time is within a preset time window determined based on wind speed and path length; When the judgment result meets the corresponding conditions, the target graph node is updated to a conflict pending re-examination state, and the original negative pathogen detection result or invalid or missing marker is retained as node attribute.

7. The method for predicting the risk of pathogen transmission in livestock farms according to claim 6, characterized in that, The confidence gating layer generates gating coefficients corresponding to node features and edge features based on the status of pathogen evidence, the confidence level of pathogen detection evidence, the confidence level of trajectory evidence, the confidence level of environmental evidence, and the effective time period of evidence. For pathogen detection features corresponding to invalid or missing states, set their cross-node gating coefficient to zero; For pathogen detection features corresponding to conflict-pending re-examination status, set their cross-node gating coefficient to be greater than zero and less than the preset first gating threshold, and retain the conflict marker; For pathogen detection features corresponding to positive anchoring states, their cross-node gating coefficient is set to be no less than a preset second gating threshold, the second gating threshold is greater than the first gating threshold, and the first gating threshold is greater than zero; For pathogen detection features corresponding to negative constraint states, only the pathogen detection features of the current graph node are adjusted, and the cross-node gating coefficients of personnel and vehicle activity features, material transfer features, airflow connectivity features, and environmental state features are not adjusted.

8. The method for predicting the risk of pathogen transmission in livestock farms according to claim 1, characterized in that, It also includes the steps for determining the main risk contribution sides: For a target graph node whose pathogen transmission risk value exceeds a preset risk threshold, while keeping the neural network parameters, other graph node features, and other dynamic directed propagation edge features unchanged, the edge activation coefficients of each dynamic directed propagation edge associated with the target graph node are sequentially set to zero, and the pathogen transmission risk value of the target graph node is obtained again. The risk contribution value corresponding to each dynamic directed propagation edge is determined by subtracting the pathogen transmission risk value after the edge activation coefficient is set to zero from the pathogen transmission risk value before the edge activation coefficient is set to zero. When the risk contribution value of a dynamic directed propagation edge exceeds the preset contribution threshold, and its propagation direction, propagation arrival time, pathogen evidence status of the source graph node, and corresponding evidence confidence meet the preset consistency conditions, the dynamic directed propagation edge is identified as the main risk contribution edge, and a propagation path to be verified is generated according to the connection order of multiple main risk contribution edges.

9. The method for predicting the risk of pathogen transmission in livestock farms according to claim 8, characterized in that, The control commands include: When a target graph node is in a conflict pending re-inspection state, a directional re-inspection instruction is sent to the corresponding automatic sampling terminal. The directional re-inspection instruction includes shortening the sampling interval and increasing the sampling number of sampling points adjacent to the position corresponding to the main risk contribution edge. When the main risk contribution side is the dynamic contact transmission side or the dynamic material transfer transmission side, a secondary disinfection verification command is sent to the corresponding access control terminal to control the access control terminal to restrict objects that have not passed the identity recognition, disinfection medium parameter verification and disinfection duration verification from entering the corresponding site area. When the main risk contributing edge is the dynamic airflow propagation edge, a high-frequency monitoring command is sent to the corresponding air sampling terminal and environmental monitoring terminal; when the temperature, humidity, ammonia concentration, pressure difference and minimum ventilation rate all meet the preset ventilation safety parameters, a ventilation parameter switching command is sent to the ventilation control terminal to adjust the fan operating status and damper opening to the preset ventilation parameter combination; when the ventilation safety parameters are not met, the ventilation parameter switching command is not sent.

10. The method for predicting the risk of pathogen transmission in livestock farms according to claim 1, characterized in that, It also includes the double-buffered model update step: Acquire re-inspection data, sampling terminal execution status, access control terminal execution status and ventilation control terminal execution status before the prediction baseline time, as well as subsequent pathogen detection results obtained in the next prediction cycle; When the re-examination data meets the requirements of normal sampling terminal status, valid detection control, sample storage time, completeness of positioning trajectory, and timestamp synchronization, the re-examination data and corresponding terminal execution status before the prediction reference time are used as input samples for the corresponding time window, and the subsequent pathogen detection results obtained in the next prediction period are written into the training sample library as risk labels. The spatiotemporal propagation graph deep neural network currently used for risk prediction is set as an online model with frozen parameters. No incremental parameter updates are performed on it during the online model's operation. The spatiotemporal propagation graph deep neural network that is trained and validated based on the training sample library is set as an offline candidate model. When the data distribution offset between the newly added training samples and the historical training samples exceeds the preset threshold, and the offline candidate model meets the preset prediction performance threshold, false alarm rate threshold, false negative rate threshold and stability threshold within the verification period determined based on the target pathogen incubation period or detection window, the offline candidate model will be switched to a new online model. After the switch, the new online model prediction performance is recorded based on the same rolling verification window. When the prediction performance drops beyond the preset rollback threshold, the online model parameter version before the switch is restored.

Images